<?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 Esteban Sánchez on Medium]]></title>
        <description><![CDATA[Stories by Esteban Sánchez on Medium]]></description>
        <link>https://medium.com/@estesan?source=rss-856d580f0c6b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*KeCvS9sAvIyFRoEpxU9S3w.png</url>
            <title>Stories by Esteban Sánchez on Medium</title>
            <link>https://medium.com/@estesan?source=rss-856d580f0c6b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 10:16:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@estesan/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[From document to graph in five steps]]></title>
            <link>https://medium.com/@estesan/from-document-to-graph-in-five-steps-048563bdda02?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/048563bdda02</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 00:00:34 GMT</pubDate>
            <atom:updated>2026-11-08T18:43:13.879Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OqcwDcJx9Dl-r8h6" /></figure><p>Published on JournalInt</p><h3>One case, several documents</h3><p>To show how JournalInt works, we use an invented case: the Harbourline file. Every person, company and place in it is fictional, but the story is a familiar one. In 2024, the port authority of Valmora awards a 48 million euro contract to Harbourline Construction to expand the container terminal. Three weeks later, the builder hires a consultancy with no employees, Greystone Consulting. And the money ends up in a company belonging to the sister-in-law of the official who signed the award.</p><p>In our demo, the case fits in a single file. In a real investigation it would arrive in pieces: the award decision, the company registers, the bank statements, each party’s statements. Each piece tells part of the story, and none of them arrives in order.</p><p><strong>What JournalInt aims for is that getting the graph of all those pieces is a matter of reviewing, not of structuring by hand.</strong> It takes five steps.</p><h3>Step 1: create the project and upload the documents</h3><p>You create an investigation and upload the documents: PDF, DOCX, plain text or Markdown, several at once, or a web address. Scanned PDFs go through OCR first.</p><p>There is nothing to prepare beforehand: no templates, no columns, no data model.</p><h3>Step 2: the AI proposes the entities</h3><p>An AI model reads the full text and proposes the entities: people, organisations and places. Each one is highlighted in the text.</p><p>In the Harbourline file, those include people such as Victor Hale, chairman of the port authority, or Marisol Crane, his sister-in-law; companies such as Harbourline Construction, Greystone Consulting or Northgate Holdings; and places such as Valmora and Corvan.</p><h3>Step 3: you confirm the entities</h3><p>You open each entity, see where it is mentioned and confirm it. This step comes before the relationships: if a person is identified wrongly, the error would carry into all of their relationships.</p><p>This is where several documents become a single graph. If Marisol Crane appears in the company register and in the bank statements, she is one node, not two. And entities of different types are never merged: a person and a company with similar names stay separate.</p><p>Once the entities are confirmed, the relationships are extracted.</p><h3>Step 4: you review each relationship with its quote</h3><p>Each relationship arrives with the literal sentence from the document that supports it, and one click highlights it in the text. In the file, the sentence that supports Harbourline Construction paying Greystone Consulting is this one:</p><blockquote><em>“Harbourline Construction paid Greystone Consulting 2.4 million euros between July and December 2024, in six monthly payments.”</em></blockquote><p>Relationships come from normalised types rather than free text: owns, is family of, funds. JournalInt checks that the sentence is in the text before showing it. No quote, no relationship. The ones that are wrong, you correct or delete.</p><p>Want to take these steps with your own documents? <a href="https://www.journalint.com/en#early-access">Request your invitation here</a>.</p><h3>Step 5: move through the graph</h3><p>What you have reviewed becomes your project’s graph, and in it the story can be seen at a glance. The money goes from the builder to the consultancy, and from there to Northgate Holdings, Marisol Crane’s company. Marisol Crane is the sister-in-law of Victor Hale, who signed the award.</p><p>Each link in that chain is in a different sentence of the file. In the graph they form a path, because each name is a single node. Every item carries its provenance label (detected by AI, confirmed by a person or detected by a person), and every line takes you back to its sentence.</p><p><strong>That is what JournalInt wants to give you: getting from documents to a graph without structuring by hand, so that your time goes on investigating.</strong></p><h3>What the graph does not do for you</h3><p>The graph shows the path, but it does not prove it. The fact that the file says Victor Hale lives in the villa Northgate Holdings bought does not make it a fact: it is a lead to check against other sources before publishing. An extraction is a lead, not a finding.</p><p>The beta is free and by invitation, and during the beta only public or already published material may be uploaded: no leaks and no drafts. If you want to take these five steps with your own documents, you can request an invitation at <a href="https://journalint.com/">journalint.com</a>.</p><p>This text was written with the help of AI tools and under human supervision.</p><p><em>Originally published at </em><a href="https://www.journalint.com/en/blog/from-document-to-graph-in-five-steps"><em>https://www.journalint.com</em></a><em> on November 5, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=048563bdda02" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[JournalInt and Linkurious: exploring a graph or building one]]></title>
            <link>https://medium.com/@estesan/journalint-and-linkurious-exploring-a-graph-or-building-one-e9218fe20874?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9218fe20874</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 00:00:24 GMT</pubDate>
            <atom:updated>2026-11-08T18:41:52.958Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1AAQ0PTX0HceREF0" /></figure><p>Published on JournalInt</p><h3>Two tools, two moments in an investigation</h3><p>Linkurious and JournalInt both work with graphs, so it is easy to mention them in the same breath. But they answer different questions.</p><p>Linkurious, owned by Nuix since April 2026, is a platform for exploring and analysing a graph that already sits in a graph database. It is used by anti-fraud, anti-money-laundering, cybersecurity and intelligence teams, and by large journalism consortia. Its starting point is data that is already structured.</p><p><strong>JournalInt starts earlier.</strong> Its starting point is unstructured documents (PDFs, notes, articles, web pages), and its job is to turn them into that graph. It is built for people who investigate on their own: investigative and data journalists, and OSINT researchers.</p><p>In one sentence:</p><blockquote><em>Linkurious explores a graph you already have; JournalInt builds that graph from your documents, and every relationship carries the sentence that proves it.</em></blockquote><h3>The bottleneck is before the graph</h3><p>The Linkurious manual puts it plainly: before you start exploring the content of your graph database, you need to import data. Somebody has to decide what is a node and what is a relationship, and load it all in. If your data already lives in a database, that step is done. If what you have is three hundred PDFs, that step is most of the work.</p><p>It happens even in the best-equipped teams. In 2024, ICIJ explained that in past projects, creating a graph database took multiple steps and meant manually exporting data from its document search tool, Datashare, to build a Neo4j database separately. It added that such graphs are usually built during the final stages of an investigation.</p><p>The graph arrives late because building it is costly. And a graph that arrives at the end can help you check, but no longer helps you discover.</p><p><strong>That step, getting from documents to a graph, is the first thing JournalInt sets out to solve.</strong></p><h3>Which relationships each one sees</h3><p>The difference is not only where each tool starts. It also changes which relationships make it into the graph.</p><p>In Linkurious, the relationships are whatever is in your data: owns, transfers, directs. If a company register says one company owns another, that line appears.</p><p>JournalInt extracts from the text structural relationships, such as owns, works for or is family of, and also intentional ones, such as criticises, pressures, funds, accuses or allies with. These are not in any register: they only appear in what someone said or wrote, which is why they can only come from documents. Today 14 relationship types are built, 11 intentional and 3 structural, out of the 40 in the taxonomy.</p><h3>Where each one wins today</h3><p>JournalInt already does the part nobody wants to do: you upload the documents, confirm the entities and the graph gets built. But it is a beta designed for the individual investigator. It has a cap of 500 nodes per read, and the full canvas for exploring the graph has not been built yet.</p><p>That is where Linkurious is far ahead. It works at a different scale, its visualisation is mature, it is made for teamwork, it has alerts and it can be installed on your own servers (on-premise). If your newsroom already has its data in a graph database and needs all of that, Linkurious is the tool.</p><p>What JournalInt does better is answer the question of where each line comes from. In Linkurious, a relationship’s provenance is the source data. In JournalInt, it is the sentence: each relationship is stored with the literal sentence from the document that supports it, and JournalInt checks that the sentence is in the text before showing it. No quote, no relationship. In front of your editor, you can go from any line in the graph to the exact sentence that justifies it, and correct or delete the ones that do not hold up.</p><p><strong>A relationship that carries the sentence that proves it is what turns a graph into something you can defend.</strong></p><p>Want to see the difference with your own documents? <a href="https://www.journalint.com/en#early-access">Request your invitation here</a>.</p><h3>Not one or the other</h3><p>One tool builds the graph and the other explores it at scale, so they can sit side by side in the same investigation. Planned: exporting the graph in standard formats, such as Neo4j CSV, with the quote as a property of each relationship, so it can be taken into tools like Linkurious. That export does not exist yet. The aim is to make the tools work together.</p><p>If what holds you back is not exploring the graph but having one, that is the gap JournalInt wants to fill. The beta is free and by invitation, and during the beta only public or already published material may be uploaded: no leaks and no drafts. If you want to try it, you can request an invitation at <a href="https://journalint.com/">journalint.com</a>.</p><p>This text was written with the help of AI tools and under human supervision.</p><p><em>Originally published at </em><a href="https://www.journalint.com/en/blog/journalint-vs-linkurious-explore-or-build-a-graph"><em>https://www.journalint.com</em></a><em> on November 4, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9218fe20874" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop preparing, start investigating]]></title>
            <link>https://medium.com/@estesan/stop-preparing-start-investigating-db512b51c12c?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/db512b51c12c</guid>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[osint]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 00:00:04 GMT</pubDate>
            <atom:updated>2026-11-03T12:04:36.905Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Q1KRpKPeA4UPjTu4" /></figure><p>Published on JournalInt</p><h3>With the graph in front of you, everything is easier</h3><p>When you have the map of who is connected to whom in front of you, investigating gets easier. You can see which company pays which, who appears in three different documents and which name keeps turning up where it should not. The questions almost ask themselves.</p><p>The problem is getting to that map. A graph with the information you need does not appear when you open a folder of PDFs: somebody has to build it. And that somebody is usually the same person who should be investigating.</p><h3>Structuring takes time you do not have</h3><p>Before you investigate, you have to structure. Read every document, note who appears, decide whether “Hale” and “Victor Hale” are the same person, write down which sentence says that one company pays another and which document it is in. It is necessary work, but it is not investigating. It is preparing.</p><p>It is also easy to get wrong. ICIJ itself warns, in its Offshore Leaks database, that what seems to be the same name can appear as two or more different nodes. A name that is not matched is a connection that gets lost.</p><blockquote><em>In Muck Rack’s State of Journalism 2026 report, 42% of journalists say they only sometimes, rarely or never have the time to do their work the way they want to.</em></blockquote><p><strong>JournalInt exists so that you do not lose that time structuring. You upload your files and start investigating on the graph.</strong></p><h3>What JournalInt structures for you</h3><p>You upload your documents (PDF, DOCX, plain text or Markdown), several at once, or paste a web address. Scanned PDFs go through OCR first.</p><p>An AI model proposes the entities: people, organisations and places. You review them and confirm them. When the same person appears under different names across several documents, they end up as a single node, and entities of different types are never merged.</p><p>With the entities confirmed, the model proposes the relationships between them, drawn from normalised relationship types rather than free text. Each relationship is stored with the literal sentence from your document that supports it, and JournalInt checks that the sentence is in the text before showing it to you. The ones that are wrong, you correct or delete.</p><p>And the graph builds up: each new document adds to what came out of the previous ones.</p><p>Want to try it with your own documents? <a href="https://www.journalint.com/en#early-access">Request your invitation here</a>.</p><h3>What it leaves you to investigate</h3><p>What you have reviewed becomes your project’s graph, and that is where the work that matters begins. You can move through it and follow paths you would not see reading document by document: one document says A pays B, another says B is owned by C, and in the graph the path from A to C appears because B is the same node in both. Each relationship takes you to its sentence in the document, and each item carries its provenance label: detected by AI, confirmed by a person or detected by a person.</p><p>But the tool is not going to stop there: a full graph to explore is planned, with filters, search and a panel showing the evidence for each relationship; two assistants that cite, one to ask the graph and one to ask your documents; analysis of who is central and which communities there are; and cross-checking against registries and sanctions lists.</p><p><strong>In the end it comes down to two things: structuring for you, and giving you tools to discover and investigate.</strong></p><h3>Prepare less, investigate more</h3><p>An extraction is still a lead, not a finding. The quote takes you to the exact sentence in your document, but checking it against other sources and deciding what gets published is still your job. What changes is when you start doing it: not at the end, once you have finally finished structuring, but as soon as the graph is in front of you.</p><p>The beta is free and by invitation, and during the beta only public or already published material may be uploaded: no leaks and no drafts. If you want to try it with your own documents, you can request an invitation at <a href="https://journalint.com/">journalint.com</a>.</p><p>This text was written with the help of AI tools and under human supervision.</p><p><em>Originally published at </em><a href="https://www.journalint.com/en/blog/stop-preparing-start-investigating"><em>https://www.journalint.com</em></a><em> on November 3, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db512b51c12c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A graph is a lead, not a finding]]></title>
            <link>https://medium.com/journalint/a-graph-is-a-lead-not-a-finding-91dcee6e80d5?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/91dcee6e80d5</guid>
            <category><![CDATA[osint]]></category>
            <category><![CDATA[data-journalism]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 00:00:25 GMT</pubDate>
            <atom:updated>2026-10-29T10:43:16.620Z</atom:updated>
            <content:encoded><![CDATA[<p>Published on The JournalInt blog</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TsSqS6yGDN20TeBqVCmg9A.png" /></figure><h3>The graph is where the work starts</h3><p>Anyone who sees a relationship graph for the first time tends to read it as a conclusion: these people are connected, this company pays that one, here is the plot. It is an understandable reading, but an incomplete one. A graph is where an investigation starts. It shows you where to look.</p><p>The problem is that getting to that starting point costs a lot. Before you can investigate, you have to structure: read every document, note who appears and under what name, who does what to whom, and in which sentence it says so. With three documents it is an afternoon. With three hundred it is weeks that almost nobody has. And those weeks never show up in the final report: nobody reads the hours it took to find out that “the director” of one page and the full name in another are the same person.</p><blockquote><em>In Muck Rack’s State of Journalism 2026 report, 42% of journalists say they only sometimes, rarely or never have the time to do their work the way they want to.</em></blockquote><p><strong>One of the reasons JournalInt exists is to cut that part down. It does not investigate for you. It reduces the time spent structuring so that you have more time left to investigate.</strong></p><h3>What it does today</h3><p>You upload your documents (PDF, DOCX, plain text or Markdown), or paste a web address. Extraction happens in two stages, with a person in between.</p><p>First, an AI model recognises entities: people, organisations and places. Then you review them and confirm them.</p><p>After that, with the entities confirmed, the model proposes the relationships between them, drawn from normalised relationship types rather than free text. Each relationship is stored with the literal sentence from your document that supports it, and JournalInt checks that the sentence is in the text before showing it to you.</p><p>We measure what it does. In an internal evaluation on a reference set, entity extraction reached a precision of 95%, a recall of 93%, and 0% invented quotes.</p><p>Want to help us improve these numbers? <a href="https://www.journalint.com/en#early-access">Request your invitation here</a> and try JournalInt with public material.</p><h3>What the time you save is for</h3><p>If the graph arrives sooner, the work that matters starts sooner. That work comes in three kinds.</p><p>The first is cross-checking. Each relationship takes you to a specific sentence in a specific document. That sentence does not prove that what it says is true: it proves that the document says it. From there, you need to look for other sources that confirm or contradict it.</p><p>The second is following second-order relationships or longer chains. One document says A pays B. Another says B hires C. Neither mentions A and C, but in the graph the path appears, because B is the same node in both. Those chains are hard to spot when you read document by document, and that is often where the story is.</p><p>The third is seeing who the actors and the narratives that carry the most weight are: who is connected to almost everyone, who links two groups that never touch each other. Today the journalist does that reading by looking at the graph. The metrics that would calculate it automatically (centrality, communities, bridge entities).</p><p><strong>And this is the second thing JournalInt was built for: making your investigation more accurate, more agile and better informed.</strong></p><h3>A lead, not a finding</h3><p>Our website has a line that sums up how we want the product to be used: an extraction is a lead, not a finding. The quote takes you to the exact sentence in your document. Checking that claim against other sources, and deciding what gets published, is still the journalist’s job.</p><p>What JournalInt wants to take off your hands is the mechanical part: the hours of reading, noting and matching names before you can start asking questions. What remains afterwards is yours: cross-checking, following the long paths, deciding who matters and what is true.</p><p>The beta is free and by invitation, and during the beta only public or already published material may be uploaded: no leaks and no drafts. If you want to try it with your own documents, you can request an invitation at <a href="https://www.journalint.com/en?utm_source=social&amp;utm_medium=medium&amp;utm_campaign=rrss-2026-w44#early-access">journalint.com</a>.</p><p>This text was written with the help of AI tools and under human supervision.</p><p><em>Originally published at </em><a href="https://www.journalint.com/?utm_source=social&amp;utm_medium=medium&amp;utm_campaign=rrss-2026-w44"><em>https://www.journalint.com</em></a><em> on October 29, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91dcee6e80d5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/journalint/a-graph-is-a-lead-not-a-finding-91dcee6e80d5">A graph is a lead, not a finding</a> was originally published in <a href="https://medium.com/journalint">Journalint.</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I moved my dev plan to Notion connected with Claude Code]]></title>
            <link>https://medium.com/journalint/why-i-moved-my-dev-plan-to-notion-connected-with-claude-code-b2407580f456?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b2407580f456</guid>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[project-management]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Mon, 10 Aug 2026 13:54:17 GMT</pubDate>
            <atom:updated>2026-08-10T13:54:56.876Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why I moved my development plan to Notion connected with Claude Code</h3><blockquote>The project is <a href="https://www.journalint.com/">JournalInt</a>, a knowledge-graph tool for journalists and investigators.</blockquote><p>Nineteen phase files. One hundred and thirty-five tasks. Thirty-six decisions. All of it written in Markdown, and by the time I looked properly, the README alone was carrying four pointers to things that no longer existed.</p><p>So on 6 August I moved the whole plan into Notion. This is what I learnt from doing it.</p><h3>A plan in Markdown is a notebook, not a system</h3><p>Markdown is very good at prose and very bad at state.</p><p>A phase file can describe a task beautifully. What it cannot do is answer a question. Which tasks are blocked right now? What does finishing task 6.4 actually unblock? Which specifications does decision D25 govern, and are any of them already half-implemented? Every one of those questions meant opening several files and holding the answer in my own head, which is precisely the thing I was trying to stop doing.</p><h3>The connection is not the MCP server, it is the four skills</h3><p>Connecting Notion to Claude Code is a few lines in .mcp.json. That part takes minutes and changes very little on its own. You end up with an agent that can technically reach your workspace, and no shared idea of when it should.</p><p>What changed the way I work is four repository-scoped skills that make the workspace the default route rather than an option.</p><ul><li>notion-read reads a task in full before any code is written, along with its phase, because the acceptance criteria that no single task covers live there.</li><li>notion-task moves work through the pipeline and closes it with the commits and the pull request.</li><li>notion-bug handles the bug I find halfway through implementing something else: stop, write it down, then carry on with what I was doing.</li><li>notion-doc handles documentation and decisions.</li></ul><p>The part I did not expect to matter as much as it does is the authority boundary. The agent changes workflow state freely, without asking: moving a task, filling start and end dates, ticking a definition-of-done box that is genuinely met, recording commits. Content is different. If a task’s context section no longer matches the code, the instruction is to report it and wait, never to silently follow a stale specification and never to quietly rewrite the specification to match the code. Knowing which of those two things is happening is the difference between a plan I trust and a plan I have to re-read.</p><h3>Blocked and depends on are not the same thing</h3><figure><img alt="A simple image describing the depends on and the blocked of a task." src="https://cdn-images-1.medium.com/max/535/1*OkkjzlqrM1QEk-3T7FWlPQ.png" /><figcaption>One task, two kinds of edge. Depends on is ordinary sequencing. Blocked by points at an unresolved decision, and the red rollup is what stops the agent from starting.</figcaption></figure><p>This is the detail that convinced me Markdown was never going to be enough.</p><p>Depends on points at another task. It is ordinary sequencing, and setting it writes the inverse automatically, so every task can also answer what finishing it unblocks. That inverse is the foresight I did not have before: not just what is next, but what opens up when it lands.</p><p>Blocked by points at a decision, and it is a categorically different state. Nobody unblocks it by shipping code. It needs an answer from me. A rollup reads the linked decision&#39;s status onto the task, and the rule is flat: never implement a task whose blocker status is pending. Stop and ask.</p><p>In a phase file, both of those were a sentence in a paragraph, indistinguishable from each other and invisible unless you happened to read that paragraph. As two different relation types they are queryable, and the distinction survives me forgetting about it.</p><h3>What actually changed</h3><p>The headline version of this would be that I connected an AI agent to my project management tool. That is the least interesting part of it. Anyone can do that in an afternoon.</p><p>What changed is that the plan and the code stopped being two artefacts I reconcile by hand. I can specify a task with its dependencies and its blocking decision in a form that holds that structure. I can file a bug in the middle of implementing something else without derailing what I was doing. I can ask what is blocked, what a decision governs, what a completed task unlocks, and get an answer from live data rather than from my memory of a file I wrote six weeks ago.</p><p>For a project built mostly by one person with an agent, that is not administrative overhead. Milestones stay honest because status is derived rather than declared, and the next move is visible rather than reconstructed. Traction and foresight, from having management and engineering pointed at the same source.</p><blockquote><strong>Keeping the project organised used to take the hours I wanted for building it. It costs a fraction of that now, and the difference goes straight back into the code.</strong></blockquote><blockquote>This text was written with the help of AI tools.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b2407580f456" width="1" height="1" alt=""><hr><p><a href="https://medium.com/journalint/why-i-moved-my-dev-plan-to-notion-connected-with-claude-code-b2407580f456">Why I moved my dev plan to Notion connected with Claude Code</a> was originally published in <a href="https://medium.com/journalint">Journalint.</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It’s Journalint.]]></title>
            <link>https://medium.com/journalint/its-journalint-61a3739fabb1?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/61a3739fabb1</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[journalism]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 10:26:28 GMT</pubDate>
            <atom:updated>2026-06-01T10:27:55.095Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This first article covers what Journalint. is, why graphs matters in investigations and what AI can bring to journalism.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JZB3w0xl2Sv02Rqsp8uY6g.png" /><figcaption>Cover image for Journalint.</figcaption></figure><p>There is a scene that appears again and again in films about investigations. A big wall full of faces, notes, newspapers, company names and places, all connected with red thread. It is dramatic because it captures something true.</p><blockquote>Stories are rarely linear, they are networks.</blockquote><p>People influence people, companies hide behind companies, narratives travel through intermediaries, and what matters is often not a single fact but the pattern between many facts.</p><p>Journalint is being built as a tool for journalists and investigators who need to follow stories <em>effeffin</em> (efficiently, effectively and intelligently). At its core, it uses knowledge graphs and AI to help turn scattered material into structured intelligence.</p><h3>Journalism already has the data, but not the right lens</h3><p>Investigative work usually does not fail because there is no information. It fails because there is too much of it, and because it arrives in forms that are difficult to connect.</p><p>In this AI era, spent time in checking names, tracing ownwership structures or detecting intention in the comunications is outdated,</p><blockquote>in this era uncovering what is hidden with trusted information is the key.</blockquote><p>This is the problem Journalint is trying to solve. Not by replacing the journalist, but by giving the investigation a better lens.</p><h3>Why graphs matter</h3><p>A graph is not just a nice visual interface. It is a way of thinking.</p><p>This is one of the ideas that has stayed with me since I studied graphs and the propagation of misinformation during a master’s in complex systems physics. From the mathematical roots of graph theory to modern network science shaped by figures such as Albert-László Barabási, this field has given us better ways to understand how complex systems behave. That period changed the way I looked at stories.</p><blockquote>Disinformation does not spread randomly. Power does not move randomly. Corruption rarely works in a straight line. These are networked phenomena, and if we want to understand them properly, we need tools that can handle networks.</blockquote><p>That is the intellectual foundation behind Journalint.</p><p>That matters because society itself behaves like a complex system. Influence is not distributed evenly. Some actors are bridges. Some are amplifiers. Some remain invisible until you see the network around them.</p><h3>What AI changes</h3><p>We have been able to detect some forms of relationship for a long time. If one document says a company owns another, or a person is listed as a director, that is already useful structured information.</p><p>But the recent wave of AI changes the scale and the type of relationships we can work with.</p><p>Now it is much more realistic to extract entities and connections across large volumes of documents, and to go beyond simple ownership or part-whole relationships. We can start identifying patterns of influence, repetition, contradiction and narrative alignment. That opens a different way of understanding stories.</p><p>For me, this is where the paradigm really shifts. The question is no longer only, “Can we collect enough information?” The question becomes, “Can we surface the relationships that actually shape the story?”</p><p>That is why I do not see Journalint as just another productivity tool for newsrooms. I see it as an intelligence layer for investigations.</p><h3>Why now</h3><p>I have been carrying this idea for a long time but tools such as GitHub Copilot and Claude have dramatically reduced the distance between an idea and a working product. They have made it possible for one person to move faster, test concepts sooner and build with a level of ambition that would have been much harder a few years ago.</p><p>In that sense, I sometimes think of myself as the AIrquitect of Journalint. The vision is human, the questions are journalistic, the direction comes from lived curiosity, but AI makes the construction velocity real.</p><blockquote>And because of the industrialization of the missinformation, the moment feels urgent to give to people better tools not as a luxury, but as infrastructure.</blockquote><h3>What I want Journalint to become</h3><blockquote>I want Journalint to help journalists see what is hidden in plain sight.</blockquote><blockquote>I want it to help them map actors and relationships faster.</blockquote><blockquote>I want it to help them understand how a narrative evolves, who is pushing it, and where the real leverage points are.</blockquote><p>Journalint is still early, but the purpose is already clear: to bring more intelligence, more clarity and less noise into investigative journalism.</p><p>If you want to follow what comes next, stay tuned at <a href="http://www.journalint.com/"><strong>www.journalint.com</strong></a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61a3739fabb1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/journalint/its-journalint-61a3739fabb1">It’s Journalint.</a> was originally published in <a href="https://medium.com/journalint">Journalint.</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can The AI Understand A Story?]]></title>
            <link>https://medium.com/@estesan/can-the-ai-understand-a-story-705cc2601318?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/705cc2601318</guid>
            <category><![CDATA[film]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Sat, 19 Sep 2020 15:50:33 GMT</pubDate>
            <atom:updated>2020-09-19T15:50:33.312Z</atom:updated>
            <content:encoded><![CDATA[<h4>Personal Projects</h4><h4>How applying NLP techniques we can play with stories.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R72_rJFu8eEdJ8v1" /><figcaption>Photo by <a href="https://unsplash.com/@chrishardyphotography?utm_source=medium&amp;utm_medium=referral">Chris Hardy</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Overview</h3><p>In this project we are going to see how using NLP techniques we can develop a small app in which given a story, like the synopsis of a film or a video game; <strong>we can predict top 3 genres</strong>.</p><p>In addition we will play a little game to see how works the app.</p><h3>Introduction</h3><p>I would like to set clear that in this article you won’t find how works NLP algorithms.</p><blockquote>The goal of this article is to show this techniques with a less technical perspective.</blockquote><blockquote>What is NLP?</blockquote><blockquote>- Humans have dreams. Even dogs have dreams, but not you. You are just a machine. An imitation of life. Can a robot write a symphony? Can a robot turn… a canvas into a masterpiece?<br>- Could you?</blockquote><p>Previous dialogue is from <em>I, Robot</em> (2004) by Will Smith and based on Isaac Asimov books; this could summarize well what is artificial intelligence area of study where NLP is.</p><blockquote>NLP is a field of AI whose aims is that machines “understand” human language.</blockquote><p>A machine understand quite well numbers, and on a more basic level this means 0 and 1, however it does not understand well if a comment is positive or negative, for example. This is where this area has its influence, to research ways to achieve that machines understand better human language.</p><p>So, in this article we are going to see:</p><ol><li><strong>Model</strong>: The geek part of the article. Here you can see the NLP techniques and how there are important.</li><li><strong>Game</strong>: An example of the app.</li><li><strong>Conclusions</strong>: How we can use this.</li></ol><h3>Model</h3><p>All related with model and code used in this project you can find on the following links: <a href="https://github.com/esan94/GeNLP">GeNLP</a> y <a href="https://github.com/esan94/movie_genre_api">movie_genre_api</a>.</p><p>To develop a model that help us to predict the genre of a story first of all we are going to need labeled data because we are going to do a multilabel clasification, so we need a file with the synopsis of films and the genres of those movies.</p><p>Before our model learn from data we need to preprocessing text data:</p><ol><li>We need to lowercase text data due to avoid the mismatching between identical words. For example, we could have “space” in a sentence or “Space” in the beginning of a sentence, so to avoid that in following techniques this words count as different words we lowercase the text.</li><li>Next step is to delete punctuation like #, !, $, @, =, … among others. This is because these words just add noise to our model.</li><li>In this model I have decided to keep only with characters from alphabet to eliminate noise from model.</li><li>Finally I have made a <strong>lemmatization</strong>. Which is to keep the root of the word to reduce noise. This is because we can put it together words like “consider” that is a verb, “consideration” which is a noun and “considerably” that is an adverb because all of those has the same root.</li></ol><p>We have pre-processed the text data but our machine will not understand this yet.</p><blockquote>As I said before, a machine understand basically numbers, so if we want that a computer understand text we need to transform the text data into a vector of numbers.</blockquote><p>For this project I have<strong> transformed text data into a vector of word concurrence</strong>. Which means that if we found word “sunshine” 17 times we will have a vector with the number 17 which is the number of time that “sunshine” appears in the text. Also in this step we drop the “stop words” which are highly concurrence words that are meaningless for the text, determinants, linking words and abbreviations are “stop words” among others.</p><p>With this we can make an idea about how the model understand text data. So at this stage we can proceed with the training step with the aim that our model learn from data.</p><h3>Game</h3><p>Click on <a href="https://github.com/esan94/GeNLP">GeNLP</a> to be familiar with the interface of the app.</p><p>Let’s try the app with different synopsis for the same film, <em>Toy Story </em>(1995). We are going to take the synopsis from <a href="https://www.imdb.com/title/tt0114709/plotsummary">here</a>.</p><p>First synopsis will be:</p><p><em>“A little boy named Andy loves to be in his room, playing with his toys, especially his doll named “Woody”. But, what do the toys do when Andy is not with them, they come to life. Woody believes that his life (as a toy) is good. However, he must worry about Andy’s family moving, and what Woody does not know is about Andy’s birthday party. Woody does not realize that Andy’s mother gave him an action figure known as Buzz Lightyear, who does not believe that he is a toy, and quickly becomes Andy’s new favorite toy. Woody, who is now consumed with jealousy, tries to get rid of Buzz. Then, both Woody and Buzz are now lost. They must find a way to get back to Andy before he moves without them, but they will have to pass through a ruthless toy killer, Sid Phillips.” — </em>John Wiggins.</p><p>As we can read is a detailed synopsis about the movie, let’s predict the genres:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6lD6xfgfEiydF9RhqpIEgQ.png" /></figure><p>Due to the app give us the top 3 genres we can check that <strong>comedy</strong>, <strong>children</strong> and <strong>animation</strong> are 3 accurate genres for this film.</p><p>Let’s try with the following synopsis:</p><p><em>“A cowboy doll is profoundly threatened and jealous when a new spaceman figure supplants him as top toy in a boy’s room.” — </em>Kenneth Chisholm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9PkFKYIICR6gWIBciYzaZg.png" /></figure><p>In this case, the fun fact is that we can see <strong>comedy</strong> and <strong>animation</strong> as in the previous synopsis but also give us <strong>horror, </strong>a genre that nobody will match with <em>Toy Story </em>which is a kid film!</p><p>If we continue trying different synopsis we will obtain same genres or others like <strong>fantasy</strong> for example but what is happening? is something wrong with the model? or is the synopsis bad?</p><blockquote>If we take distance from the movie and read both synopsis without any judgement it is true that the second synopsis could be the description of a drama or horror film, this is because our model is trained words like “threatened” o “supplants” with horror films in this case, so this is the reason for this category.</blockquote><h3>Conclusions</h3><p>As we can see through this techniques we can make that a machine understand human language, however, there are more than we have seen here. This is a simple model that worked well with movie data but we could get a better model by using neural networks.</p><p>In 2017 <a href="https://medium.com/u/c3aeaf49d8a4">Netflix Technology Blog</a> published the following article:</p><p><a href="https://netflixtechblog.com/artwork-personalization-c589f074ad76">Artwork Personalization at Netflix</a></p><p>To sum up the article it tell us that Netflix fits the image of a serie or a movie based on the user, which means that if you like romance movies and there is a comedy and romance movie, Netflix will show you an image more romance like than comedy like.</p><p>Following this idea I will make this question:</p><blockquote>A story can be told in different ways and still be the same, i.e. a film, for example, of romance, adventure and suspense can be told in at least 3 different ways, so can the description of a story be customized according to the user, thus favoring customer acquisition?</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=705cc2601318" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[¿Puede La IA Interpretar Una Historia?]]></title>
            <link>https://medium.com/@estesan/puede-la-ia-interpretar-una-historia-25913f8227a3?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/25913f8227a3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[film]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[movies]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Sat, 19 Sep 2020 15:49:22 GMT</pubDate>
            <atom:updated>2020-09-19T15:49:22.171Z</atom:updated>
            <content:encoded><![CDATA[<h4>Proyectos Personales</h4><h4>Como aplicando NLP (Procesamiento de Lenguaje Natural) podemos jugar con historias.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4nIFlZu9m91OZD8O" /><figcaption>Photo by <a href="https://unsplash.com/@chrishardyphotography?utm_source=medium&amp;utm_medium=referral">Chris Hardy</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Resumen</h3><p>En este proyecto vamos a ver cómo usando técnicas de procesamiento de lenguaje natural podemos elaborar una pequeña aplicación en la que, dada una historia, como puede ser el resumen de una película, o un videojuego; podemos <strong>predecir los 3 géneros más acertados</strong>.</p><p>Además haremos un pequeño juego con el que podremos ver cómo funciona la aplicación y qué provecho se le podría sacar.</p><h3>Introducción</h3><p>Me gustaría empezar dejando claro que este artículo no va de explicar cómo funcionan los algoritmos de NLP (Procesamiento de Lenguaje Natural, según sus siglas en inglés) o qué hacen.</p><blockquote>El objetivo de este artículo es dar a conocer este tipo de herramientas desde una perspectiva menos técnica o más cercana a una persona que no tenga (mucha) experiencia en el uso de las mismas.</blockquote><blockquote>¿Que es el NLP?</blockquote><blockquote>- Los humanos tienen sueños. Hasta los perros tienen sueños, pero no tú. Tú eres solo una maquina. Una imitación de la vida. ¿Puede un robot escribir una sinfonía? ¿Puede un robot convertir… un lienzo en una obra maestra?<br>- ¿Podría usted?</blockquote><p>Esta frase está sacada de la película <em>Yo, Robot</em> (2004) protagonizada por Will Smith y basada en el libro de Isaac Asimov; y podría resumir muy bien que es esto de la inteligencia artificial, campo de estudio donde se encuentra el NLP.</p><blockquote>El NLP es una disciplina de la Inteligencia Artificial que busca conseguir que las máquinas (los ordenadores) “entiendan” el leguaje humano.</blockquote><p>Una máquina entiende muy bien de números y a un nivel más básico esto se traduce en 0 y 1, pero no entiende nada bien el sentido de una frase o si un comentario es positivo o negativo. Por lo que ahí aparece esta rama, para avanzar en esta investigación y conseguir que las máquinas <strong>interpreten</strong> lo mejor posible el lenguaje humano.</p><p>Una vez sabido esto en este artículo vamos a ver:</p><ol><li><strong>Modelo</strong>: O la parte más “geek” del artículo. Aquí mostraré las técnicas utilizadas y explicaré porqué son importantes.</li><li><strong>Juego</strong>: Un ejemplo de uso de la aplicación.</li><li><strong>Conclusiones</strong>: Veremos qué podemos sacar en claro de todo esto y para qué puede servir.</li></ol><h3>Modelo</h3><p>Todo lo relacionado con el modelo y el código usado para este proyecto lo podéis ver en los siguientes links: <a href="https://github.com/esan94/GeNLP">GeNLP</a> y <a href="https://github.com/esan94/movie_genre_api">movie_genre_api</a>.</p><p>Para desarrollar un modelo que nos prediga el género de una historia lo primero que vamos a necesitar son unos datos etiquetados, ya que vamos a realizar lo que en “machine learning” se conoce como una tarea de multi-etiquetado (o multilabeling en inglés). Por lo que necesitaremos un archivo que contenga los resúmenes de las películas y el género que corresponden.</p><p>Antes de hacer que nuestro modelo aprenda de los datos que les pasamos tenemos que hacer que el modelo entienda el lenguaje. Para ello vamos a realizar las siguientes tareas de estandarización o preprocesado de textos:</p><ol><li>Vamos a pasar todo el texto a minúsculas, así evitaremos tomar como palabras diferentes aquellas que son las mismas. Por ejemplo en el texto podemos tener “space” como una palabra dentro de una frase o “Space” como inicio de una frase. Para evitar que en técnicas que usaremos más adelante se cuenten como palabras diferentes, escribimos todo en minúscula y solventaríamos ese problema.</li><li>El siguiente paso es eliminar los signos de puntuación como pueden ser: #, !, $, @, =, … entre otros. Esto se hace debido a que dichos signos no añaden información al significado del lenguaje y por lo tanto los eliminamos para quitar ruido al modelo.</li><li>En este caso he optado por quedarme solo con caracteres del alfabeto, eliminando así los número con el mismo objetivo que el punto anterior. quitar ruido para luego entrenar el modelo.</li><li>Finalmente he realizado una <strong>lematización</strong>. El proceso de lematizar un texto es básicamente quedarnos con la raíz de la palabra así podemos reducir el ruido en el modelo ya que reduciríamos el número de palabras de todas aquellas que comparten la misma raíz. Por ejemplo “consider” como verbo “consideration” como nombre y “considerably” como adverbio podríamos unirlas todas en su raíz ya que la comparten.</li></ol><p>Hasta aquí habríamos pre-procesado el texto, pero todavía no hemos conseguido que un ordenador entienda texto.</p><blockquote>Como he comentado antes un ordenador entiende básicamente números, por lo que para que un texto sea interpretado por un ordenador la técnica es transformar ese texto en un vector de números.</blockquote><p>En este caso hemos optado por <strong>transformar el texto a un vector de ocurrencia de palabras</strong>, es decir que si en un texto aparece la palabra “sunshine” 17 veces en el vector tendríamos ese 17 que es el número de veces que aparece la palabra. Con este paso además se hace un eliminado de lo que son las “stop words” palabras que tienen mucha ocurrencia en un texto pero que no añaden nada de información, como pueden ser, entre otras: determinantes, conjunciones o abreviaturas.</p><p>Con esto ya nos podemos hacer una idea de cómo interpreta el texto el modelo de clasificación que estamos haciendo. Por lo que con esto ya podríamos decirle a nuestro modelo que aprenda de los datos que le hemos pasado para clasificar los géneros.</p><h3>Juego</h3><p>Click en <a href="https://github.com/esan94/GeNLP">GeNLP</a> para conocer sobre la interfaz de la aplicación.</p><p>Probemos ahora la aplicación con diferentes resúmenes de la misma película, en este caso vamos a coger la película <em>Toy Story</em> (1995) y los resúmenes los vamos a coger de <a href="https://www.imdb.com/title/tt0114709/plotsummary">aquí</a>.</p><p>El primero que vamos a probar es el siguiente:</p><p><em>“A little boy named Andy loves to be in his room, playing with his toys, especially his doll named “Woody”. But, what do the toys do when Andy is not with them, they come to life. Woody believes that his life (as a toy) is good. However, he must worry about Andy’s family moving, and what Woody does not know is about Andy’s birthday party. Woody does not realize that Andy’s mother gave him an action figure known as Buzz Lightyear, who does not believe that he is a toy, and quickly becomes Andy’s new favorite toy. Woody, who is now consumed with jealousy, tries to get rid of Buzz. Then, both Woody and Buzz are now lost. They must find a way to get back to Andy before he moves without them, but they will have to pass through a ruthless toy killer, Sid Phillips.” — </em>John Wiggins.</p><p>Como podemos leer es una descripción muy bien detallada de la película, veamos qué géneros nos predice la aplicación:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6lD6xfgfEiydF9RhqpIEgQ.png" /></figure><p>Dado que la aplicación nos devuelve el top 3 de géneros de la historia podemos comprobar cómo <strong>comedia</strong>, <strong>niños</strong> y <strong>animación</strong> son 3 géneros donde fácilmente podríamos categorizar esta película.</p><p>Probemos ahora utilizando otro resumen de la misma película, a ver que ocurre y qué predicción hace nuestra aplicación:</p><p><em>“A cowboy doll is profoundly threatened and jealous when a new spaceman figure supplants him as top toy in a boy’s room.” — </em>Kenneth Chisholm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9PkFKYIICR6gWIBciYzaZg.png" /></figure><p>En este caso podemos ver cómo dado el resumen anterior la aplicación nos devuelve <strong>comedia</strong>, <strong>animación</strong>, los cuales casan bien con la película, pero además nos devuelve <strong>horror,</strong> género que nadie en su sano juicio tildaría Toy Story como ¡una película de horror para niños!</p><p>Si siguiéramos probando con otros resúmenes, por norma general nos darían los mismos géneros que en la primera predicción o quizá podría añadir otro de <strong>fantasía</strong>. Pero, ¿qué ocurre?, ¿está mal el modelo?, ¿está mal el resumen?.</p><blockquote>Si nos abstraemos de la película y leemos ambas descripciones con una mentalidad sin juicio, si que es cierto que la segunda descripción bien podría tratarse de un drama u horror cómo se predice. Esto es debido a que el modelo tiene entrenado, por la aparición de las palabras, que palabras como “threatened” o “supplants” son palabras más comunes en películas de horror, sobre todo si no se acompaña de una explicación más detallada. De ahí que se categorice como tal.</blockquote><h3>Conclusiones</h3><p>Como se ha podido ver, mediante estas técnicas podemos hacer que una máquina interprete el lenguaje humano, sin embargo hay mucho más que lo que hemos visto, esto es un modelo muy sencillo que ha funcionado bien para los datos que tenía, pero podríamos mejorarlo añadiendo redes neuronales.</p><p>En 2017 <a href="https://medium.com/u/c3aeaf49d8a4">Netflix Technology Blog</a> publicó el siguiente artículo:</p><p><a href="https://netflixtechblog.com/artwork-personalization-c589f074ad76">Artwork Personalization at Netflix</a></p><p>En el que, en resumen, se habla de que las portadas de las series/películas se eligen en función de tus gustos, es decir si te gustan las películas románticas y una película se puede categorizar como comedia y romántica, a ti te aparecerán como portada imágenes más románticas que cómicas.</p><p>Siguiendo esta idea lanzo la pregunta:</p><blockquote>Una historia puede contarse de diferentes maneras y seguir siendo la misma. Una película, por ejemplo, de romance, aventuras y suspense puede contarse como mínimo de 3 maneras diferentes, por tanto, ¿se puede personalizar la descripción de una historia en función de los gustos, favoreciendo así la captación de clientes?</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25913f8227a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Venciendo al mercado]]></title>
            <link>https://medium.com/@estesan/venciendo-al-mercado-8e7f21fd4c76?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e7f21fd4c76</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[trading]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Tue, 18 Aug 2020 14:23:48 GMT</pubDate>
            <atom:updated>2020-08-23T09:01:50.522Z</atom:updated>
            <content:encoded><![CDATA[<h4>Proyectos Personales</h4><h4>o intentándolo.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bnj-CS-RMYRTRjcvM7Z6IA.jpeg" /><figcaption>El Palmar (Vejer de la Frontera, Cádiz) — Spain, by: Esteban Sánchez</figcaption></figure><h3>Resumen</h3><p>Este es un proyecto personal en el que he querido desarrollar una aplicación de trading haciendo uso de herramientas de machine learning. A partir de un modelado de los datos, junto con una categorización basada en distribuciones y técnicas de machine learning, he desarrollado una estrategia de trading para que, ayudado con la aplicación, un inversor novato pueda tener beneficios con poco riesgo.</p><h3><strong>Introducción</strong></h3><p>El análisis del mercado es tan interesante como complejo, en el siguiente link podemos atisbar un poco de esta complejidad <a href="https://en.wikipedia.org/wiki/Efficient-market_hypothesis">[1]</a>. Pese a ello hay muchos trabajos hechos con machine-learning para tratar de arrojar luz a este campo.</p><p>En este trabajo he creado una aplicación que tiene dos puntos principales:</p><ol><li>Una pantalla para poder analizar índices bursátiles en diferentes horizontes temporales. Aquí podremos encontrar un <strong>gráfico de velas</strong>, un gráfico para el <strong>análisis de indicadores técnicos</strong> <a href="https://www.investopedia.com/technical-analysis-4689657">[2]</a>, un gráfico de <strong>porcentaje de cambio del precio</strong> entre días y un <strong>box-plot</strong> de esto último para conocer la distribución.</li><li>Una pantalla para realizar el análisis de la estrategia de trading que he desarrollado (<em>Strategyone</em>). Esta estrategia a su vez se divide en dos partes, una de <strong>predicción del movimiento de los índices</strong> bursátiles mediante técnicas de machine learning y una segunda de <strong>comparación de los vectores de datos actuales con sus predicciones y con lo que ocurrió en el pasado</strong>. Los horizontes temporales para predecir elegidos han sido a 7, 14, 21 y 28 días.</li></ol><p>Este segundo apartado se explica en detalle en las secciones <em>Cómo vencer al mercado </em>y <em>Estrategia de trading.</em></p><p>Los datos los he obtenido de la API de Alpha Vantage <a href="https://www.alphavantage.co/">[3]</a> mientras que para obtener una lista de índices bursátiles he usado la API de Finnhub <a href="https://finnhub.io/">[4]</a>.</p><h3>Contexto</h3><p>Como físico siempre me ha fascinado el mundo de los sistemas complejos, el como ciertas fórmulas se pueden aplicar tanto a sistemas biológicos como a sistemas financieros o a conjunto de muchos electrones dando resultados interesantes en todos estos sectores.</p><blockquote>O que el estudio individual de una componente del sistema tenga un comportamiento diferente a cómo se comporta cuando se encuentra dentro del sistema.</blockquote><p>Por ello, este proyecto nace de la inquietud por conocer el mercado bursátil además del reto intelectual y de software que supone entender un sistema tan complejo como es el mercado.</p><p>El proyecto ha pasado por tres estadios. La primera versión de este trabajo fue desarrollada como trabajo final de el Máster en ciencia de datos que cursé en <a href="https://kschool.com/">[5]</a>, el objetivo aquí fue solo crear un modelo de clasificación, con técnicas de machine learning, que predijera que iba a hacer una acción en bolsa. La segunda versión, y ya a parte del Máster, fue un intento de mejorar la versión antecedente. Mientras que la tercera versión, que es la que se discute aquí, tiene además de una mejora de las anteriores versiones, el desarrollo de una estrategia de trading.</p><h3>Cómo vencer al mercado</h3><p>Con el fin de usar un <strong>modelo de clasificación</strong> para predecir los movimientos bursátiles, he tenido que categorizar los datos. Estas categorías las he denominado “Strong bull” para predicciones de subidas de precios fuertes, “Bull” para predecir subidas de precio, “Keep” para saber cuándo un precio se mantiene, “Bear” para bajadas de precio y “Strong bear” para bajadas de precios fuertes <a href="https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/">[6]</a>.</p><blockquote>¿Cómo se deciden las categorías de los índices bursátiles?</blockquote><p>Para ello he usado las <strong>distribuciones del porcentaje de cambio</strong> en el precio de la acción. Como buscamos predecir a futuro, en los registros, la columna de porcentaje de cambio tiene que tener, para cada día, la <strong>información de cómo cambia el precio de ese día con respecto al horizonte temporal sobre el que se quiere predecir</strong>. Por tanto, se comprueba</p><blockquote>cómo es el porcentaje de cambio del registro que se quiere categorizar con respecto a la distribución de los últimos 4 meses, y en función de entre qué percentiles se encuentre el valor con respecto a esa distribución, la categoría será una de las señaladas anteriormente.</blockquote><p>Con esta lógica vamos a poder categorizar todos los datos dado un horizonte temporal y siempre será sobre lo que ocurra a futuro.</p><p>Tras esta categorización el siguiente paso fue conocer cuál era la mejor manera de aplicar un algoritmo de clasificación para obtener la mayor precisión. Tras varias ideas y pruebas, el proceso elegido fue el escalar los datos mediante la técnica del <strong>escalado robusto</strong> y usar un <strong>Random Forest </strong>como algoritmo de clasificación. Esta elección se debió a que estas técnicas arrojaban una precisión mayor, sobre todas las categorías, de media.</p><p>Solo con estos pasos se logra obtener un modelo que predice “Strong bull” con una <strong>precisión de casi un 40%</strong></p><h3>Estrategia de trading</h3><p>Para la estrategia de trading nos vamos a basar en lo que ocurrió en el pasado y nos vamos a soportar sobre la idea de que acertamos siempre y cuando se gane, <strong>olvidamos ya que para ganar hay que acertar la categoría predicha</strong>.</p><blockquote>Esto es que si predecimos “Bull”, realizamos una operación en largo pero luego lo que ocurre es “Strong bull” tomaremos como un acierto la operación, lo mismo ocurre si predecimos “Strong bull” y ocurre “Bull” o si por el contrario predecimos “Strong bear”, realizamos una operación en corto y sale “bear” y, por ende a la inversa.</blockquote><p>Consideraremos una operación fallida si no ocurre nada de lo descrito anteriormente.</p><p>Con esto en mente, <strong><em>la estrategia va a ser solo realizar operaciones en largo y cuando nuestro modelo prediga “Strong bull”</em></strong>, dado que es la categoría que en el modelo de clasificación mayor precisión arrojó.</p><blockquote>¿Cómo es la estrategia?</blockquote><p>Una vez tenemos en todos los registros aplicado el robust scaler, predicha la categoría y la categorización real, <strong>aplicamos un PCA</strong> para reducir el número de dimensiones a 4 manteniendo más de un 95% la variabilidad de los datos. Con esto tenemos ahora 4 variables más la predicción asociada al registro y la categoría real de este. Así que <strong>podemos conocer cómo son las variables cuando se predice algo con respecto a la realidad de la categoría,</strong> por lo que agrupamos por predicción y categoría y calculamos la mediana asociada a cada perfil de curva para saber cómo describir cada una.</p><blockquote>Es decir, vamos a tener descritas las variables de que se predijera “Strong bull” y realmente ocurrió “Strong bull” u otra categoría.</blockquote><p>Todo ello limitado a los últimos 6 meses de datos con respecto al día de predicción, para así evitar que un estado del mercado más antiguo influya en la estrategia. En este punto obtenemos lo siguiente:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*FS4d75FM1q-DbHvTLclt5g.png" /><figcaption>Descripción de las variables para cada par predicción-clase después del PCA.</figcaption></figure><p>Esto lo que quiere decir es que en los último 6 meses antes de la predicción cuando se predijo “Strong bull” y realmente se acertó en la categoría, las variables de las componentes principales tenían esa pinta de mediana.</p><p>Por lo tanto para poder realizar una operación, tenemos que aplicar a los datos del día en el que estamos prediciendo un robust scaler y un PCA,</p><blockquote>si la predicción que hemos obtenido ha sido <strong>“Strong bull”</strong> habríamos pasado el primer paso para realizar la operación. El segundo paso es comprobar a qué perfil de las anteriores curvas se parecen los datos sobre los que estás prediciendo, para ello, <strong>usamos la semejanza del coseno</strong> con el fin de observar a qué vector se parecen más los datos. Si al que más se parece es a <strong>“strong bull-strong bull”</strong> nos estaría dando la segunda llave para realizar una operación más segura.</blockquote><p><strong>Con esta estrategia de trading vamos a obtener casi un 50% de acierto</strong>, pero como hemos dicho al principio, acertar no es solo acertar la categoría.</p><blockquote>acertar no es solo acertar la categoría.</blockquote><p>Acertar es también, en nuestro caso, predecir “Strong bull” y que realmente sea “Bull”. <strong>Cuando se tiene esto en cuenta la precisión de la estrategia aumenta hasta casi a un 58%.</strong></p><h3>Conclusión</h3><p>El objetivo de este trabajo era realizar una estrategia que permita a un pequeño inversor realizar operaciones con el menor riesgo posible al fallo, para evitar al máximo las pérdidas. Como he expuesto, la estrategia obtiene casi un 58% de acierto en las condiciones descritas, a mi juicio esto no es una estrategia recomendable para realizarla de manera automática puesto que asumes que vas a fallar casi un 40%.</p><p>Sin embargo es interesante ver cómo siguiendo una estrategia basada solo en datos, con el mínimo conocimiento bursátil, ya se obtiene más de un 50% de precisión en las operaciones realizadas.</p><p>Todo el código del proyecto se puede leer en: <a href="https://github.com/esan94/bsm03">GitHub/esan94/bsm03</a></p><h3>Siguientes pasos</h3><p>Siguientes pasos posibles para mejorar esto podrían pasar por:</p><ul><li>Cambiar el modelo de datos.</li><li>Mejorar el algoritmo de clasificación.</li><li>Añadir conocimiento bursátil al proyecto.</li><li>Asignar pesos a las componentes principales para aplicar la semejanza del coseno.</li></ul><h3>Recursos</h3><ul><li>[1] <a href="https://en.wikipedia.org/wiki/Efficient-market_hypothesis">https://en.wikipedia.org/wiki/Efficient-market_hypothesis</a></li><li>[2] <a href="https://www.investopedia.com/technical-analysis-4689657">https://www.investopedia.com/technical-analysis-4689657</a></li><li>[3] <a href="https://www.alphavantage.co/">https://www.alphavantage.co/</a></li><li>[4] <a href="https://finnhub.io/">https://finnhub.io/</a></li><li>[5] <a href="https://kschool.com/">https://kschool.com/</a></li><li>[6] <a href="https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/">https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/</a></li></ul><p>Podéis seguir mis proyectos en <a href="https://www.linkedin.com/in/estebanmsg/">LinkedIn</a>, <a href="https://github.com/esan94">GitHub</a> o <a href="https://medium.com/@emsg94">Medium</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e7f21fd4c76" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beating Stock Market]]></title>
            <link>https://medium.com/data-science/beating-stock-market-8b33c5afb633?source=rss-856d580f0c6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b33c5afb633</guid>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[trading]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[Esteban Sánchez]]></dc:creator>
            <pubDate>Tue, 18 Aug 2020 14:21:07 GMT</pubDate>
            <atom:updated>2020-08-23T15:26:58.292Z</atom:updated>
            <content:encoded><![CDATA[<h4>Personal Proyects</h4><h4>or trying to.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bnj-CS-RMYRTRjcvM7Z6IA.jpeg" /><figcaption>El Palmar (Vejer de la Frontera, Cádiz) — Spain, by: Esteban Sánchez</figcaption></figure><p><strong><em>Note from Towards Data Science’s editors:</em></strong><em> While we allow independent authors to publish articles in accordance with our </em><a href="https://towardsdatascience.com/questions-96667b06af5"><em>rules and guidelines</em></a><em>, we do not endorse each author’s contribution. You should not rely on an author’s works without seeking professional advice. See our </em><a href="https://towardsdatascience.com/readers-terms-b5d780a700a4"><em>Reader Terms</em></a><em> for details.</em></p><h3>Summary</h3><p>This is a personal project in which I have tried to develop a trading application using machine learning tools. Starting with data modelling along with a categorisation based on distribution and machine learning techniques, I have developed a trading strategy for beginner investors to generate low-risk profit with the help of this application.</p><h3>Introduction</h3><p>The market analysis is both interesting and complex as it can be seen in the following link <a href="https://en.wikipedia.org/wiki/Efficient-market_hypothesis">[1]</a>. Nevertheless, there are several works carried out with machine-learning which try to shed light on this field.</p><p>In this piece of work, I have created an application consisting of two main points:</p><ol><li>A screen where stock market index may be analysed over different temporal horizons. Here it can be found a <strong>candlestick chart</strong>; a chart to <strong>analyse technical indicators</strong> <a href="https://www.investopedia.com/technical-analysis-4689657">[2]</a>; a line chart which shows the percentage <strong>of price change between days</strong>, as well as a <strong>box-plot</strong> representing this last chart in order to understand that distribution.</li><li>A screen where the analysis of the trading strategy which I have developed (<em>Strategyone</em>) can be done. This strategy is divided into two different parts: the first one consisting of the <strong>prediction of stock market index</strong> movements by means of machine learning, whereas the second one involves the <strong>comparison between the current data vectors prediction and what happened in the past</strong>. The chosen temporal horizons range from 7, 14, 21 to 28 days.</li></ol><p>This last section is explained thoroughly in “How to beat the market” and “Trading strategy”</p><p>Data has been obtained through the Alpha Vantage API <a href="https://www.alphavantage.co/">[3]</a>, while a list of the stock market index from the Finnhub API <a href="https://finnhub.io/">[4]</a>.</p><h3>Context</h3><p>As a physicist I have been always fascinated by the complex systems world: how certain formulae can be applied to and have interesting results either to biological systems or financial ones, as well as to a group of several electrons.</p><blockquote>Likewise, how the individual study of an element of the system might result into a different behaviour when it is studied within the system.</blockquote><p>Consequently, this project emerges from the curiosity about the stock market in addition to the software and intellectual challenge that implies to understand such a complex system as the market is.</p><p>The project has gone through three stages: the first version of this work was developed as the final thesis of the Master’s Degree in Data Science which I attended in <a href="https://kschool.com/">[5]</a>, and whose aim was only the creation of classification model which could predict the future of an stock in the market using machine learning. The second version was designed externally to the Master’s and it tried to improve the first one. Finally, the third version is the one here discussed, and it offers a significant improvement, the development of a trading strategy.</p><h3>How to beat the market</h3><p>In order to use a <strong>classification model</strong> to predict market movements, I needed to categorise the data. These prediction categories have been called “Strong bull”, predictions in which the price increase is significant; “Bull”, when there is a price increase; “Keep”, the price remains the same; “Bear”, a decrease on the price, and “Strong bear”, the price decrease is significant <a href="https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/">[6]</a>.</p><blockquote>How are the stock market index categories chosen?</blockquote><p>This have done through the <strong>distribution of percentage variation in the stock price</strong>. As our aim is predicting the future, in the registers, the percentage variation column needs the <strong>daily information about how the price varies in relation to the temporal horizon that we want to predict</strong>.</p><blockquote>Therefore, the variation percentage to be categorised is compared to the last 4-month distribution, and one of the categories abovementioned will be selected based on the range of the percentiles in relation to that distribution.</blockquote><p>In this way, we could categorise all the data given a temporal horizon, and this will always be about the future.</p><p>Once the categorisation is done, the next step was getting to know which the best way to apply an algorithm of classification with more precision is. After a number of trials and different ideas, the selected process was scaling the data by means of the <strong>robust scaler</strong> technique and <strong>Random Forest</strong> as classification algorithm. These were the chosen ones since they provide an average higher precision upon all the categories.</p><p>Only following these steps, we can obtain a model which is able to predict “Strong bull” <strong>with a 40 % level of accuracy.</strong></p><h3>Trading Strategy</h3><p>The trading strategy will be based on what happened in the past and on the idea that we guess correctly provided that we win, <strong>omitting that in order to win we must also guess the right predicted category</strong>.</p><blockquote>That is, if the prediction is “Bull”, we carry out a long position operation and the resulting outcome is actually “Strong bull”, our prediction will be considered as accurate. Likewise, if we predict “Strong bull” and the result is “Bull” or when the prediction is “Strong bear”, we carry out a short position movement and the outcome achieved is “Bear” and the other way round.</blockquote><p>If none of the abovementioned cases take place, the operation will be considered as a fail.</p><p>Having this in mind, <strong><em>the strategy will only consist of long position operation and when the model predicts “Strong bull”</em></strong> given that it is the category with higher accuracy from the classification model.</p><blockquote>How does the strategy work?</blockquote><p>Once the robust scaler is applied to all the registers, the category is predicted and the actual categorisation, a <strong>PCA is applied</strong> to reduce the number of dimensions to 4 maintaining the 95 % of data variability. Therefore, we have other 4 variables together with the prediction linked to the register and its actual category. <strong>How the variables are can be known when something is predicted in relation to the real category</strong>, so we arrange the prediction and the category, and we calculate the median associated to each profile curve to understand how to describe each one.</p><blockquote>As a result, we will have described the variables in which “Strong bull” is predicted” and the actual outcome was “Strong bull” or any other category.</blockquote><p>All of this will be limited to the last 6-month-data in relation to the prediction day in order to avoid the influence of an old market state on the strategy. The results obtained are summarised below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*FS4d75FM1q-DbHvTLclt5g.png" /><figcaption>Description of the variables for each prediction-category after the PCA.</figcaption></figure><p>The interpretation of this table is that in the last 6 months before the prediction of “Strong bull” and the category was guess correctly, the variables of the main components had as the median.</p><p>Consequently, in order to carry out a operation, we must apply the data of the day in which we are doing the prediction a robust scaler and a PCA,</p><blockquote>If the prediction obtained is “Strong bull”, we will have reached the first step to carry out the operation. The second step is checking which profile of the previous curves is more similar to the data that is being predicted. This will be done using the cosine similarity which will allow us to observe the more similar vector to the data. If it corresponds to “Strong bull-strong bull”, we will have the key to perform a safer operation.</blockquote><p><strong>Following this trading strategy, we will obtain almost a 50 % level of accuracy</strong>, but, as it was mentioned at the beginning, guessing correctly does not imply guessing the category too.</p><blockquote>Guessing correctly does not imply guessing the category too</blockquote><p>Under our circumstances, a correct guessing will be also the prediction of “Strong bull” and obtaining “Bull” as a final result. <strong>The strategy level of accuracy reaches 58 % when this is taken into account.</strong></p><h3>Conclusion</h3><p>The aim of this piece of work was the development of a strategy which allows a beginner investor to carry out to generate low-risk profit without suffering a total loss. As I have mentioned, the strategy ensures a 58 % level of accuracy under the described conditions, but, on a personal note, it is not a strategy to be implemented automatically because the error level assumed raises up to 40%.</p><p>However, it is interesting to see how a level of accuracy over 50 % is obtained in the performed operations, following a strategy based only on data and with a limited and minimal knowledge about the stock market.</p><p>All the project code can be read on: <a href="https://github.com/esan94/bsm03">GitHub/esan94/bsm03</a></p><h3>Following Steps</h3><p>The possible next steps to improve might include:</p><ul><li>The change of the data model.</li><li>The improvement of the classification algorithm.</li><li>The addition to the project of more knowledge about the stock market.</li><li>The assignation of value to the main components to apply the cosine similarity.</li></ul><h3>Resources</h3><ul><li>[1] <a href="https://en.wikipedia.org/wiki/Efficient-market_hypothesis">https://en.wikipedia.org/wiki/Efficient-market_hypothesis</a></li><li>[2] <a href="https://www.investopedia.com/technical-analysis-4689657">https://www.investopedia.com/technical-analysis-4689657</a></li><li>[3] <a href="https://www.alphavantage.co/">https://www.alphavantage.co/</a></li><li>[4] <a href="https://finnhub.io/">https://finnhub.io/</a></li><li>[5] <a href="https://kschool.com/">https://kschool.com/</a></li><li>[6] <a href="https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/">https://www.investopedia.com/insights/digging-deeper-bull-and-bear-markets/</a></li></ul><p>You can follow me on <a href="https://www.linkedin.com/in/estebanmsg/">LinkedIn</a>, <a href="https://github.com/esan94">GitHub</a> o <a href="https://medium.com/@emsg94">Medium</a>.</p><p>Translation made by <a href="https://www.linkedin.com/in/paloma-s%C3%A1nchez-narv%C3%A1ez-420517109/">Paloma Sánchez Narváez</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b33c5afb633" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/beating-stock-market-8b33c5afb633">Beating Stock Market</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>