<?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 Shivani Ponturu on Medium]]></title>
        <description><![CDATA[Stories by Shivani Ponturu on Medium]]></description>
        <link>https://medium.com/@shivaniponturu?source=rss-a6a0e93c8b63------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*L_CS0aY-TELr4ff-vh12Xg.png</url>
            <title>Stories by Shivani Ponturu on Medium</title>
            <link>https://medium.com/@shivaniponturu?source=rss-a6a0e93c8b63------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 07:19:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shivaniponturu/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[The machines NASA left behind have one last job]]></title>
            <link>https://medium.com/@shivaniponturu/the-machines-nasa-left-behind-have-one-last-job-816a7c1f1c8e?source=rss-a6a0e93c8b63------2</link>
            <guid isPermaLink="false">https://medium.com/p/816a7c1f1c8e</guid>
            <category><![CDATA[nasa]]></category>
            <category><![CDATA[edtech]]></category>
            <category><![CDATA[space-exploration]]></category>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[hackathons]]></category>
            <dc:creator><![CDATA[Shivani Ponturu]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 08:34:53 GMT</pubDate>
            <atom:updated>2026-11-16T08:34:53.612Z</atom:updated>
            <content:encoded><![CDATA[<h4>How I built Pass It On for NASA Space Apps 2026: six machines that tell kids their own stories, films built from NASA records, and an AI that answers questions with a NASA source.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ie9F_6a9k6lHhQcW.jpg" /><figcaption>The start page: the Moon and Mars side by side, with Quill the rookie astronaut</figcaption></figure><p>Since the 1960s, NASA has left landers, rovers, science stations and even a car on the Moon and Mars. Each one did its job and stayed behind. Kids today mostly hear about the missions flying now, like Artemis and Perseverance, and almost never about the machines that made those missions possible.</p><p>If you want to meet the older machines, you have to read the archive. NASA’s NSSDCA (the NASA Space Science Data Coordinated Archive) keeps a record for every spacecraft: launch date, mass, landing site, last contact. The records are accurate and complete, and they have no story. The NASA Space Apps Challenge 2026 challenge <em>Abandoned but not Forgotten</em> asked for a way to tell those stories. Pass It On is my answer.</p><h3>The idea: let the machines speak</h3><p>Pass It On is a website for kids aged 9 to 13. Quill, a rookie astronaut about to fly for the first time, visits six machines: three on the Moon and three on Mars. Each machine tells its own story in a three-minute film. At any moment a child can say “Hey Quill” (or type) to stop the film and ask a question. The machine answers in its own voice, shows the NASA page the answer came from, and carries on.</p><p>At the end of each film the machine passes on one lesson. Once Quill has all six, Quill launches toward the next frontier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kc1A2lu6rbvOsdYn.jpg" /><figcaption>The map: three machines on the Moon and three on Mars</figcaption></figure><p>The six machines are:</p><ul><li><strong>Surveyor 3</strong>, a robot lander that has sat in a small crater in the Moon’s Ocean of Storms since April 1967.</li><li><strong>The Apollo 12 science station</strong> (ALSEP), left by the astronauts in November 1969 to keep doing science after they went home.</li><li><strong>The Apollo 15 Lunar Roving Vehicle</strong>, the moon buggy, parked at the foot of the Moon’s Apennine mountains since 1971.</li><li><strong>Viking 1 and Viking 2</strong>, twin landers that reached Mars in 1976 and asked whether Mars was ever home to life.</li><li><strong>Sojourner</strong>, the very first rover to drive on Mars, in 1997.</li><li><strong>Spirit and Opportunity</strong>, twin robot geologists that landed on Mars in 2004.</li></ul><h3>Turning records into stories</h3><p>Everything starts in one facts file. For each machine it holds its dates, landing site, mass, instruments, parts, builders, cost, what it found and how its mission ended. Every fact carries a link to the NASA page it came from. I began with the seven NASA links on the challenge page and added NASA fact sheets, NASA History, NASA Science and Space Place pages.</p><p>From those facts I wrote 99 short story lines. 90 of them are spoken by the machines, and each of those shows its NASA source on screen while it plays. The other 9 belong to Quill, who asks the questions a kid would ask, like “But how could you tell what a rock was made of, just by driving up to it?”</p><p>Every film follows the same eight chapters: hello, why NASA sent me, who built me, my journey, my parts, how I worked, what I found, and why I stayed. Kids quickly learn what comes next, and it is easy to compare one machine with another.</p><p>The machines speak as “I”, and numbers become things a kid can picture. Sojourner says it weighs 10.5 kilograms, “about as much as a small dog”, and is “about the size of a microwave oven”. Its landing gets the drama it deserves: “Our landing was wild! Giant airbags puffed up around my lander.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kKGRpakpHs_mlfbB.jpg" /><figcaption>The real Pathfinder lander on Mars with its airbags deflated, after bouncing to a stop. Photo: NASA/JPL</figcaption></figure><p>Each film is a cartoon drawn with code, with a real NASA photo on a board beside it. I used 32 NASA photos, looked at each one before adding it, and credited every one. The painted Moon and Mars backgrounds were made once with Google Gemini, and the site labels them as AI illustration.</p><h3>One question, start to finish</h3><p>Here is a real question on the live site. During Sojourner’s “my parts” chapter, I asked: “Why is Mars red?”</p><ol><li>The browser’s Web Speech API (built-in speech-to-text) turned my voice into text, and the film paused.</li><li>The site sent the question to a small server function, along with which machine was on screen and the line I had just heard.</li><li>The function gave the AI a pack of notes: Sojourner’s whole NASA record, seven checked facts about Earth, the Moon and Mars from NASA’s fact sheets, and a numbered list of sources.</li><li>Qwen3 235B, running on Amazon Bedrock, wrote an answer of at most four short sentences, ending with the number of the source it used.</li><li>The function turned that number into a link, and ElevenLabs read the answer aloud in Sojourner’s voice.</li></ol><p>Sojourner explained that Mars looks red because its surface is covered in rusty dust: iron in the rocks reacted with oxygen, “just like an old bike left in the rain”. Under the answer: Source: NSSDCA, a link to NASA’s Mars fact sheet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jWFuefh2Mi63d5VO.jpg" /><figcaption>Sojourner’s page: the film paused, the answer, and its NASA source</figcaption></figure><h3>Keeping the answers true</h3><p>An AI talking to children has to be careful with facts. These rules do most of the work:</p><ul><li><strong>About itself, NASA facts only.</strong> If the facts don’t cover a question about the machine or its mission, it says it isn’t sure, then shares one true, related fact. Facts I couldn’t confirm on a NASA page are marked, and the server removes them before the AI ever sees them.</li><li><strong>NASA numbers for the worlds.</strong> For the size, gravity, temperature and distance of Earth, the Moon and Mars, it may only use the numbers from NASA’s fact sheets.</li><li><strong>Comparisons stay exact.</strong> If the notes say “as heavy as a small dog”, it can never say “a big dog”.</li><li><strong>A link only when it’s earned.</strong> General questions, like how rockets work, are answered from general knowledge, and those answers show no NASA link.</li><li><strong>Low temperature.</strong> The model runs at 0.2, which makes answers steadier and less creative.</li><li><strong>A plan B.</strong> If the AI can’t be reached, the site answers from the facts file on its own.</li></ul><p>There are no accounts and no tracking, and children’s questions are never stored. The API keys stay on the server, and each visitor can only ask so many questions a minute. The one thing to know is that Chrome’s speech-to-text sends audio to Google, which is how the browser feature works.</p><h3>Listening without a button</h3><p>Kids shouldn’t have to find a button in the middle of a story, so the site listens for “Hey Quill”. That turned out to be harder than it sounds. Chrome often hears “Quill” as “will”, “kill” or “quail”, so the site accepts the common mishearings.</p><p>The films caused the second problem. A narrator saying “Hello, rookie!” sounds a lot like a wake word. While a film or an answer is playing, the site only accepts a clear “Hey Quill”, and it ignores anything that mostly repeats the words coming out of the speakers. Across 204 spoken lines, it never went off by mistake.</p><h3>How well it works</h3><p>I tested the live site on 15 November 2026 with eight questions: one per machine, plus two tricky ones.</p><ul><li><strong>8 of 8</strong> got an answer with a NASA source.</li><li>The average answer took <strong>1.1 seconds</strong> (0.7 to 1.9 seconds).</li><li><strong>7 of 8</strong> were fully correct when I checked them against the NASA facts. One mixed the Moon’s night temperature into an answer about Mars.</li><li><strong>0</strong> false “Hey Quill” alarms in 204 spoken lines.</li></ul><p>I haven’t tested it with children yet. That is the first next step.</p><h3>What I learned</h3><ul><li><strong>An AI needs numbers to hold on to.</strong> Early on, the model made up its own comparisons. Giving it seven checked NASA facts about Earth, the Moon and Mars, each with a source number, fixed most of that. It can still mix up worlds, so the next step is to check every number in an answer against the facts in code before it is spoken.</li><li><strong>NASA pages sometimes disagree.</strong> One NASA page gives Pathfinder’s landing speed as 18 metres per second and another as 14. Dates depend on time zones too: Spirit landed on 3 January 2004 in US time, which was 4 January in UTC. My rule: use UTC dates as NSSDCA gives them, pick one source per line, and show it on screen.</li><li><strong>Make it once, and use AI only when it’s needed.</strong> The films and their voices were recorded once, with the timing of every word for the captions. The AI runs only when a child asks something. The site stays fast, costs little to run, and the stories still play if the AI is down.</li></ul><h3>Tech stack</h3><ul><li>Website: React 19, TypeScript, Vite 8 and Tailwind CSS 4</li><li>Server: two Vercel Functions, one for answers and one for speech</li><li>AI answers: Qwen3 235B on Amazon Bedrock</li><li>Voices: ElevenLabs (Multilingual v2 for the films, with word timings for captions; Flash v2.5 for live answers)</li><li>Listening: the browser’s Web Speech API</li><li>Backgrounds: Google Gemini image generation, made once</li><li>Data: NSSDCA records and fact sheets, NASA Science, NASA History and other NASA pages</li><li>Hosting: Vercel</li></ul><h3>Try it</h3><p>The live site is at <a href="https://passiton-nasa.duckdns.org/">Pass It On</a>, with no login. Voice works best in Chrome or Edge, and typing works everywhere. The code is on <a href="https://github.com/SHIVANI-PONTURU/Pass-It-On-">GitHub</a>. Here is the 30-second demo:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FjCOas2XO0cM%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DjCOas2XO0cM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FjCOas2XO0cM%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f50979a3d526a50aca9fbf17731bbd3d/href">https://medium.com/media/f50979a3d526a50aca9fbf17731bbd3d/href</a></iframe><p>I built Pass It On solo, as Team Lunars, for NASA Space Apps Challenge 2026, in the challenge Abandoned but not Forgotten: Storytelling about NASA&#39;s Discarded Equipment on the Moon and Mars. Photos: NASA and NASA/JPL. NASA does not endorse this project.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=816a7c1f1c8e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Freeze the stolen money, not the whole account]]></title>
            <link>https://medium.com/@shivaniponturu/freeze-the-stolen-money-not-the-whole-account-6840c6656db1?source=rss-a6a0e93c8b63------2</link>
            <guid isPermaLink="false">https://medium.com/p/6840c6656db1</guid>
            <category><![CDATA[anti-money-laundering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[falkordb]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[hackathons]]></category>
            <dc:creator><![CDATA[Shivani Ponturu]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 07:10:16 GMT</pubDate>
            <atom:updated>2026-10-18T07:46:58.786Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How I built TraceHold for Graph Hacks 2026: a FalkorDB graph, three sandboxed AI agents, and a person who signs off every hold.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5JksWRxO3lZc7ueb3SMeRQ.png" /><figcaption>The TraceHold dashboard shows where every rupee a victim disputed ended up</figcaption></figure><p>When someone in India loses money to a scam, they report it on the 1930 helpline or the National Cyber Crime Reporting Portal. By then the money has usually moved through several mule accounts (accounts used to pass stolen money on), mixed with other people’s money on the way. Banks often respond by freezing every account the money touched. That stops the money, and it also locks innocent people out of all their savings over a small disputed sum.</p><p>India’s rules are moving towards holding only the disputed amount. The Ministry of Home Affairs’ procedure of 2 January 2026 limits holds to the disputed sum, and the Reserve Bank of India has a draft to freeze only the disputed amount from 1 April 2027. The tools in use today can flag a mule account, but they cannot say how much of that account’s balance is the victim’s money. TraceHold works out that amount.</p><h4>The idea: hold only the victim’s share</h4><p>TraceHold tracks stolen funds step-by-step through every transfer, calculating how much of each account’s balance belongs to the victim. That amount is the “taint.” Stolen funds are split proportionally: if 30% of an account’s total balance is stolen money, then 30% of every outbound transfer is considered stolen money too.</p><p>For example, imagine an account has ₹2,00,000 in original funds and receives ₹1,00,000 in stolen money. Since stolen funds make up one-third of the new ₹3,00,000 balance, one-third of any outgoing payment is treated as stolen. If the account pays out ₹90,000, ₹30,000 of that is stolen money moving to the next account, while ₹70,000 stays behind. TraceHold places a hold on that remaining ₹70,000, leaves the account holder free to use their own ₹1,40,000, and keeps tracking the ₹30,000 wherever it goes next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VNZUwA2KzkE1WIO0jBNtEg.png" /><figcaption>Pro-rata taint. The lien holds only the victim’s share.</figcaption></figure><h4>Why a graph</h4><p>Money moving between accounts is already a graph. Accounts are the points and transfers are the links between them. Tracing stolen money means walking those links forward in time, layer after layer, which is the job a graph database (a database that stores points and links and follows them quickly) is built for.</p><p>TraceHold keeps the whole bank network in FalkorDB: 518,581 accounts, 5,078,345 transfers, 30,470 banks, 166,207 customers, and the devices, phones and IP addresses they use. The data is IBM’s public synthetic anti-money-laundering dataset with a synthetic Indian layer on top (rupee amounts, account ages, devices, phones and fraud complaints). There is no real customer data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JtQqipviR2-s9nfyGRrDWA.png" /><figcaption>The network graph. Each bank also gets a memory graph of its own.</figcaption></figure><p>The hard part is time. Money can only move on after it arrives, so every hop must use a transfer that happened after the one that brought the money in. In Cypher (FalkorDB’s query language), one time-respecting hop looks like this:</p><pre>UNWIND $ids AS tid<br>MATCH (v:Account)-[f:TRANSFERRED {id: tid}]-&gt;(m:Account) WITH f, m<br>MATCH (m)-[n:TRANSFERRED]-&gt;(x:Account) WITH f, n WHERE n.ts &gt;= f.ts<br>RETURN count(n)</pre><p>The agents also use three graph algorithms that run inside FalkorDB. Label propagation and WCC (weakly connected components) both find groups of accounts tied together, which is how TraceHold spots mule communities that share devices, phones or IP addresses. Betweenness (how many paths pass through an account) finds the hub accounts in a community.</p><h4>Three agents and a person</h4><p>A complaint goes through three AI agents, all running Claude Opus 4.6 on Amazon Bedrock, and then a person.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*03ioOgnitW5xz8vR076tJw.png" /><figcaption>Every lien passes a person before it reaches a bank</figcaption></figure><ul><li>The <strong>Tracer </strong>follows the money from the victim’s disputed transfer until it is held, cashed out, or too small to follow.</li><li>The <strong>Adjudicator </strong>judges each traced account as likely mule, likely genuine or uncertain. It reads signals such as how fast money passes through, the account’s age, shared devices and phones, mule communities, and what the bank decided in earlier cases. Then it proposes a lien or a review.</li><li>The <strong>Verifier </strong>re-checks every proposed lien against the evidence and approves or rejects it with a reason.</li><li>At the HITL gate (human in the loop), a person approves or overrides each lien. Nothing reaches a bank before that.</li></ul><p>Three rules keep the agents in line. They never write database queries: They call fixed tools over MCP (Model Context Protocol, a standard way for an AI agent to call tools), and each tool runs a fixed Cypher query. Every amount comes from code, never from the model. And seven deterministic checks (rules in code, not AI) run on every proposal before it can reach the gate.</p><h4>One complaint, start to finish</h4><p>Here is a real run on the live demo. In complaint CMP-0011 at Finland Bank #0, a victim disputes ₹8,67,196.01 that left their account in one fraudulent transfer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HwsXUX2gzdQhojaw2I6OTg.png" /><figcaption>The trail, layer by layer. Each card is an account the stolen money reached</figcaption></figure><p>The three agents finished in about a minute, with 8 model calls costing about 23 US cents. The books balance to the paisa: Of the ₹8,67,196.01, ₹5,60,646.17 is still held on traced accounts, ₹2,82,171.90 was cashed out or left the bank, and ₹24,377.94 sits on branches too small to follow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ax2aDJXkZY9mpfafo3Cu1w.png" /><figcaption>The case at a glance: the disputed amount, the agents’ progress and the books.</figcaption></figure><p>TraceHold proposed two liens, each exactly the victim’s money still in the account: ₹5,47,021.57 on an account holding ₹2.55 crore, so the lien holds 2.1% of the balance, and ₹11,212.35 on a large business account. Four accounts holding small amounts of the victim’s money were labelled uncertain and sent to review, with document requests to their banks. The Verifier approved both liens and explained why, and I approved them at the gate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LyROb8LBKYAyB8BTED_uiw.png" /><figcaption>The HITL gate: the Verifier’s reasons, and two lien requests sent</figcaption></figure><p>Freezing those two accounts outright would have locked up more than ₹1,200 crore; the liens hold ₹5,58,233.92.</p><h4>Watching FalkorDB do the work</h4><p>Every step of a run is recorded, so the app can replay it: which agent ran, which tools it called, and the Cypher those tools sent to FalkorDB, while the trail grows layer by layer. A live feed shows the queries FalkorDB is running at that moment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7EsaLz3au-iyIpQqCrh8pw.png" /><figcaption>Replaying CMP-0011: the Adjudicator gathering signals, with the query behind them</figcaption></figure><h4>Memory: each bank learns from its own decisions</h4><p>Each bank keeps its own memory graph in FalkorDB, with its cases, decisions, liens and lessons. FalkorDB cannot join across graphs, so memory nodes point at network accounts by their ID. When a later case touches the same account, device or phone, the agents see what the bank decided before.</p><p>When a person rejects a lien with a reason, or overturns the Verifier, that becomes a lesson. Its confidence goes up 0.1 each time a later decision agrees and down 0.2 each time one disagrees, and below 0.3 it stops being used. At China Bank #6, a lesson taught at the gate on CMP-0283 rose from 0.70 to 0.80 when a review in a later case agreed with it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x2vKzgmAnx97kaZKzDk0ng.png" /><figcaption>A bank’s memory: one lesson, and how its confidence changed</figcaption></figure><h4>Keeping the agents in a box</h4><p>Agents that can freeze money need tight limits. Each agent runs in its own NVIDIA OpenShell sandbox (a locked-down container) with a deny-by-default policy. The Tracer, for example, can reach its own tools, the handoff routes it needs, the model and the tracing service. Everything else is blocked, including FalkorDB itself, the bank, the other agents’ tools, the gate and the internet. The model key never enters a sandbox. OpenShell’s proxy adds it to requests on the way out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wlYbpPeg9EjLfdue9JZpvQ.png" /><figcaption>What the Tracer’s sandbox can and cannot reach</figcaption></figure><h4>How it compares</h4><p>I compared TraceHold with three simpler ways of freezing money on 36 test complaints: freezing every account the money touched, a greedy trace (holding up to the disputed amount at every hop), and fixed rules with no AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CjDb9CfPmM48LpLgc5sUJg.png" /><figcaption>TraceHold against the alternatives on 36 test complaints</figcaption></figure><p>TraceHold froze ₹1.85 crore of innocent money (money in accounts the dataset labels genuine). A full freeze froze ₹15,725 crore, the greedy trace ₹215.7 crore and the fixed rules ₹76.4 crore. TraceHold also held more stolen money than the rules, ₹309.8 crore against ₹265.6 crore.</p><p>It caught fewer mule accounts than the rules, though: 59% against 70%. There were two causes. Accounts holding a large share of the loss were sometimes only sent to review, and an early “absorbed” stop could cut off mules further down the trail. I changed both rules and tested the new version on 41 different complaints, so it was not scored on the complaints that shaped it. In that set one complaint dominated the totals: an account the dataset labels genuine held ₹5,085 crore of stolen money, and every method froze it. Leaving thatrat complaint out, TraceHold froze ₹2.48 crore of innocent money against ₹2.98 crore for the rules and ₹2,404.7 crore for a full freeze, and caught 57% of mules against 63%.</p><h4>What I learned</h4><ul><li>Where a filter sits in a Cypher query can change its speed ten thousand times over. Written the obvious way, one time-respecting hop took 77 seconds for a complaint’s 16 fraud transfers. A WITH (which passes results from one part of a query to the next) between the two MATCH parts brought it to 6.5 seconds, and moving the time filter behind a second WITH, so FalkorDB walks only that account’s own transfers, brought it to 2 milliseconds. On FalkorDB 6.0.2, two MATCH … WHERE parts with no WITH between them even returned no rows at all, so every query in TraceHold uses the WITH shape.</li><li>The model’s judgement needs a code check behind it. In an early run the Tracer stopped at an account it called “absorbed” (the stolen money stayed there) when 99.9% of that money had in fact left. Now code checks every “absorbed” stop: it stands only if the account still holds at least half of the stolen money it received and passed on less than a tenth of the disputed amount. Otherwise code follows the branch and records that it overrode the model.</li><li>Keep money in whole paise. Every amount is an integer number of paise, and every run checks that the disputed amount equals what is held, plus what was cashed out, plus what was not followed. A rounding slip shows up at once.</li></ul><h4>Tech stack</h4><ul><li><strong>Graph database</strong>: FalkorDB 6.0.2</li><li><strong>AI model</strong>: Claude Opus 4.6 on Amazon Bedrock</li><li><strong>Agents</strong>: LangGraph, with tools served over MCP</li><li><strong>Sandboxes</strong>: NVIDIA OpenShell on Podman (containers)</li><li><strong>Backend</strong>: Python 3.13, Starlette and FastAPI</li><li><strong>Web app</strong>: React 19, Vite, TypeScript, Tailwind CSS, React Flow and Recharts</li><li><strong>Tracing</strong>: Langfuse (records each agent step and model call)</li><li><strong>Hosting</strong>: one AWS EC2 server (4 CPUs, 16 GB of memory) behind Caddy (HTTPS)</li></ul><p>It runs on EC2 because the full graph needs about 6 GB of memory and the sandboxes need Linux, more than my laptop could give. The server stays on through the judging period.</p><h4>Try it</h4><p>The live demo is at <a href="https://tracehold.duckdns.org/">Tracehold</a>, with no login. Pick bank 11, open Investigations, start an investigation on CMP-0011, and approve or override the liens yourself when the agents finish. “New here?” in the header gives a guided tour. The code is at <a href="https://github.com/SHIVANI-PONTURU/Tracehold">Github</a>.</p><p>I built TraceHold for Graph Hacks 2026, in Track 1 (Agents That Act on Connected Data) and Track 2 (Agent Memory and Coordination).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6840c6656db1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>