<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Paolo Delano on Medium]]></title>
        <description><![CDATA[Stories by Paolo Delano on Medium]]></description>
        <link>https://medium.com/@Desk_Rider?source=rss-2bfa3d9425c8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_dc46f6dRqEdGH72csLmjg.jpeg</url>
            <title>Stories by Paolo Delano on Medium</title>
            <link>https://medium.com/@Desk_Rider?source=rss-2bfa3d9425c8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 00:37:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Desk_Rider/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Stop Token-Incinerating Your Risk Engine]]></title>
            <link>https://medium.com/neo4j/stop-token-incinerating-your-risk-engine-b184a177694e?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/b184a177694e</guid>
            <category><![CDATA[semantic-layer]]></category>
            <category><![CDATA[fraud]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 20:42:12 GMT</pubDate>
            <atom:updated>2026-09-15T21:19:28.664Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UM47QmFKtBLmgpDYwdlXJA.jpeg" /></figure><h4>The winners in fintech won’t buy larger context windows, they will build context rich semantic layers. Here is the math proving why.</h4><p>The global financial system spends over <strong>$200 billion annually</strong> on anti-money laundering (AML) and compliance. Yet, law enforcement recovers less than <strong>1%</strong> of illicit financial flows globally.</p><p>The companies that win the next decade of fintech won’t be the ones buying larger context windows or burning millions on bigger models. The winners will build <strong>context rich semantic layers, </strong>capable of<strong> </strong>translating raw network structures into instant structural intuition.</p><h3>The Three Evolutions of Context</h3><p>In his book <em>Blink</em>, Malcolm Gladwell writes about “thin-slicing”, human’s ability to spot macro patterns from micro signals in a fraction of a second using intuition. Modern AI models possess extraordinary processing power, but when we hand them financial transaction records, we trap them in a blind spot.</p><p>To understand why AI struggles with financial fraud, you have to look at the three ways we represent the world to a machine:</p><h4>1. Flat Data</h4><p>For decades, compliance has relied on relational databases (flat, tabular spreadsheets). <em>Account A sent $5,000 to Account B at 10:01 AM.</em> When you feed flat data to an AI agent, you are asking it to solve a complex mystery by reading isolated names out of a phone book. It can analyze individual events, but it has no awareness of the relationships between things.</p><h4>2. Standard Knowledge Graphs</h4><p>To fix this, the industry upgraded to Knowledge Graphs. We mapped entities as nodes and transfers as edges. Now the AI knows that <em>Account A is connected to Account B</em>. But here is the trap: dumping raw graph connections into an LLM’s context window forces the model to manually trace every edge, step-by-step. And while the patterns exist, there is a chance it will miss it or it will hallucinate.</p><h4>3. Enriched Knowledge Graphs</h4><p>This is the breakthrough. Instead of forcing the AI to manually trace every transaction between accounts, we pre-calculate the underlying structure of the network using graph algorithms (e.g. Community Detection, Centrality, Shortest-Path, Similarity)<strong> </strong>and save the results as node properties to enrich the knowledge graph (KG).</p><h3>The Hypothesis</h3><p>An AI agent provided with an <em>enriched graph context, </em>augmented with pre-computed <em>graphy</em> metrics (PageRank centrality, Louvain community clusters, node similarity, and shortest-path distance matrices), will identify synthetic identities, suspicious fraudulent accounts, and central mules significantly faster and with higher precision than an identical agent given raw graph data.</p><p>Over the next three sections, we put this hypothesis to a head-to-head trial using a synthetic dataset of 30,000 accounts, 5 million transactions, hidden mule rings, and stitched identities. We challenge our agents with the three hardest questions in modern financial crime:</p><ul><li><strong>Spotting Synthetic Identities:</strong> Can the agent isolate stitched fake profiles operating across shared PII clusters?</li><li><strong>Identifying Suspicious Fraud Targets:</strong> Can the agent spot low-volume accounts acting as critical structural bridges before a bust-out occurs?</li><li><strong>Isolating Influential Mules:</strong> Can the agent trace multi-hop circular laundering loops?</li></ul><p>Look at this first image. What you’re staring at is the digital footprint of a modern financial institution: a tangled, high-entropy web of clients, merchants, banks, and wire transfers. Somewhere inside this labyrinth of ordinary life, millions of illicit dollars are silently moving. To a traditional rule-based monitoring engine, it looks like ambient noise. To a human investigator, it looks like a haystack with an invisible needle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8h-YyTkv8Ev5mtP6zw8w4g.png" /><figcaption>Raw Network</figcaption></figure><p>Zooming closer into the network reveals the raw connective tissue. You begin to see how individual entities anchor themselves to the grid , not just through financial transfers, but through shared Personally Identifiable Information (PII) like Social Security Numbers, phone numbers, and email addresses. On paper, these are isolated transactional records; in reality, they are the subtle, structural threads connecting real people to digital phantoms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oVe_SmrXC5mWprwRIbLtoQ.png" /><figcaption>Individual entities with shared PII</figcaption></figure><p>Before Large Language Models entered the auditor’s toolbelt, unmasking this hidden topology required a deliberate, multi-step data science workflow. Forensic analysts couldn’t simply query a database for “fraud”; they had to systematically reshape the graph’s geometry. First, you applied community detection algorithms, like Louvain or Connected Components, to slice through the ambient noise and group tightly interconnected clusters into distinct digital neighborhoods. Once those boundary lines were drawn, you ran centrality algorithms, such as PageRank or Betweenness Centrality , to measure topological gravity. Suddenly, out of a chaotic mass of thousands of client nodes, the math isolated the high-traffic hubs and critical structural bridges. This was the traditional playbook: convert raw, overwhelming scale into a tight, highly targeted short-list of suspicious suspects for human sleuths to audit.</p><p>The magic of <em>graphy </em>metrics, like centrality, distance, or clusters, isn’t that it gives you more data; it’s that it gives you instant spatial intuition. When you look at this cluster, your eyes don’t wander across hundreds of transactions, they drop straight onto the bright pink nodes bridging these communities. In a matter of milliseconds, pre-computed centrality transforms a complex, invisible money-laundering web into an obvious pattern you can’t unsee.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wlTaJBXFXifSl-cVrMOlxQ.png" /><figcaption>Network enhanced with clustering and centrality</figcaption></figure><p>And this is what we expect the LLM to use when we prompt it to answer the three questions above. Our main suspects from the image above when we filter by community size and centrality are Lillian Nichols, Nicholas Quinn, Evelyn Frost, Charlotte Mccoy, Nathaniel Mcdaniel, Kimberly Alvarado, Thomas Gomez, Olivia Charles, Chase Dunlap, and Samuel Petty.</p><h4><strong>Spotting Synthetic Identities</strong></h4><p><strong>Prompt</strong>: <em>Identify potential synthetic identities using personal identifiable information like email, phone number, social security number, and customer name</em></p><p><strong>Raw knowledge graph results</strong></p><blockquote>Claude: Virtually every case of a shared identifier comes with a different name each time — never the same person’s name reappearing on the same SSN/email/phone. That pattern (one anchor identifier + a rotating cast of names) is the textbook signature of synthetic identities: a real or fabricated core credential stretched across multiple constructed personas.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*kJBAWwiWmclwZDrcVK4ScQ.png" /><figcaption>Synthetic ID results from raw KG</figcaption></figure><h4><strong>Enhanced knowledge graph results</strong></h4><blockquote>Claude: Using GDS Weakly Connected Components (WCC) over the identity graph, I found accounts linked by shared PII — the classic synthetic-identity signature: distinct customer names recycling a limited pool of real/stolen emails, phone numbers, and SSNs. 855 clusters of 2+ accounts sharing at least one identifier, covering 3,042 accounts total (average cluster size 3.6, largest cluster 11 accounts). 15 clusters have 8 or more members.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/390/1*P5yMLKwU83qDdkVtdiI85Q.png" /><figcaption>Synthetic ID results from enhanced KG</figcaption></figure><p>Look closely at the raw KG results, our AI agent walked right past the biggest fire in the building. It completely missed an 11-person cluster built on top of just 3 recycled Social Security Numbers , the exact smoking gun of a syndicate operating in plain sight. The agent didn’t get a better model, a larger context window, or a smarter prompt. <strong>The only difference was the pre-calculation of clusters using community algorithms.</strong> By simply shifting from raw graph to pre-computed properties, the invisible identity factory instantly lit up like a neon sign.</p><p>On the other hand, the agent using the enhanced KG accurately spotted the largest cluster with 11 clients and 3 shared SSNs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pi2pxXE7EYawb9S_EK5PfQ.png" /><figcaption>Largest cluster with shared PII</figcaption></figure><h4><strong>Identifying Suspicious Fraud Targets</strong></h4><p><strong>Prompt</strong>: identify “bridge” client nodes within suspicious transaction communities that include 7 or more client nodes.</p><p><strong>Raw knowledge graph results</strong></p><blockquote>Claude: I then searched for pure Client nodes that transact with members of 2+ of these large rings — same method as before (pure Cypher traversal + aggregation, no gds.*).<br><strong>Excluded again as noise:</strong> Andrew Harvey, Thomas Gomez, Samuel Petty — the three mega-hub clients keep surfacing because their sheer transaction volume (50K–106K transactions) makes them touch dozens of rings incidentally, not deliberately.<br>Of 262 remaining candidates, most connect exactly 2 communities. A few stand out by connecting 3–4 communities:</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rFExi_eqBZ2H_EYEGRCQcw.png" /><figcaption>Bridge account results from raw KG</figcaption></figure><p><strong>Enhanced knowledge graph results</strong></p><blockquote>Claude: Expanding the search to <strong>every</strong> synthetic-identity ring of size ≥7 (65 such rings found via GDS WCC on the graph) and checking, for each, which external account receives confirmed-fraud transfers from 2+ of its members.<br><strong>Headline numbers</strong>: 65 of the 65+ rings have an identifiable bridge account, 66 unique bridge accounts total, funneling <strong>≈ $16.44M</strong> combined, averaging 7.4 ring members routed through each bridge.<br>Top 15 bridges by amount moved:</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XtLKp9dfn1sE4ODR7qEzBA.png" /><figcaption>Bridge account results from enhanced KG</figcaption></figure><p>The empirical divide between these two approaches couldn’t be starker.</p><p>When we call the <strong>Enriched Agent, </strong>it didn’t just stumble onto our primary suspect accounts; it mapped the entire battlefield. Cross-referencing its findings against our manual forensic baseline revealed a stunning 100% capture rate: <strong>every single client node possessing an above-average betweenness centrality score was identified instantly.</strong> The agent didn’t need to guess where the structural risk lay , the network properties pointed straight at the bottlenecks. Notice in the image below, a subsection of the biggest suspect community, where Lillian Nichols and Nicholas Quinn (among others) act as central nodes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zJEEaq5XLgE0no33J-wAeg.png" /><figcaption>Correct flagging of bridge client accounts</figcaption></figure><p>Meanwhile, the <strong>Raw Agent</strong> was completely lost in the fog.</p><p>Operating on un-enriched graph data, it failed to identify a <em>single</em> primary suspect from our target list. Out of the four clients it flagged as “suspicious,” only two actually belong to the suspect communities, and both possess negligible betweenness scores. The raw model was effectively chasing ghost signals: blowing tokens on low-value noise while walking right past the high-leverage bridges running the entire syndicate. Take Dylan Collier as an example. Although he belongs to one of our suspect communities, he is clearly not influencing its behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RIHp_kJRtGtZmie8I8urvg.png" /><figcaption>Incorrect flagging of a bridge client account</figcaption></figure><p>This isn’t an incremental upgrade in AI performance. It is the definitive proof that without pre-computed topological context, even the smartest models are functionally blind to systemic financial crime.</p><h4>Identifying Money Laundering Loops</h4><p><strong>Prompt</strong>: Give me the top 10 most suspicious accounts and the accounts in the biggest money laundering loop ranked by transaction amount.</p><p><strong>Raw knowledge graph results</strong></p><blockquote>Claude: <strong>Part 1 — Top 10 most suspicious accounts </strong>I pulled the strongest signal from every angle we’ve explored this session — confirmed fraud-ring leadership, cross-community bridging, and literal circular money loops — and ranked by severity of evidence, then amount.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*adYHizPBUilVBpa-pBXPug.png" /><figcaption>Top 10 suspicious accounts</figcaption></figure><blockquote><strong>rows 1–3 and 10 are a genuinely new finding — a real circular fund flow, invisible to every fraud flag or ring-detection method we’d used so far, because none of these six people (across the two loops) are labeled </strong><strong>Mule .</strong></blockquote><blockquote><strong>Part 2 — Biggest money-laundering loop, ranked by transaction amount</strong><br>Searching for actual directed cycles (not just fan-in rings) turned up none inside the confirmed-fraud mule network — that network is a pure fan-in/collector structure with zero reciprocal flow. But widening to <em>all</em> Client/Mule transfers, real 3-party loops exist. The largest by total value moved:<br><strong>Ethan Vinson → Parker Hester → David Browning → Ethan Vinson</strong>, total $13,363,364.27 circulated.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MycABg_YJGHX5N5ykyzh-g.png" /><figcaption>Biggest laundering loop by txn amount</figcaption></figure><blockquote>All three are unflagged Client accounts with moderate, unremarkable transaction volume (152–407 total transactions each) — nothing about them would trip a standard fraud filter individually. It&#39;s only the closed-loop shape of the money (funds leaving and eventually returning to the origin through two intermediaries) that reveals it as a classic layering/round-tripping pattern. Runner-up loop: Emma Sherman → Molly Santos → Ariana Wilder → Emma Sherman, $10,227,463 total.</blockquote><p><strong>Enhanced knowledge graph results</strong></p><p>Claude:<strong> Top 10 most suspicious accounts</strong> (ranked by total dollar amount touching confirmed fraud transfers, sent + received):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XKlrkpp4pFs1Jy1_vpT2Ug.png" /></figure><p>All ten are pure receivers except Brooklyn Welch, who’s the only one that <em>sent</em> rather than received. Gianna Kane is also the #1 bridge account from the previous analysis (funneling an 8-member synthetic ring). None of these 10 belong to the 5 originally-tagged big_fraud_group rings, confirming again that the highest-dollar fraud activity sits outside current monitoring tags.<br><strong>On the &quot;biggest laundering loop&quot;</strong>: I checked rigorously for an actual circular structure (A→B→C→A) using GDS Strongly Connected Components on the fraud-only transaction graph, plus explicit 2-hop and 3-hop cycle searches across every account touched by confirmed fraud — <strong>there is no closed loop</strong>. The confirmed-fraud network has zero cycles; money only flows one direction, from synthetic identities into a single collector (hub-and-spoke, not circular). <br>Given that, the closest legitimate answer to &quot;biggest loop&quot; is the <strong>largest fraud ring</strong> we identified (cluster 31271), ranked by transaction amount:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1005/1*ufSiN-zOQ9PxeHMamfUTWA.png" /><figcaption>Largest Fraud Ring Members</figcaption></figure><p>The takeaway from this trial delivers a fascinating, counter-intuitive plot twist, one that completely reshapes how, I at least, think about human-AI collaboration in financial forensics, or any other investigation really.</p><p>When we turned the <strong>Enriched Agent</strong> loose to hunt for laundering loops, something unexpected happened. Because we strictly instructed it to rely on pre-computed graph properties, it was mathematically constrained: it actually failed to trace the closed-loop transaction chains step-by-step.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RxX3ogM6X2VtirTWNyy5sg.png" /><figcaption>Confirmed Closed-Loop Transaction</figcaption></figure><p>Yet, it pulled off something far more valuable. Instead of getting bogged down in individual money trails, it leveraged the macro-geometry of the network to isolate the single largest cluster of initial suspects, instantly identifying it as a coordinated laundering cartel.</p><p>Meanwhile, the <strong>Raw Agent</strong> played the role of the meticulous, microscopic investigator.</p><p>By filtering raw transaction flows tied to known fraud, it successfully mapped a classic, closed-loop laundering sequence (image above), the exact multi-hop circuit bad actors use to layer illicit cash. Out of our ten baseline hypotheses, it hit on two, while surfacing a fresh batch of previously unsuspected accounts that warrant a deep-dive audit. Where the enriched model saw the whole forest, the raw model actually managed to trace a specific tree.</p><p>The winning strategy isn’t choosing one agent over the other. It’s building a hybrid context engine, one that uses pre-computed graph data science properties to narrow the search, and raw cypher to catch the thieves in the act.</p><p>I’d like to thank my colleague <a href="https://medium.com/u/94de2efeae4f">Laura Parra</a> for her help in creating this experiment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b184a177694e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/neo4j/stop-token-incinerating-your-risk-engine-b184a177694e">Stop Token-Incinerating Your Risk Engine</a> was originally published in <a href="https://medium.com/neo4j">Neo4j Developer Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Precise Economic Policy in a World of Chaos]]></title>
            <link>https://medium.com/neo4j/precise-economic-policy-in-a-world-of-chaos-8e9b68588442?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e9b68588442</guid>
            <category><![CDATA[supply-chain-risk]]></category>
            <category><![CDATA[economic-policy]]></category>
            <category><![CDATA[graphrag]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[neo4j]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 23:21:51 GMT</pubDate>
            <atom:updated>2026-04-22T18:27:53.740Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why the Future of Macro-Risk is Agentic and Interconnected</h3><p>The phone rang at 3:14 AM. For Roberto, who owns a high-volume beverage distribution firm in a bustling <strong>Commercial Hub</strong> in San Salvador, that sound usually means one of two things: a warehouse alarm went off, or his inventory is stuck.</p><p>“Don Roberto,” the voice on the other end crackled. It was his logistics lead. “The crossing at <strong>El Amatillo</strong> is backed up for miles. Protests on the Honduran side. Our shipment of soft drinks isn’t moving for at least forty-eight hours.”</p><p>In the old days, Roberto would have poured a stiff coffee and started guessing. He’d call the beverage manufacturer in Honduras to see if they could reroute. He’d check his warehouse levels for raw sugar, wondering if his supplier’s production would halt too. He’d pray the shelves in San Salvador didn’t go empty.</p><p>But this morning, Roberto didn’t guess. He opened his laptop.</p><h4>Mapping the Ripple: From Raw Sugar to the Hub</h4><p>Roberto doesn’t just see a “truck.” He sees a <strong>Knowledge Graph</strong>. To him, the delay at the border isn’t a static event; it’s a node in a living web. The visual below isn’t just a map; it’s a semantic layer where the relationship is the primary data point and treats every border crossing, commercial hub, and supplier as an interconnected node.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F1PRzFOdy1vPgk3WlIxGXw.png" /><figcaption>Border Crossings (yellow), Commercial Hubs (purple)</figcaption></figure><p>By mapping these dependencies, the system can instantly propagate a disruption at El Amatillo (pointed at) across the entire network, transforming a mess of unstructured news, macroeconomic indicators, and logistics logs into a predictive, traversable digital twin of the Central American economy.</p><p>Take a look at the graph model below. This isn’t just a collection of nodes, it’s a blueprint of economic causality.</p><p>At the edge of the network, <strong>News Articles</strong> (tan nodes) act as the primary triggers — sentiment and events that ripple through <strong>Commodities</strong>, <strong>Suppliers</strong>, <strong>Border Crossings</strong>, or <strong>Commercial Hubs</strong>. In this ecosystem, commodities serve as the essential raw inputs for the <strong>Products</strong> that <strong>Suppliers</strong> bring to market. <strong>Businesses</strong> (light blue nodes) are anchored to specific <strong>Commercial Hubs</strong>, and the entire system culminates in a <strong>Transaction</strong>: the moment value is exchanged between a business and a supplier as a product physically traverses a <strong>Border Crossing</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MeXxC_I8_cGuLBpihYbGdQ.png" /><figcaption>Blueprint of Economic Casualty</figcaption></figure><p>When that border node turns red, the graph calculates the velocity of the disaster. Because the truck is stuck, Roberto’s inventory in San Salvador will hit “critical low” in 48 hours. But the graph shows him something deeper: the Honduran manufacturer is also running low on raw sugar due to the same bottleneck. Even if the trucks move tomorrow, the supplier’s production line is about to stall.</p><p>When we zoom out from Roberto’s warehouse, the graph reveals a much more systemic vulnerability. This isn’t just one man’s inventory problem; it’s a structural threat to the city’s commercial pulse.</p><p>In the model, <strong>Raw Sugar</strong> acts as a highly dependent node. By traversing the graph, we see that this single commodity is the lifeblood for <strong>three distinct finished goods.</strong> These three products alone power <strong>hundreds of individual transactions</strong> daily across the district.</p><p>The ripple effect terminates at <strong>seven major businesses</strong> within the San Salvador hub that depend on these goods to keep their doors open. When the sugar node is “choked” at the border, the graph doesn’t just show a delay; it calculates a synchronized downturn. If those seven businesses lose their primary products, the commercial hub sees a measurable drop in economic velocity, potentially leading to localized labor hour reductions or price hikes that affect the end consumer within days. Through the lens of the graph, a single stalled truck at El Amatillo is directly linked to the empty shelves of seven different retailers in the heart of the capital.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*myBMEQrqBxoiledxUvV5Ig.png" /><figcaption>Risk of high dependency on raw sugar</figcaption></figure><h4>Agentic Synthesis: Autonomous Intelligence in the Graph</h4><p>Moving from a static map to a predictive engine requires more than just data; it requires an Agentic Framework that understands the specific context of each economic entity. Instead of a single AI trying to solve everything, we deploy a squad of specialized domain agents that act as the digital eyes of the network.</p><p>Each agent is primed with a specific mission and a set of <strong>Relationships </strong>that define how they interact with incoming information:</p><ul><li><strong>The Border Crossing Agent:</strong> Monitors news for civil unrest, weather alerts, or administrative strikes at specific transit points.</li><li><strong>The Provider Agent:</strong> Scans for corporate news, labor disputes, or manufacturing delays. It connects news about a parent company to a subsidiary supplier.</li><li><strong>The Commodity Agent:</strong> Watches global markets and environmental reports. It knows that a drought in the region isn’t just “weather”, it’s a future shortage of the raw sugar needed for production.</li><li><strong>The Commercial Hub Agent:</strong> Focused on the “last mile.” It monitors local municipal news in San Salvador, such as road closures or power grid failures that could stall distribution.</li></ul><p>Here is the sample code for the commodity agent. Notice how we first find all news nodes with a relationship to a commodity node, then traverse the graph.</p><pre>commodity_risk_traversal_query = &quot;&quot;&quot;<br>        MATCH (n:Noticia)-[:AFECTA_A]-&gt;(c:Commodity)<br>        WITH score, c, n<br>        MATCH (c)-[:INSUMO_PARA]-&gt;(p:Producto)&lt;-[:CONTIENE]-(t:Transaccion)-[:ENTREGADA_A]-(e:Empresa)<br>        RETURN  <br>        n.texto AS news_content, <br>        n.severidad as severity,<br>        n.titulo as headline,<br>        c.nombre as commodity,<br>        collect(p.descripcion) as ProductsImpacted,<br>        count(distinct e.id) AS businessesImpacted, <br>        sum(distinct t.valorUSD) as dollarImpact,<br>        score AS SimilarityScore,<br>        elementId(n) AS Metadata_ID<br>&quot;&quot;&quot;<br><br>commodity_retriever = VectorCypherRetriever(<br>    driver,<br>    index_name=&quot;news_vector_index&quot;,<br>    embedder=embedder,<br>    retrieval_query=commodity_risk_traversal_query<br>)</pre><p>The result of this traversal is then used as context for the agent to reason and provide an answer. Here is a sample code using OpenAI</p><pre>response = embedding_client.chat.completions.create(<br>        model=&quot;gpt-4o&quot;,<br>        messages=[<br>            {&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: prompts[risk_type]},<br>            {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;Analyze these news alerts for risk:\n\n{news_context} Indicate dollar amount, product, and business impact&quot;}<br>        ],<br>        temperature=0<br>    )<br></pre><p>And here is a visual representation of that context: it begins with a <strong>News</strong> node (the external trigger) linked to a <strong>Commodity</strong> (the raw input). That commodity is mapped to a <strong>Product</strong>, which serves as the anchor for one ore more <strong>Transactions</strong>, ultimately terminating at <strong>Businesses</strong> within a commercial hub.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H8jO4hZyqbUe7jhq1P5N4A.png" /><figcaption>Visual Representation of the context for our commodity agent</figcaption></figure><p>Ultimately, when Roberto prompts the <strong>Commodity Agent</strong> to filter for specific high-signal events , like a transport strike or a harvest failure, he isn’t just getting a list of headlines. He’s getting a <strong>weighted impact analysis</strong>. Because the agent understands the graph’s topology, it doesn’t just find the news, it identifies the specific “dependency path” from the strike to his bottom line, providing an answer that is as strategically sound as it is technically precise.</p><pre>Based on the news alerts provided, here is an analysis of the potential risk and price volatility for the impacted commodities:<br><br>1. **Hard Red Winter Wheat (ZW)**<br>   - **Headline:** Geopolitical Tensions Impact ZW Pricing<br>   - **Severity:** High<br>   - **Dollar Impact:** $815,754.97<br>   - **Products Impacted:** Wheat Flour, Sweet Biscuits<br>   - **Business Impact:** No specific businesses mentioned as impacted.<br>   - **Analysis:** The uncertainty in international markets has led to speculative buying, which is destabilizing local costs. This could lead to increased price volatility for products derived from Hard Red Winter Wheat, such as wheat flour and sweet biscuits. The geopolitical tensions are likely to cause fluctuations in supply and demand, impacting prices.<br><br>2. **Corn (Maize) (ZC)**<br>   - **Headline:** Logistics Crisis at ZC Infrastructure<br>   - **Severity:** High<br>   - **Dollar Impact:** $1,065,036.03<br>   - **Products Impacted:** Carbonated Soft Drinks, Poultry Meat<br>   - **Business Impact:** No specific businesses mentioned as impacted.<br>   - **Analysis:** The breakdown in specialized transport systems has left cargo stranded, which could lead to supply chain disruptions. This is likely to cause short-term price increases and volatility for corn-based products, including carbonated soft drinks and poultry meat, as supply becomes constrained.<br><br>3. **Raw Sugar (SB)**<br>   - **Headline:** Transport Strike Targets SB Carriers<br>   - **Severity:** High<br>   - **Dollar Impact:** $1,543,928.18<br>   - **Products Impacted:** Carbonated Soft Drinks, Sweet Biscuits, Cane Sugar<br>   - **Business Impact:** No specific businesses mentioned as impacted.<br>   - **Analysis:** The strike by trucking unions demanding higher insurance coverage is likely to disrupt the transportation of raw sugar. This could lead to supply shortages and increased costs for products containing sugar, such as carbonated soft drinks, sweet biscuits, and cane sugar. The strike may result in significant price volatility until the issue is resolved.<br><br>Overall, the high severity of these events suggests a substantial risk of price volatility for the affected commodities and their derivative products. Businesses involved in the production or sale of these products should prepare for potential cost increases and supply chain disruptions.</pre><p>Notice the difference when asking the <strong>Border Agent</strong> for news that mention <em>strikes, delays, or closures</em></p><pre>The news alerts indicate significant disruptions at various Aduana facilities, each with varying degrees of severity and impact on border crossings and transactions. Here&#39;s an analysis of the potential risks and impacts:<br><br>1. **Aduana La Hachadura Facility Explosion**<br>   - **Severity:** Critical<br>   - **Impact:** The indefinite shutdown of this key site due to an explosion will severely disrupt the handling and processing of goods. This will likely lead to delays in border crossings and transactions for the affected products.<br>   - **Dollar Amount:** $934,019.55<br>   - **Products Impacted:** Sweet Biscuits, Cane Sugar, Cement Clinkers, Portland Cement, Wheat Flour, Ethyl Alcohol (Ethanol), Poultry Meat<br>   - **Number of Businesses Impacted:** 7<br><br>2. **Aduana San Cristóbal Flooding**<br>   - **Severity:** High<br>   - **Impact:** The destruction of storage facilities due to flooding will cause significant supply chain disruptions and potential shortages in the region. This could lead to increased prices and delays in transactions.<br>   - **Dollar Amount:** $975,010.44<br>   - **Products Impacted:** Cane Sugar, Poultry Meat, Sweet Biscuits, Wheat Flour, Cement Clinkers, Portland Cement, Ethyl Alcohol (Ethanol)<br>   - **Number of Businesses Impacted:** 7<br><br>3. **Aduana Anguiatú Warehouse Fire**<br>   - **Severity:** Medium<br>   - **Impact:** The fire has destroyed a major distribution center, affecting the supply of several products. This will likely result in temporary shortages and increased costs for businesses relying on these supplies.<br>   - **Dollar Amount:** $1,140,795.04<br>   - **Products Impacted:** Portland Cement, Sweet Biscuits, Cement Clinkers, Poultry Meat, Ethyl Alcohol (Ethanol), Cane Sugar, Wheat Flour<br>   - **Number of Businesses Impacted:** 6<br><br>Overall, these events pose significant risks to border crossings and transactions, with potential delays, increased costs, and market instability. Businesses involved in the affected products should prepare for disruptions and consider alternative supply chain strategies.</pre><pre>border_risk_traversal_query = &quot;&quot;&quot;<br>MATCH (n:Noticia)-[:AFECTA_A]-&gt;(c:Frontera)<br>WITH c, n<br>MATCH (c)&lt;-[:ATRAVIEZA_FRONTERA]-(t:Transaccion)<br>MATCH (t)-[:CONTIENE]-&gt;(p:Producto)<br>MATCH (t)-[:ENTREGADA_A]-&gt;(e:Empresa)<br>RETURN n.texto AS news_content, <br>        n.severidad as severity,<br>        n.titulo as headline,<br>        collect(distinct p.descripcion) as Products,<br>        count(distinct e) AS BusinessCount,<br>        sum(distinct t.valorUSD) as DollarAmount<br>&quot;&quot;&quot;</pre><p>Beyond the individual warehouse, the true breakthrough lies in the <strong>Hidden Risk Agent, </strong>a high-level supervisor designed for macroeconomic oversight. While Roberto focuses on his soda, this agent treats the entire graph as its context, scanning for structural fragility that no human analyst could spot. It doesn’t just wait for a strike; it identifies hidden dependencies , like a single bridge in Guatemala that, if closed, would halt 40% of all beverage transactions in San Salvador. For a Minister of Economy or a central banker, this is the ultimate policy tool. It transforms risk mitigation from a game of whack-a-mole into precise strategy, allowing leaders to fortify the specific nodes (commodities, border crossings, suppliers) that hold the entire regional economy together. By the time a crisis hits the news, the policy response has already been simulated, tested, and deployed.</p><p>Here is one last example using the <strong>Hidden Risk Agent </strong>to look for risks in the textile industry</p><pre>hidden_risk_query = &quot;&quot;&quot;<br>  MATCH (n:Noticia)-[:AFECTA_A]-&gt;(direct_entity)<br>  OPTIONAL MATCH (direct_entity)-[:INSUMO_PARA|PRODUCE|CONTIENE|UBICADA_EN*1..2]-(hidden_target)<br>  WHERE hidden_target &lt;&gt; direct_entity<br>  WITH n, direct_entity, hidden_target, labels(hidden_target)[0] as type<br>  RETURN <br>    n.titulo AS NewsHeadline,<br>    direct_entity.nombre AS DirectImpact,<br>    collect(DISTINCT {<br>        entity: hidden_target.nombre, <br>        type: type,<br>        risk_chain: labels(direct_entity)[0] + &#39; -&gt; &#39; + type<br>    }) AS HiddenDependencies,<br>    count(DISTINCT hidden_target) AS TotalHiddenEntities<br>&quot;&quot;&quot;</pre><pre>The disruptions in the textile commodity sector are highlighted by <br>the news of a speculative bubble bursting for Gildan Activewear. <br>This incident directly impacts Gildan Activewear and is associated <br>with hidden dependencies related to products, transactions, and commodities, <br>specifically Upland Cotton. <br><br>The situation may lead to broader implications for the textile industry, <br>affecting supply chains and pricing.</pre><h4>From Business Survival to National Strategy: A Tool for Economic Policy</h4><p>Traditional economic policy is often like steering a massive tanker by looking at the wake behind it. You see the GDP dip <em>after</em> the quarter ends. You feel the inflation <em>after</em> the prices have already spiked. A knowledge graph changes the temporal nature of governance from reactive to predictive.</p><p>By leveraging the graph, policymakers can move beyond broad, “blunt-force” interventions, as mentioned above, and instead utilize precise strategy application:</p><ul><li><strong>Precision Infrastructure:</strong> Instead of general road funding, the graph identifies which specific border crossings or commercial hubs are “single points of failure” for the highest volume of downstream transactions. Policy follows the path of maximum economic impact.</li><li><strong>Tariff Impact Modeling:</strong> Before a trade agreement is signed, the graph can simulate the “Blast Radius” of a new tariff. It identifies exactly which domestic businesses will face higher input costs allowing the government to negotiate exemptions or prepare targeted relief before the shock hits.</li><li><strong>War and Geopolitical Stress Tests:</strong> By adding a “News Article” node representing a global conflict into the model, the graph maps how a disruption in Baltic grain or Middle Eastern energy cascades through local suppliers and commodities, surfacing hidden vulnerabilities in national food or energy security.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e9b68588442" width="1" height="1" alt=""><hr><p><a href="https://medium.com/neo4j/precise-economic-policy-in-a-world-of-chaos-8e9b68588442">Precise Economic Policy in a World of Chaos</a> was originally published in <a href="https://medium.com/neo4j">Neo4j Developer Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How a $28 Part Can Bring Down a $5M Machine, and How GenAI Stops It]]></title>
            <link>https://medium.com/neo4j/how-a-28-part-can-bring-down-a-5m-machine-and-how-genai-stops-it-cc88d00faf43?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc88d00faf43</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[mining]]></category>
            <category><![CDATA[neo4j]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Tue, 30 Sep 2025 17:24:27 GMT</pubDate>
            <atom:updated>2025-10-10T14:35:11.958Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Create a free graph database instance in </em></strong><a href="https://login.neo4j.com/u/signup/identifier?state=hKFo2SA4YXB2VkxUSjN1eWNZN25RbXZaRExfcjdteW4wUEd5WKFur3VuaXZlcnNhbC1sb2dpbqN0aWTZIExxUEYxb2V5OGU5amRVTGJ2c2RxZXJieUJyZWJmUDhFo2NpZNkgV1NMczYwNDdrT2pwVVNXODNnRFo0SnlZaElrNXpZVG8"><strong><em>Neo4j AuraDB</em></strong></a></p><p>In mining, fortunes are won and lost not in boardrooms but in the quiet moments when a machine decides not to start. Picture a $5 million haul truck, engineered to move mountains, sitting idle because of a missing part that costs less than a dinner for two. The paradox is absurd, but it happens every day. And the reason isn’t the cost of the material; it’s that no one really knows which parts are the ones the whole system can’t live without.</p><p>The challenge isn’t just having materials in stock. It’s knowing which materials matter most, where to keep them, how they connect across machines, and what repair actions to take when failures occur. Traditionally, this knowledge is scattered across three silos:</p><ul><li><strong>Bills of materials (BoM):</strong> The blueprint of assemblies and components</li><li><strong>OEM manuals</strong>: The technical instructions buried in thousands of PDF pages</li><li><strong>Maintenance logs:</strong> The record of what actually failed, how it was repaired, and what materials were consumed</li></ul><p>By connecting these silos through <a href="https://neo4j.com/books/developers-guide-how-to-build-knowledge-graph/"><strong>knowledge graphs</strong></a> and making them explorable with <strong>GenAI</strong>, mining companies can finally unlock a <strong>digital twin</strong> of their equipment — one that reduces downtime, optimizes spare parts, and empowers technicians in the field.</p><p>Have a look at this graph model connecting these silos. You will likely identify some fields (now nodes) from common ERP platforms like SAP.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hyc7nXwtQoLZHrTodbzaRQ.png" /><figcaption>Graph model</figcaption></figure><p>The model is centered around the equipment from where we connect the BoM structure, the maintenance and breakdown events, and the technical documents. My colleagues <a href="https://medium.com/u/aaa89f5d832">Alex Gilmore</a> and <a href="https://www.linkedin.com/in/chris-upkes/">Chris Upkes</a> developed a fantastic Python library to help chunk and extract entities from PDF documents (hopefully made public soon).</p><h3><strong>The Hidden Complexity of BoMs</strong></h3><p>A single mining machine can have:</p><ul><li><strong>Thousands of parts</strong> across multiple levels of nested assemblies</li><li><strong>Multiple BoMs</strong> depending on version, retrofit, or regional variant</li><li><strong>Dependencies across fleets</strong>, where the same component appears in haul trucks, loaders, and drills</li></ul><p>ERP systems like SAP capture this data, but make it hard to navigate. Maintenance teams often spend hours searching for part numbers or repair steps — time the business can’t afford.</p><p>With our knowledge graph, teams can navigate complex equipment structures to find answers quickly and uncover hidden dependencies. For example, a material is used so widely across machines that running out of it creates a massive operational risk.</p><p>The following image is a visualization of a BoM structure (assemblies and material in brown and teal) and the OEM manuals (green nodes) that describe the equipment (blue nodes).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tQTPieBTIXUllqVnE09DVQ.png" /><figcaption>BoM structure and OEM manuals</figcaption></figure><p>And below is a visualization of how a specific type of lubricant oil (blue node) is shared across BoMs (orange nodes) and ultimately equipment (brownish nodes).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nEalO2nrM8ldXvIzalQVIA.png" /></figure><p>With this structure, GenAI can answer natural language questions with grounded evidence:</p><ul><li><em>“Show me the procedure to replace the hydraulic pump on Drill 320XPC, with part numbers.”</em></li><li><em>“List the top five most frequently replaced components on Cat 16M graders and their total downtime cost.”</em></li><li><em>“Which components are common to both haul trucks and loaders and how often have they failed?”</em></li><li><em>“What is the load capacity of the 739 mining truck?”</em></li></ul><p>Neo4j Aura makes this much easier with the Agent toolbar, where you can design tools tailored to different types of questions and let the agent automatically choose the best one to retrieve the answer. Here’s a sample of how that looks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fb3AAo9b-xfgXq7z1lGHCA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gZX544vlomzJjH0BbbWggg.png" /><figcaption>Sample questions using GenAI on our knowledge graph</figcaption></figure><h3>The Hidden Cost of Downtime</h3><p>The economics of maintenance are counterintuitive:</p><ul><li><strong>High cost of downtime:</strong> A single failure can cost hundreds of thousands of dollars.</li><li><strong>Low cost of materials: </strong>The failing component may be worth $50 or less.</li></ul><p>The real challenge is identifying <strong>which materials are truly critical</strong>. In isolation, a gasket looks unimportant. In a network, that gasket may have <strong>high centrality</strong>, meaning it’s used across dozens of machines.</p><p>Graph algorithms expose these blind spots:</p><ul><li><strong>Degree centrality: M</strong>aterials that appear most often across BoMs</li><li><strong>Betweenness centrality:</strong> Materials that connect multiple machine types, acting as single points of failure</li><li><strong>PageRank:</strong> Shows which materials are <strong>most influential </strong>in the entire network of equipment dependencies, even if they don’t appear frequently</li><li><strong>Impact analysis:</strong> Centrality and downtime cost combined to prioritize stocking strategies</li></ul><p>In our dataset, we found several striking examples, like a $28 motor oil lubricant with the highest centrality, shared across 131 pieces of equipment.</p><h3>Summary</h3><p>In the end, it was never really about the machines. It was about the small things that kept them running: the lubricants, the sensors, the bolts that nobody noticed until they weren’t there. Mining companies spend billions on fleets that look like rolling fortresses, but it turns out their greatest vulnerability costs less than a tank of gas. What knowledge graphs and GenAI give us isn’t just a new way to manage data; it’s a way to see the mine for what it truly is: a network of fragile dependencies, where knowledge is as valuable as ore. Miss that, and the $23 lubricant becomes the most expensive item in the pit.</p><h3>Resources</h3><ul><li><a href="https://login.neo4j.com/u/signup/identifier?state=hKFo2SA4YXB2VkxUSjN1eWNZN25RbXZaRExfcjdteW4wUEd5WKFur3VuaXZlcnNhbC1sb2dpbqN0aWTZIExxUEYxb2V5OGU5amRVTGJ2c2RxZXJieUJyZWJmUDhFo2NpZNkgV1NMczYwNDdrT2pwVVNXODNnRFo0SnlZaElrNXpZVG8">Neo4j AuraDB</a></li><li><a href="https://neo4j.com/blog/developer/managing-risk-in-a-manufacturing-plant-with-neo4j-aura-graph-analytics/">Managing Risk in a Manufacturing Plant With Neo4j Aura Graph Analytics</a></li><li><a href="https://neo4j.com/use-cases/supply-chain-management/">Neo4j Use Case: Supply Chain</a></li><li><a href="https://neo4j.com/blog/knowledge-graph/how-to-build-knowledge-graph/">How to Build a Knowledge Graph in 7 Steps</a></li><li><a href="https://neo4j.com/generativeai/">Neo4j for GenAI</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc88d00faf43" width="1" height="1" alt=""><hr><p><a href="https://medium.com/neo4j/how-a-28-part-can-bring-down-a-5m-machine-and-how-genai-stops-it-cc88d00faf43">How a $28 Part Can Bring Down a $5M Machine, and How GenAI Stops It</a> was originally published in <a href="https://medium.com/neo4j">Neo4j Developer Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Criminals Use Whatsapp. We Use Graphs]]></title>
            <link>https://medium.com/@Desk_Rider/criminals-use-whatsapp-we-use-graphs-834e62c2d39d?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/834e62c2d39d</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[organized-crime]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 25 Jun 2025 16:55:45 GMT</pubDate>
            <atom:updated>2025-06-25T16:55:45.449Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How Legal Documents and AI Reveal the Hidden Networks Behind Crime</strong></p><p>Julián thought selling mangoes would be simple. He wasn’t laundering money, he wasn’t smuggling weapons, he wasn’t even dodging taxes. He just wanted to run a fruit stand in a quiet Lima neighborhood, where he could weigh papayas and talk soccer with regulars. Then one day, a WhatsApp message popped up on his phone, short, ugly, unmistakable: <em>“We know where you live. Pay $1,300 every two weeks or you and your family die.”</em> A bullet appeared at his door. A masked man followed. And like thousands of small business owners across Latin America, Julián found himself pulled into a violent economy where criminal organizations operate more efficiently than the local tax office, and where the price of doing business is your peace, your profit, and sometimes your life.</p><p>It is well known that organized crime causes widespread financial, social, and commercial harm worldwide. Beginning in the 1990s, small criminal groups restructured or merged, resulting in more agile <strong>networked enterprises</strong>. What was then a focus on drug trafficking has diversified into human smuggling, extortion, illegal mining and other crimes often more profitable than cocaine profits.</p><p>This <a href="https://www.reuters.com/markets/mexico-central-bank-has-growing-inflation-challenge-extortion-2024-08-27">Reuters article</a> explains how Mexico’s Central Bank not only has currency fluctuations and supply shocks to consider for its monetary policy but also the “protection money” cartels demand from local business owners resulting in an unintended consequence: inflation.</p><p>In <a href="https://www.reuters.com/markets/commodities/fuel-theft-violence-siphoning-215-million-ecuador-oil-industry-2024-12-12">this article</a> by Reuters, the author estimates the losses incurred by Ecuador’s state-run oil company, Petroecuador, to $215 million; caused by criminals tapping fuel pipelines to finance drug transportation and manufacturing.</p><h4>Organized Crime is a Graph Problem</h4><p>Cartels thrive on connections: people, companies, trade routes, assets. They don’t operate in silos, they operate in networks; and if investigators continue to work with spreadsheets, the hidden connections that enable these organizations to operate and evade justice will never surface.</p><p>What if Julián’s case wasn’t just another isolated incident buried in a police report? What if the names, places, and threats in his story could be connected to hundreds of other extortion attempts, to a network of enforcers, to the businesses being shaken down block by block across Lima, Medellín, or Culiacán? In this blog post, we explore how combining cutting-edge AI, graph technology, and data science can do exactly that: uncover patterns hidden in judicial documents, identify central figures within criminal groups, track which assets are financing operations and who’s supplying them, and pinpoint which locations are most vulnerable to each type of crime. Because the real power of these tools isn’t just in making sense of complex data — it’s in giving people like Julián a fighting chance.</p><h3>The Technology that Fights Back</h3><p>Crime thrives in the gaps between data points: names spelled differently, shell companies with no apparent owner, the same motorcycle used in five different assaults. That’s where technology comes in.</p><p>Judicial documents are messy. Names are misspelled. Dates are inconsistent. Criminal roles are vague or overlapping. Our dataset consists of 165 documents, over 121,000 pages, written by the Office of the Attorney General of Colombia. In them we found more than 160,000 unique names and 30,000 types of crime but the majority of these were combinations of the same name, for example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*0ql6D2tRpXPXqUsl0K_Qhg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/410/1*CYRM2fVfe13HyF6JTG7q2g.png" /><figcaption>Similar name and crime type combinations</figcaption></figure><p>Here’s how the right combination of Data Science, Generative AI, and Neo4j can turn scattered documents into actionable insight.</p><h4>GenAI: Reading What Humans Can’t</h4><p>We start by extracting key information from each document, one page at a time. From there, we identify and tag critical entities: people, crimes, locations, vehicles, weapons, phone numbers, and more. These entities are then connected in a graph model that captures the relationships and context within each case. Take a look below to see how these connections come to life.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j9kn_pb3z7cBEUDOUb7hSw.png" /><figcaption>Graph Model</figcaption></figure><p>This model allows us to interpret context more clearly, distinguishing between the accused and the victim, identifying which assets were used to finance a crime versus those that were targeted or impacted by it.</p><h4>Cleaning the Chaos</h4><p>One of the main tradeoffs I encountered using LLMs for entity extraction was speed versus precision. Even after multiple prompt iterations, the model struggled to reliably distinguish between names and nicknames, or to group similar crime types under a single category. Still, the ability to process thousands of pages and extract entities in just a few hours makes that compromise more than acceptable.</p><p>We focused our effort in cleaning the Person and Crime nodes. First we assigned a similarity score among names using the <a href="https://neo4j.com/docs/apoc/2025.05/overview/apoc.text/apoc.text.sorensenDiceSimilarity/">sorensenDiceSimilarity algorithm</a>. Here is a sample outcome</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/577/1*vqu4LXS7rD0pK7upQjQdIw.png" /><figcaption>SorensenDiceSimilarity Scores</figcaption></figure><p>Next, we linked nodes with a similarity score above 0.9, taking a conservative approach for this initial pass. To group these similar names, we applied a <a href="https://neo4j.com/docs/graph-data-science/current/algorithms/wcc/">weakly connected components algorithm</a>, allowing us to identify clusters and merge them into unified nodes. Here’s an example of one such cluster:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wp4JI6IMnLrs-roGmEPMCg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qXHPa4pXWmQoqdVCa3cXRQ.png" /></figure><p>After just a single iteration, we reduced the number of unique person names from 160,000 to 13,000, and consolidated 30,000 crime variations down to just 3,000 distinct types.</p><h4>Revealing Criminal Networks</h4><p>We can start by looking at a specific sentencing document and revealing the general network of people (red) investigated by crimes (brown). As expected, all people and crimes belong to the same cluster. But once we apply centrality measures we can spot influencers in the group as seen in the visualization on the right.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lvocJAwBqByhHlvcILeIPg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bzMyH5MgK5W9wM1vViP5aQ.png" /><figcaption>Visualization of people investigated by crimes (left) vs same group filtered by centrality (right)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/367/1*W4AFUQnhdqw-80Ki_tal5g.png" /><figcaption>People with highest centrality scores</figcaption></figure><p>When we apply the same process across all documents, the results shift slightly. We begin to see more clusters where individuals are linked through a variety of crime types, revealing broader patterns of association. Focusing only on the large cluster, which is only partially shown, we can again apply centrality measures to reveal the influencers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jcWzbPe5ylwVBXg9N6TOfA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qbLvAUyABdxwCy7QUpXs6w.png" /><figcaption>Clusters of people associated by same crimes (left) vs revealed influencers for the large cluster (right)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/485/1*OZe0aEBwhfhY2Klj8APsrg.png" /><figcaption>People with highest centrality scores across all documents</figcaption></figure><p>When we isolate the graph around extortion cases, a familiar pattern emerges: a small set of names with outsized influence, sitting at the center like spider queens in a web spun from threats and tribute.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Yfa8HCqzgFk5yXgP258smg.png" /><figcaption>Top influencers in extorsion crimes</figcaption></figure><p>Let’s zoom in on the extortion network and see not only who pulls the strings but <em>where</em> it’s happening most. When we layer location data onto the graph, hotspots begin to emerge: districts, neighborhoods, even specific streets where threats are more than routine, they’re systemic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/462/1*M9pfckgCq9JAitJAoyWYfQ.png" /><figcaption>Top 10 locations for extorsion related crimes</figcaption></figure><p>Geotagging opens the door to identifying spatial patterns tied to specific types of crime. By linking offenses to their locations, we can uncover geographic trends that aren’t visible in raw text alone. Here are a few examples that show how certain crimes tend to cluster in particular regions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/357/1*ef3_0e9lzqLuOIcgb-ReVw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/360/1*-hgReqSIjZbkb9-CtvC7hg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*kwWj9-cITy9E9EUkLIAa7g.png" /><figcaption>Extorsion | Kidnapping | Homicide</figcaption></figure><h4>Assets That Enable</h4><p>Every operation needs fuel. For criminal networks, that fuel isn’t ideology, it’s financing. Behind the threats, the violence, the extorsion, there’s always someone managing the resources: drugs, vehicles, properties, cattle, or cold hard cash. With this graph model we can begin to surface the individuals responsible for funding each organization’s operations. These aren’t the trigger-pullers or the street-level dealers , they’re the ones who make sure the bills get paid and the weapons get delivered. In this section, we map out who’s bankrolling the cartels and what they’re using to do it.</p><p>When we examine the assets that enable criminal activity, it’s no surprise that drugs and cash account for nearly 80% of all financing. But they’re not the only resources in play — cattle, land, and vehicles also show up across cases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1UN8hv8hv-TdqI4WiBlGIQ.png" /><figcaption>Assets used to fund criminal activities</figcaption></figure><p>The more revealing question is: <strong>who</strong> is providing these assets, and <strong>which criminal organizations</strong> are they supporting? At first glance, the financing network appears to involve thousands of individuals. But when we apply centrality measures to cut through the noise, a smaller set of key players emerges, those truly driving the financial engine behind the operations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wAXD4OfgHgA6zn637f7wOA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MVTucVC6DeGHDhf9GuJzdw.png" /><figcaption>All individuals (red) financing organizations (yellow) | Top 20 individuals with highest centrality scores</figcaption></figure><p>What we see is clear: the most influential individuals aren’t tied to just one group, they provide resources across multiple criminal organizations, acting as financial hubs in a much larger network.</p><h4>Graphs Don’t Guess, They Expose</h4><p>Organized crime thrives on complexity: multiple actors, scattered evidence, and connections hidden in plain sight. But by combining generative AI, data science, and graph technology, we can bring structure to chaos. From identifying key players and understanding how crimes are financed, to uncovering geographic hotspots and mapping criminal ecosystems, this approach transforms raw legal data into actionable intelligence. For people like Julián, that means more than insights, it means the possibility of real protection. Because the fight against organized crime isn’t just about catching criminals. It’s about seeing the network and breaking it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=834e62c2d39d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Transform Static Risk Assessment Into a Dynamic Data-Driven Strategy]]></title>
            <link>https://medium.com/neo4j/graphrag-because-your-credit-risk-model-deserves-a-brain-upgrade-fcb4f15d4334?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcb4f15d4334</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[credit-risk-analysis]]></category>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[graphrag]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 16 Apr 2025 17:10:20 GMT</pubDate>
            <atom:updated>2025-04-21T15:31:19.140Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Commercial Credit Risk Assessment Is Complex</strong></h3><p>Loan origination is a complex process that requires managers to analyze financial statements, leverage ratios, and profitability metrics. Beyond financial data, they also assess the company’s industry and business model.</p><p>These insights, along with other key factors, feed into internal risk rating models, which generate a risk rating that influences loan pricing and approval. Additional factors affecting the risk rating include industry trends and macroeconomic conditions, such as market fluctuations, regulatory risks, and the broader economic environment.</p><p>Much of this information is stored as <a href="https://neo4j.com/blog/developer/construct-knowledge-graphs-unstructured-text/">unstructured data</a>, such as annual report filings, quarterly earnings reports, balance sheets, and income statements. This not only adds to the complexity of the process but also introduces challenges like bias in data interpretation, lack of transparency in assessments, and the risk of overlooking critical factors due to human error.</p><h3>Large Language Models to the Rescue</h3><p>Large language models (LLMs) bring significant benefits across industries by automating tasks, improving decision-making, and increasing efficiency. In credit risk rating, one immediate advantage is their ability to extract vast amounts of unstructured data and organize it into a database for faster, more efficient analysis.</p><p>But why stop there? Imagine how much deeper a manager’s analysis could be if the database also integrated news and market sentiment, legal documents, and even weather events — all with real-time insights at their fingertips. This would transform risk assessment from a static process into a dynamic, data-driven strategy. <strong><em>This is the brain upgrade our model needs.</em></strong></p><h3>Graph or Vector Database?</h3><p>Graph databases and vector databases serve different purposes, and the idea of one being better than the other depends on the specific use case. However, graph databases can be considered better than pure vector databases in certain contexts, especially when dealing with complex relationships between entities. Without going into much detail, here are a few reasons why a graph database would be better in this context:</p><h4>1. Complex Relationships</h4><ul><li><strong>Graph databases</strong> are designed to represent and query complex relationships and networks.</li><li><strong>Vector databases</strong> are primarily used for storing high-dimensional vectors representing objects. While optimized for similarity search (finding items similar to a given vector), they don’t inherently capture or manage relationships between entities in a structured way.</li></ul><h4>2. Efficient Relationship Queries</h4><ul><li><strong>Graph databases</strong> are optimized for querying and navigating the relationships between entities.</li><li><strong>Vector databases</strong> don’t have the same natural support for traversing relationships in the data. Relationships are typically inferred based on similarity rather than explicitly defined, as in a graph.</li></ul><h4>3. Data Flexibility</h4><ul><li><strong>Graph databases</strong> support a variety of relationships and can evolve dynamically. You can add new types of relationships between existing entities without major restructuring.</li><li><strong>Vector databases</strong> are usually more rigid in structure. While vectors can represent complex data, modifying the way vectors are stored or interpreted often requires retraining models or reformatting the data.</li></ul><p>My colleagues <a href="https://medium.com/u/57f13c0ea39a">Tomaz Bratanic</a> and <a href="https://medium.com/u/8931e2a35281">Zachary E. Blumenfeld</a> put together two great blog posts showing how to build <a href="https://neo4j.com/blog/genai/what-is-knowledge-graph/">knowledge graphs</a> from unstructured data: <a href="https://medium.com/data-science/building-knowledge-graphs-with-llm-graph-transformer-a91045c49b59">Building Knowledge Graphs with LLM Graph Transformer</a> and <a href="https://neo4j.com/blog/news/graphrag-python-package/">GraphRAG Python Package: Accelerating GenAI With Knowledge Graphs</a>.</p><h3>The Brain Meets the Graph</h3><p>Now, let’s explore how connected data can unlock deeper intelligence. I’ll skip the details of how the graph was built, but in short, it combines both structured and unstructured data. The model integrates financial insights from 10-K annual report filings, commodities data, company profiles, and even weather events to provide a comprehensive view (kudos to my colleague <a href="https://medium.com/u/e3693e50614f">Alex Fournier</a> for helping put this together).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2kaXUmNEacc7VfDpdUqSJQ.png" /></figure><p>As shown in the graph model above, each document has been broken into chunks, with key entities extracted, such as commodities, construction materials, and elements like copper, nickel, and lithium. In addition to the raw text, each chunk includes a vector embedding that captures its semantic meaning. A full-text index has also been created based on the text properties of these chunks.</p><p>This model demonstrates its potential to answer questions that were once impossible — or extremely time-consuming — to tackle. And the best part? In many cases, we can bypass the need for LLMs altogether and simply query the graph to drive business decisions.</p><p>For example, leveraging our full-text index, we can search for terms such as <em>“</em>dependency on lithium<em>”</em> and traverse our graph to provide more context to the result:</p><pre>CALL db.index.fulltext.queryNodes(&quot;ftChunkContent&quot;, &quot;dependency on lithium&quot;) <br>YIELD node, score<br>WITH node, score<br>WITH elementId(node) as chunkId, score<br>CALL(chunkId)<br>{<br>    MATCH (c:Chunk)&lt;-[:HAS_CHUNK]-(d:Document)&lt;-[:FILED]-(co:Company)<br>    WHERE elementId(c) = chunkId<br>    return c.text as Text, co.name as Company<br>}<br>return Text, Company, score<br>order by score desc<br>limit 10</pre><p>A visualization of the results is shown below, with Tesla, Exxon, and Ford having the highest score proximity to the search term.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*qQRbyKLU_KNVnJ5i3INAIg.png" /></figure><p>And here’s the text matching our search term in Tesla’s 10-K filing:</p><p>“<em>In addition, the battery packs that we produce make use of </em><strong><em>lithium</em></strong><em>-ion cells. On rare occasions, </em><strong><em>lithium</em></strong><em>-ion cells can rapidly release the energy they contain by venting smoke and flames in a manner that can ignite nearby materials as well as other </em><strong><em>lithium</em></strong><em>-ion cells. While we have designed our battery packs to passively contain any single cell’s release of energy without spreading to neighboring cells, there can be no assurance that a field or testing failure of our …</em>”</p><p>With this model in place, we can explore the benefits of an LLM in the hands of our risk analyst.</p><h4>Retrieval-Augmented Generation</h4><p>By leveraging the vector embeddings in our chunk nodes, we can treat our model like a vector database and perform similarity searches. The <a href="https://neo4j.com/docs/neo4j-graphrag-python/current/">Neo4j GraphRAG Python package</a> abstracts the complexity of creating retrievers. For example:</p><pre>vector_retriever = VectorRetriever(<br>   neo4j_driver,<br>   index_name=&quot;chunk-embedding-index-openai&quot;,<br>   embedder=openai_embedder,<br>   return_properties=[&quot;text&quot;],<br>)</pre><p>Now we can ask:</p><p><strong>Question: </strong>What companies are exposed to risk of lithium shortages?</p><p><strong>Response:<br></strong><em>“Ford is highly dependent on its suppliers to deliver components in accordance with Ford’s production schedule and specifications, and a shortage of or inability to acquire key components or raw materials, such as lithium, cobalt, nickel, graphite, and manganese, can disrupt Ford’s production of vehicles.”</em></p><p>“… <em>availability of raw materials such as lithium, nickel, cobalt and/or other metals. The prices for these materials fluctuate and their available supply may be unstable, depending on market conditions and global demand for these materials. For example, as a result of increased global production of electric vehicles and energy storage products, suppliers of these raw materials may be unable to meet our [Tesla] volume needs. Additionally, our suppliers may not be willing or able to reliably meet our timelines or our cost and quality needs, which may require us to replace them with other sources.</em>”</p><p>Since we also have weather events, we can try incorporating that into our risk analysis. For example:</p><p><strong>Question: </strong>What companies might have weather affecting them and what are the materials at risk?</p><p><strong>Response:<br></strong>“<em>The occurrence of natural disasters or extreme weather events, such as hurricanes, tropical storms, floods, wildfires, earthquakes, tsunamis, cyclones, typhoons, extended winter storms, droughts and tornadoes; epidemics, pandemics or disease outbreaks and other extreme events and man-made disasters, such as nuclear or biological attacks or other acts of violence, such as active shooter situations, whether as a result of war or terrorism or otherwise, can have a material adverse effect on the U.S. economy in general, our industries and us specifically.</em>”</p><p>While similarity search provides decent results, it falls short when it comes to reasoning beyond individual chunk nodes to deliver more accurate answers. Let’s see how these same questions perform when using a <a href="https://neo4j.com/blog/genai/what-is-graphrag/">GraphRAG</a> retriever instead.</p><h4>GraphRAG</h4><p>This retriever combines vector search and Cypher traversals to incorporate additional nodes and relationships, which result in more accurate answers. This time, we add a parameter to our retriever code (a traversal query that guides the LLM through our graph model):</p><pre>graph_retriever = VectorCypherRetriever(<br>   neo4j_driver,<br>   index_name=&quot;chunk-embedding-index-openai&quot;,<br>   embedder=openai_embedder,<br>   result_formatter=vector_cypher_formatter, <br>   retrieval_query = traversal_query<br>)<br><br>traversal_query = &quot;&quot;&quot;<br>WITH node as chunk<br>MATCH (co:Company)-[:FILED]-&gt;(:Document)-[:HAS_CHUNK]-&gt;(chunk)-[:MENTIONS]-&gt;(n)<br>      <br>WITH COLLECT(Distinct chunk) as chunks,<br>COLLECT(Distinct co) as companies,<br>COLLECT(Distinct n) as elements<br><br> RETURN                 <br>      &#39;\n\n=== text ===\n&#39; + <br>       apoc.text.join([c in chunks | c.text], &#39;\n---\n&#39;) + <br>       &#39;\n\n=== graph ===\n&#39; + <br>       &#39;\n\n=== companies ===\n&#39; +<br>        apoc.text.join([c in companies | c.name], &#39;\n---\n&#39;) +<br>       &#39;\n\n=== elements ===\n&#39; +<br>        apoc.text.join([e in elements |labels(e)[0] + &#39;: &#39; + e.name], &#39;\n---\n&#39;)  <br>        AS info<br>&quot;&quot;&quot;<br></pre><p><strong>Question: </strong>What companies are exposed to risk of lithium shortages?</p><p><strong>Response:<br></strong><em>1. Ford: Ford is highly dependent on the supply of lithium and other raw materials for the production of electric vehicle batteries. A shortage of lithium can disrupt Ford’s production of vehicles and increase production costs.</em></p><p><em>2. Tesla: Tesla relies on lithium-ion battery cells for its vehicles and energy storage products. A reduced availability of lithium may impact Tesla’s access to battery cells and its growth, potentially reducing profitability.</em></p><p><em>3. Panasonic: As a supplier of lithium-ion battery cells, Panasonic is exposed to the risk of lithium shortages, which could affect its ability to meet the demand from companies like Tesla and other clients.</em></p><p><strong>Question: </strong>What companies might have weather affecting them and what are the materials at risk?</p><p><strong>Response:<br></strong><em>1. Amazon<br> — Weather events: Flood Warning<br> — Materials at risk: Glass<br> — Summary: Amazon’s operations could be impacted by flood warnings, which may affect their logistics and supply chain, particularly concerning the handling and transportation of glass materials.</em></p><p><em>2. CVS Health<br> — Weather events: None specified, but extreme events are mentioned.<br> — Materials at risk: Carbon (in the context of transition risks from carbon-based energy)<br> — Summary: While no specific weather events are mentioned, CVS Health is concerned about extreme events that could impact operations and increase healthcare costs. The transition from carbon-based energy also poses risks.</em></p><p><em>3. AT&amp;T<br> — Weather events: None specified, but natural disasters and extreme weather conditions are mentioned.<br> — Materials at risk: Not specified<br> — Summary: AT&amp;T’s infrastructure and operations could be disrupted by natural disasters and extreme weather conditions, which could lead to significant service disruptions and financial impacts.</em></p><p><em>These companies face potential risks from weather-related events that could disrupt their operations and supply chains, affecting their materials and overall business performance.</em></p><h3>Summary</h3><p>At first glance, GraphRAG might seem like just another tool in our GenAI toolkit. But dig deeper, and you start to see something more fundamental: a new way of assessing risk: not just by collecting data but by understanding its structure.</p><p>For our risk analyst, the challenge has never been the lack of information. It’s the fragmentation, the endless documents that bury insights under layers of irrelevant detail. GraphRAG changes that equation. It doesn’t just retrieve data; it connects the dots. It identifies patterns across seemingly unrelated disclosures, commodities reports, and even environmental events — surfacing what a traditional model might miss entirely.</p><p>And that’s the real story here — not that we’ve found a better retriever but that we’ve built a smarter brain, one that thinks a little more like an analyst and a little less like a machine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcb4f15d4334" width="1" height="1" alt=""><hr><p><a href="https://medium.com/neo4j/graphrag-because-your-credit-risk-model-deserves-a-brain-upgrade-fcb4f15d4334">Transform Static Risk Assessment Into a Dynamic Data-Driven Strategy</a> was originally published in <a href="https://medium.com/neo4j">Neo4j Developer Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Uncovering Corruption in Public Office]]></title>
            <link>https://medium.com/@Desk_Rider/uncovering-corruption-in-public-office-how-knowledge-graphs-and-data-science-are-shining-a-light-30dc73c43078?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/30dc73c43078</guid>
            <category><![CDATA[corruption]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Tue, 15 Oct 2024 16:47:21 GMT</pubDate>
            <atom:updated>2024-10-18T16:24:03.905Z</atom:updated>
            <content:encoded><![CDATA[<h3>How Knowledge Graphs and Data Science are Shining a Light on the Dark Side of Governance</h3><p>Corruption in public office has plagued societies for centuries, stifling economic growth, eroding trust in institutions, and undermining the rule of law. While traditional methods of investigating corruption — such as audits, whistleblowers, and investigative journalism — have exposed countless schemes, the sheer scale and complexity of modern corruption networks demand more sophisticated tools. Cutting-edge technologies, like knowledge graphs and data science, are revolutionizing how corruption is detected, analyzed, and ultimately, curbed.</p><h4><strong>The Complexity of Corruption Networks</strong></h4><p>Corruption in public office often operates within highly intricate networks, involving a web of individuals, shell companies, financial transactions, contracts, and sometimes even entire governments. These networks are difficult to untangle due to their obfuscation tactics, which may involve layers of intermediaries, false documentation, and offshore entities.</p><p>Given the vast amounts of data generated by government activities, such as procurement contracts, political donations, lobbying efforts, and financial disclosures, it becomes increasingly challenging for human analysts alone to pinpoint corruption schemes within these datasets. This is where data science and knowledge graphs come into play.</p><h4><strong>How Knowledge Graphs and Data Science Enhance Corruption Detection</strong></h4><p>In the context of corruption investigations, knowledge graphs allow us to visualize and analyze relationships between individuals, government officials, businesses, and financial flows. With this structure, we can track how public funds are distributed, how decision-makers are linked to contractors, and how officials may be connected to suspicious entities, both locally and globally.</p><p>By applying machine learning algorithms, data scientists can identify anomalies in financial transactions, such as sudden, unexplained wealth or repeated no-bid contracts awarded to the same entities, or the reverse, entities repeatedly bidding on contracts but never getting them. These anomalies often serve as red flags for potential corruption, signaling that a deeper investigation is warranted.</p><p>One of the major challenges in corruption investigations is identifying when different names, aliases, or organizations refer to the same individual or group. For example, a corrupt official might use variations of their name or operate through multiple companies to hide their involvement. Data science techniques such as entity resolution help identify these hidden connections.</p><h4><strong>Real World Application</strong></h4><p>In this section we are going to show real world examples of anomalies we helped a south American government agency discover within their procurement data.</p><p>The knowledge graph below represents the workflow the government has in place to award contracts and all parties involved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KSjx2Z9KukEAIzU6LEe4aw.png" /></figure><p>We begin with a government entity responsible for managing a bidding process that results in a project being adjudicated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/792/1*Pf-cU-zMKz16M8pmL89jzw.png" /><figcaption>An entity manages processes to adjudicate projects</figcaption></figure><p>These entities are related to identifiers (email, phone number, address, tax id) that help us uncover possible ties to suspicious entities, including non-government ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6lhuwF8TA6YkTDdye3P0tg.png" /><figcaption>Shared identifiers tying government and non-government entities</figcaption></figure><p>Non-government entities can have one or more roles in the process as the supplier, contractor, legal representative, or adjudicator. This structure helps us untangle the complicated network that could be designed to hide conflicts of interest where corrupt officials are handing out contracts to benefit themselves. Or known fraudsters attempting to bid on contracts using fake identity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q-Qn2cGQ2EaRRmmC3CdHDg.png" /><figcaption>Untangling conflicts of interest or fake identity in contract assignment</figcaption></figure><h4><strong>Shining a Light on the Dark Side</strong></h4><p>Now that we understand the process and the model, lets focus on a few examples.</p><ol><li><strong>Same entities bidding for multiple projects but never getting them</strong> <br>In the graph below we see entities acting as a suppliers, participating in several processes, but never actually being awarded the contract. Although this is expected, high recurrence is a sign of entities participating with the sole purpose of inflating the contract bid amount.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r823Iln_Dayd3Zg7Wou_cA.png" /><figcaption>Contract Inflators</figcaption></figure><p><strong>2. Contracts with above average value awarded by a government entity<br></strong>The objective here is not only to focus on contracts that deviate from the entity’s average value but at also identify the person and roles associated with them. The graph is the perfect tool to find patterns where the same non-government entity, perhaps with different roles, is regularly being awarded these high value contracts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PLMoUCF-gTQmYkuuC818Jg.png" /></figure><p><strong>3. Same non-government entity playing 3 or more roles<br></strong>This is the pattern we want to use to identify known fraudsters attempting to bid on contracts using fake identity. Here in our dataset we see hundreds of these patterns</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jLOQMY4MJ29ZevTsTgOYRA.png" /><figcaption>Non-government entities playing several roles</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wl55AtorxCj_j82og9hi6g.png" /><figcaption>Zooming into the potential fake identities</figcaption></figure><p>But more importantly, we can navigate the graph and visually inspect the lineage from the government entity that awarded the contracts to these potential fraudsters or those with conflict of interest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*biIO9zn99pf_PSE6D2KsrA.png" /><figcaption>Lineage between government and awarded contract</figcaption></figure><h4>Conclusion</h4><p>Uncovering corruption in public office is a complex and arduous task, but with the power of knowledge graphs and data science, we are now better equipped than ever to map out, investigate, and dismantle these networks. By making the invisible visible — by revealing the hidden connections, the offshore accounts, and the intricate webs of influence — these technologies are helping to restore integrity to public institutions and build a more just society.</p><p>As knowledge graphs and data science continue to evolve, so too will our ability to combat corruption. In this new era of data-driven governance, the truth will always have a way of coming to light.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30dc73c43078" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Spreading the Savings Like Mayo: How Knowledge Graphs Can Transform Invoice Data and Lower Costs]]></title>
            <link>https://medium.com/neo4j/spreading-the-savings-like-mayo-how-knowledge-graphs-can-transform-your-invoice-data-and-lower-6501c61de6a5?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/6501c61de6a5</guid>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[sales]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 11 Sep 2024 22:45:26 GMT</pubDate>
            <atom:updated>2024-09-17T14:39:10.379Z</atom:updated>
            <content:encoded><![CDATA[<h3>Spreading the Savings Like Mayo: How Knowledge Graphs Can Transform Your Invoice Data and Lower Costs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wzEvfqwLXLzGaXTDipWjrg.jpeg" /></figure><p>In the world of business, where every penny counts, companies are constantly looking for ways to cut costs, streamline operations, and edge out the competition. But what if I told you the secret to lowering costs wasn’t hidden in complex algorithms or revolutionary tech but instead in the humble world of … mayo?</p><p>Yes, mayo.</p><p>Much like mayo is the unsung hero of a great sandwich, knowledge graphs can be the secret ingredient to better supplier relationships and cost efficiency. Let’s dive into how you can use them to achieve these goals — and spread the savings like never before.</p><h3><strong>What Are Knowledge Graphs?</strong></h3><p>A knowledge graph is a powerful data model that organizes information in a web of relationships between entities, rather than in isolated rows and columns like traditional databases. Think of it as connecting dots between people, companies, products, and transactions in a meaningful way, allowing you to visualize and analyze how they interact.</p><p>When applied to invoice data, knowledge graphs make it easier to spot trends, detect anomalies, and understand patterns in your procurement and supply chain. Whether you’re dealing with thousands of suppliers or millions of invoices, knowledge graphs give you a bird’s-eye view of your entire operation, helping you zero in on opportunities to save money, mitigate risks, and optimize relationships.</p><p>In the case of our customer, one of the largest corporate conglomerates in Mexico, the model included entities from the procurement process (where suppliers submit their offers) and from the accounting process (the actual purchase of goods). They are responsible for 2 percent of all invoices generated in the country.</p><p>The complete model is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PNNjAeW2QR6lNtwBuRtldQ.jpeg" /></figure><p>The bottom section of the graph represents the procurement process; a supplier sends a purchase order (PO), which then generates a payment for an item. The top section represents the accounting section, revolving around the invoice; a tax ID, registered as a supplier, submits an invoice, which is then paid by another tax ID, owned by a line of business (LoB). The invoice is related to a transaction and the purchased item.</p><p>The result is the combination of two data domains in a single model that can now be used to analyze the interactions between these entities and find interesting insights.</p><h3><strong>So Where’s the Mayo?</strong></h3><p>One of our customer’s critical questions was whether there were certain high-frequency items that were being purchased by different cost centers at different prices. Blinded by the volume of data and siloed procurement teams, price differences are certainly easy to miss.</p><p>Our analysis began by ranking the top 10 suppliers by total invoiced amount. Within those, we then looked at the top 10 most purchased items, along with the lowest and highest purchase price, and total spend.</p><p>As it turns out, our customer purchased 33,000 mayonnaise containers (among other items) from a single supplier with as much as a $5 price difference. Their total spend on mayo that year: $1,509,834. Had they purchased all the mayo at the lowest price, there would’ve been savings of around $410,000.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S3Jpcv9O7Ms1jSJYJBf6cQ.png" /><figcaption>Top Five Purchased Items and Purchase Price Difference</figcaption></figure><p>For this specific scenario, a knowledge graph could have helped as an enabler of data-driven negotiations: You have the data to negotiate smarter and ensure that you’re getting the best price and terms from your suppliers.</p><h3><strong>Finishing Condiments</strong></h3><p>Now, let’s break down how knowledge graphs can help you manage your mayo (and your costs) more efficiently:</p><ol><li><strong>Negotiating better prices with suppliers — </strong>For example, identify suppliers that tend to raise prices after long-term contracts or fail to provide discounts when volumes increase.</li><li><strong>Lowering supply-chain risks — </strong>For example, visualize which suppliers are crucial to your operations, and see if there’s an over-reliance on any single one of them.</li><li><strong>Identifying high-centrality suppliers</strong> — For example, if a high-centrality supplier is high-risk or underperforming, use your knowledge graph to explore other options and diversify your supply base.</li></ol><p>So, go ahead — spread the savings like mayo, and let knowledge graphs help you make every dollar count!</p><p>To learn more about knowledge graphs, check out <a href="https://neo4j.com/blog/graphrag-manifesto/">The GraphRAG Manifesto: Adding Knowledge to GenAI</a> and <a href="https://neo4j.com/developer-blog/fine-tuning-vs-rag/">Knowledge Graphs and LLMs: Fine-Tuning vs. Retrieval-Augmented Generation</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6501c61de6a5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/neo4j/spreading-the-savings-like-mayo-how-knowledge-graphs-can-transform-your-invoice-data-and-lower-6501c61de6a5">Spreading the Savings Like Mayo: How Knowledge Graphs Can Transform Invoice Data and Lower Costs</a> was originally published in <a href="https://medium.com/neo4j">Neo4j Developer Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why doesn’t my application scale?]]></title>
            <link>https://medium.com/@Desk_Rider/why-doesnt-my-application-scale-1c5c6d039407?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c5c6d039407</guid>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Thu, 08 Mar 2018 02:04:24 GMT</pubDate>
            <atom:updated>2018-03-08T14:32:18.295Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X77ZGRK8xNcjIKkvBoYo-g.jpeg" /></figure><p>Writing applications that scale while demand increases is critical for their success in production. Whether it is an increase in user requests, computing operations, or handling data, developers must ensure their applications perform regardless of it.</p><p>While I’ve developed business applications for many years I recently acquired an interest in scientific performance of algorithm implementation. In this article I look at how the <a href="https://en.wikipedia.org/wiki/Merge_sort">mergesort </a>algorithm’s performance will vary when implemented using .NET System.Collection objects (i.e. List&lt;T&gt;) or the primitive types (i.e. Int [] ).</p><h4>Simulation</h4><p>I ran a <a href="https://en.wikipedia.org/wiki/Monte_Carlo_method">Monte Carlo Simulation</a> with sets of random data to get the average running time for each implementation. Starting with a set of 100 all the way to 10,000,000 items.</p><p>Implementation A (credit to <a href="https://www.youtube.com/watch?v=5NEQw1fhYsY">Daniel Knight</a> ) uses 2 List&lt;int&gt; objects to store the data while it is being sorted and then merged.</p><p>Here is a small snippet of the code</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/439/1*TwW5SJcOkaGQKC4ZzE211g.png" /><figcaption>Mergesort using Collections</figcaption></figure><p>Implementation B (credit to <a href="https://www.coursera.org/instructor/~250165">Robert Sedgewick</a> and <a href="https://www.coursera.org/instructor/~246867">Kevin Wayne</a> from Princeton University) uses Int[] arrays. Here is a small snippet of the code</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*HL3X2HDELm0aNxXb7oCqQw.png" /><figcaption>Mergesort using primitive types</figcaption></figure><p>While both implementations yield the same results, a sorted list, the difference in performance is evident when data increases. Have a look at this chart</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/551/1*Rx78PGCo05pqFFAQGkcUFA.png" /></figure><p>The running time for 10 MM items using implementation took so long I had to kill the process!</p><p>Now, I’ll be the first to admit that .NET does help developers a lot by providing objects that already sort, merge, resize, etc. and I often write code using its classes without knowing what’s happening behind the scenes.</p><h4>Analysis</h4><p>So why is the performance so different if the algorithm is the same? It has to do with the fact that List&lt;T&gt; is a generic version of an ArrayList, whose size is dynamically increased when needed. This is great when the size of your data is not fixed. But resizing the array comes with a cost. In Implementation A we are basically using the Add and the Remove methods. If we add enough items to the list, .NET will automatically resize it; the same happens when we remove items, .NET will shrink the array to reuse memory.</p><p>The List.Add method is a O(1) operation when the item count is less than the list’s capacity. However, if the List is resized it becomes an O(n) operation, where n is the item count. This matters when we are talking about 100K, 1MM items as you can tell by the chart above.</p><p>The List.Remove method performs a linear search, it is an O(n) operation.</p><p>On the other hand, operations using the int[ ] object are constant O(1).</p><p>So while the .NET Collection objects provide great functionality to write code, we need to understand how operations are implemented before we decide to solve problems with them.</p><p>When dealing with big volumes of data, paraphrasing Robert Sedgwick, you can spend a lot of money in hardware or a lot of time to solve a problem, or you can implement a good algorithm and be much more effective.</p><h4>References</h4><ul><li><a href="https://msdn.microsoft.com/en-us/library/cd666k3e(v=vs.100).aspx">List(T).Remove Method (T) (System.Collections.Generic)</a></li><li><a href="https://msdn.microsoft.com/en-us/library/3wcytfd1(v=vs.100).aspx">List(T).Add Method (T) (System.Collections.Generic)</a></li><li><a href="https://msdn.microsoft.com/en-us/library/system.collections.generic(v=vs.110).aspx">System.Collections.Generic Namespace</a></li><li><a href="https://msdn.microsoft.com/en-us/library/system.collections.arraylist(v=vs.110).aspx">ArrayList Class (System.Collections)</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c5c6d039407" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What happens after you run a marathon…]]></title>
            <link>https://medium.com/@Desk_Rider/what-happens-after-you-run-a-marathon-9faed390162b?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/9faed390162b</guid>
            <category><![CDATA[running]]></category>
            <category><![CDATA[marathon]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 31 Aug 2016 17:43:36 GMT</pubDate>
            <atom:updated>2016-08-31T17:43:36.039Z</atom:updated>
            <content:encoded><![CDATA[<p>You will either love it and run more or hate it and never run that distance again. I became the latter; here’s my story…</p><p>Like most self-motivated amateur athletes I wanted to run the 42.2 Km race to check it off my bucket list. This would be my first long distance race ever; the most I had ran was 4 Km. So I consulted with the wisest of my coaches: Google, “Marathon training plan”, click and done. There I had it, a 3 month step by step, day to day plan that would get me from 0 to 42.2 Km.</p><p>I ran 4 days a week and did some other activities 2 of the other days. The plan was good, I saw my endurance increase. After a few weeks I wasn’t scared of a 10K, and a few weeks later I ran my first half-marathon in under 2 hours. I thought that was pretty decent. So I stuck to the training plan.</p><p>Soon after, and within a few weeks before the marathon I had to call a friend to pick me up from an 18K training run. Severe pain on my right knee prevented me to even walk. The diagnosis was damage to the iliotibial band which required physio and rehab twice a week for at least 3 weeks. The most I had ran by this point was 30 Km, and I wouldn’t get to do another long distance run until the actual race.</p><p>The night before the marathon came and all I could think of was the injury. At the hotel room I laid out my gear, making sure I had everything I needed. I set the alarm clock, and reviewed the route to get to the start line. It seemed quite straightforward. Then after what only seemed like an instant, the alarm went off I jumped off the bed got dressed and headed out to the subway station.</p><p>Subway is not running! Of course, the subway doesn’t run at 5:00 AM. Alright, no need to panic, get a cab. Most main streets were closed off for the race but I managed to get there with some time before the start time. so first thing first, stretch that iliotibial band and warm up. Breakfast? No, no time for that; I’m sure they’ll have bars and stuff along the trail. So the race starts, and this is how every 10 Km went in my head:</p><p><strong>0 to 10 Km: </strong>Let’s run this one at a medium pace just to get the hang of it. Whatever these guys next to me are saying sounds reasonable. Oh shit, I didn’t eat anything. I guess I’ll grab a snack at the 5 Km station. <br>OMG, there are no snacks at the station… and all the energy drink is gone! Ok, water will do until the next one. Split time 0:58:00</p><p><strong>11 to 21 Km:</strong> Ok, that 10 K wasn’t that bad. Let’s try to increase the pace so I don’t get bored. I’m also on target to a sub 4 hour race. Water only, again?! Well, at least the weather is nice and cool, and my knee is holding up. Split time 0:54:00</p><p><strong>22 to 32 Km: </strong>Good job, half way there. F**k that last hill was hard, hopefully that guy in the crowd wasn’t kidding when he said it’s all downhill from here. Hmm, or did he mean I’d feel like crap from here on? I’m so hungry, and look at all these convenience stores. If I only had some money, I’d grab a slice of pizza and keep running. Ok, let’s take these 10 K at a more moderate pace to conserve energy. I wonder if anyone dropped a gel pack, even a partly used one. Holy shit I’m hungry. Split time 0:58:00</p><p><strong>32 to 42.2 Km:</strong> Alright, ten clicks to go. Still on track for target time, this is it! Every step I take is the farthest I’ve ever run. Didn’t they say there’s a wall at the 30K mark? Who are these fools? #*$@ I’m hungry. Why did I sign up for this? Are those bananas up ahead? I can’t believe it, 2 girl scouts handing out bananas. I’ll take them all, thank you!</p><p>With 4 Km to go I slowed down so much that I was pretty much walking. I had no energy left and I was bored out of my mind. Simple things like hearing my name from the spectators or reading a motivating line kept me going. First hearing and then actually seeing the finish line gave me the final boost to jog across. My total time was 4:23:39 and I have never run a marathon since.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9faed390162b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I am a terrible swimmer]]></title>
            <link>https://medium.com/@Desk_Rider/i-am-a-terrible-swimmer-4edbefc4f542?source=rss-2bfa3d9425c8------2</link>
            <guid isPermaLink="false">https://medium.com/p/4edbefc4f542</guid>
            <category><![CDATA[triathlon]]></category>
            <category><![CDATA[swimming]]></category>
            <dc:creator><![CDATA[Paolo Delano]]></dc:creator>
            <pubDate>Wed, 22 Jun 2016 15:07:57 GMT</pubDate>
            <atom:updated>2016-06-22T15:07:57.687Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*90dGhmimm_-d1kab6yYx0g.jpeg" /></figure><p>This might come as a surprise if you’ve read my previous stories. How can a triathlete be a terrible swimmer, right? The truth is I was a terrible swimmer; now I consider myself ok, which means I will complete the swim leg of the race.</p><p>The biggest problem for me was exhaustion from a lack of oxygen. I found it extremely difficult to keep a steady heart rate while holding my breath for 2 or 3 strokes. Growing up, every coach would tell me to keep my head underwater for 4 strokes and then roll to breathe. Using this technique I was unable to swim more than 50m without resting. How was this possible? How can someone run 10 or 20Km any given day, but not swim more than 50m?</p><p>During every swim session I was completely drained after the warm up. I just thought I had limited lung capacity or I needed to train harder. Neither of those made me a better swimmer. The answer was total immersion swimming. Tim Ferris talks about it in his book “<a href="http://fourhourbody.com/">The 4 hour body</a>” and also describes it in <a href="https://www.youtube.com/watch?v=1AQA-ofcdrU">this video</a>.</p><p>The biggest difference is that you are not kicking anymore. This sounds completely counterintuitive, but just think about it… your leg muscles (quads, hamstrings) are huge. The amount of oxygen needed to move them, and move them fast, will leave you breathless. There are other differences that I encourage you to review in videos.</p><p>Using this technique I was able to swim long distances without elevating my heart rate, and at the same speed! In a triathlon this is a huge advantage because it means that I’m conserving energy for the bike or run stage without compromising the first stage.</p><p>Here is a 2,100m swim routine you can use to practice total immersion and improve your lung capacity:</p><p>600m: warm up<br>500m: freestyle with pull &amp; paddles<br>400m: freestyle at race tempo<br>200m: alternate 50m backstroke/ 50m breast stroke<br>100m: kickboard<br>4x50m: breath control (breathe every 3 strokes for 25m, and every 5 strokes for 25m) rest for 10&quot;<br>100m: cool down</p><p>Happy swimming!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4edbefc4f542" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>