1<!DOCTYPE html> 2<html lang="en" dir="ltr"><head><title>Papers2Dataset - AI agent that browses scientific papers and extract structured information</title><meta charset="utf-8"/><link rel="preload" href="../index-47d9ba67.css" as="style"/><link rel="preload" href="../prescript-2bfc6315.js" as="script"/><style type="text/css">@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:400;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin/wght/normal.woff2);unicode-range:U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:400;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin-ext/wght/normal.woff2);unicode-range:U+0100-02BA,U+02BD-02C5,U+02C7-02CC,U+02CE-02D7,U+02DD-02FF,U+0304,U+0308,U+0329,U+1D00-1DBF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:700;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin-ext/wght/normal.woff2);unicode-range:U+0100-02BA,U+02BD-02C5,U+02C7-02CC,U+02CE-02D7,U+02DD-02FF,U+0304,U+0308,U+0329,U+1D00-1DBF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:700;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin/wght/normal.woff2);unicode-range:U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD;font-display:swap;}</style><link rel="preconnect" href="https://cdnjs.cloudflare.com" crossorigin="anonymous"/><meta name="viewport" content="width=device-width, initial-scale=1.0"/><meta name="og:site_name" content="Dmitrii's personal blog"/><meta property="og:title" content="Papers2Dataset - AI agent that browses scientific papers and extract structured information"/><meta property="og:type" content="website"/><meta name="twitter:card" content="summary_large_image"/><meta name="twitter:title" content="Papers2Dataset - AI agent that browses scientific papers and extract structured information"/><meta name="twitter:description" content="I built a tool that uses AI agents to walk the citation graph and extract structured information."/><meta property="og:description" content="I built a tool that uses AI agents to walk the citation graph and extract structured information."/><meta property="og:image:alt" content="I built a tool that uses AI agents to walk the citation graph and extract structured information."/><meta property="twitter:domain" content="eamag.me"/><meta property="og:url" content="https://eamag.me/2025/papers-to-datasets"/><meta property="twitter:url" content="https://eamag.me/2025/papers-to-datasets"/><link rel="icon" href="../static/icon.png"/><meta name="description" content="I built a tool that uses AI agents to walk the citation graph and extract structured information."/><meta name="generator" content="Quartz"/><link href="../index-47d9ba67.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-735924e0.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-589c1cea.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-bbc3f7c0.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-2911f5d1.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-e427079e.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-53b25b2b.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-c93c6a44.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-8a06014f.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-77d6a441.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-6ee21cee.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-1ea6ad18.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-34bdfded.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-caf7c009.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-788c9ca3.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-642859cd.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-4abc06bd.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-274a3dfe.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-94cb6c84.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-58d5ff56.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-7262b5a7.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../static/resource-style-a5c806d2.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../static/resource-style-6bd36952.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" rel="stylesheet" type="text/css" data-persist="true"/>
2<script src="../prescript-2bfc6315.js" type="3af50e427fb83ca387164fec-application/javascript" data-persist="true"></script>
2<script type="3af50e427fb83ca387164fec-application/javascript" data-persist="true">const fetchData = fetch("../static/contentIndex.json").then(data => data.json())</script>
2<meta property="og:image:width" content="1200"/><meta property="og:image:height" content="630"/><meta property="og:image" content="https://eamag.me/2025/papers-to-datasets-og-image.webp"/><meta property="og:image:url" content="https://eamag.me/2025/papers-to-datasets-og-image.webp"/><meta name="twitter:image" content="https://eamag.me/2025/papers-to-datasets-og-image.webp"/><meta property="og:image:type" content="image/webp"/>
2<script data-cfasync="false" nonce="5051a6fe-ca91-4484-b1a8-f065c51914b0">try{(function(w,d){!function(j,k,l,m){if(j.zaraz)console.error("zaraz is loaded twice");else{j[l]=j[l]||{};j[l].executed=[];j.zaraz={deferred:[],listeners:[]};j.zaraz._v="20";j.zaraz._n="5051a6fe-ca91-4484-b1a8-f065c51914b0";j.zaraz.q=[];j.zaraz._f=function(n){return async function(){var o=Array.prototype.slice.call(arguments);j.zaraz.q.push({m:n,a:o})}};for(const p of["track","set","debug"])j.zaraz[p]=j.zaraz._f(p);j.zaraz.init=()=>{var q=k.getElementsByTagName(m)[0],r=k.createElement(m),s=k.getElementsByTagName("title")[0];s&&(j[l].t=k.getElementsByTagName("title")[0].text);j[l].x=Math.random();j[l].w=j.screen.width;j[l].h=j.screen.height;j[l].j=j.innerHeight;j[l].e=j.innerWidth;j[l].l=j.location.href;j[l].r=k.referrer;j[l].k=j.screen.colorDepth;j[l].n=k.characterSet;j[l].o=(new Date).getTimezoneOffset();if(j.dataLayer)for(const t of Object.entries(Object.entries(dataLayer).reduce((u,v)=>({...u[1],...v[1]}),{})))zaraz.set(t[0],t[1],{scope:"page"});j[l].q=[];for(;j.zaraz.q.length;){const w=j.zaraz.q.shift();j[l].q.push(w)}r.defer=!0;for(const x of[localStorage,sessionStorage])Object.keys(x||{}).filter(z=>z.startsWith("_zaraz_")).forEach(y=>{try{j[l]["z_"+y.slice(7)]=JSON.parse(x.getItem(y))}catch{j[l]["z_"+y.slice(7)]=x.getItem(y)}});r.referrerPolicy="origin";r.src="/cdn-cgi/zaraz/s.js?z="+btoa(encodeURIComponent(JSON.stringify(j[l])));q.parentNode.insertBefore(r,q)};["complete","interactive"].includes(k.readyState)?zaraz.init():j.addEventListener("DOMContentLoaded",zaraz.init)}}(w,d,"zarazData","script");window.zaraz._p=async cc=>new Promise(cd=>{if(cc){cc.e&&cc.e.forEach(ce=>{try{const cf=d.querySelector("script[nonce]"),cg=cf?.nonce||cf?.getAttribute("nonce"),ch=d.createElement("script");cg&&(ch.nonce=cg);ch.innerHTML=ce;ch.onload=()=>{d.head.removeChild(ch)};d.head.appendChild(ch)}catch(ci){console.error(`Error executing script: ${ce}\n`,ci)}});Promise.allSettled((cc.f||[]).map(cj=>fetch(cj[0],cj[1])))}cd()});zaraz._p({"e":["(function(w,d){})(window,document)"]});})(window,document)}catch(e){throw fetch("/cdn-cgi/zaraz/t"),e;};</script>
2</head><body data-slug="2025/papers-to-datasets" data-basepath><div id="quartz-root" class="page" data-frame="default"><div id="quartz-body"><div class="left sidebar"><h2 class="page-title"><a href="..">Dmitrii's personal blog</a></h2><div class="mobile-only"><div class="spacer"></div></div><div class="flex-component" style="flex-direction: row; flex-wrap: nowrap; gap: 0.5rem;"><div style="flex-grow: 1; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><div class="search"><button class="search-button" aria-label="Search" aria-expanded="false"><svg role="img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 19.9 19.7"><title>Search</title><g class="search-path" fill="none"><path stroke-linecap="square" d="M18.5 18.3l-5.4-5.4"></path><circle cx="8" cy="8" r="7"></circle></g></svg><p>Search</p></button><div class="search-container"><div class="search-space"><input autocomplete="off" class="search-bar" name="search" type="text" aria-label="Search for something..." placeholder="Search for something..."/><div class="search-layout" data-preview="true" data-field-priority="["title","content","tags"]"></div></div></div></div></div><div style="flex-grow: 0; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><button class="darkmode" aria-label="Dark mode"><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="dayIcon" x="0px" y="0px" viewBox="0 0 35 35" style="enable-background:new 0 0 35 35" xml:space="preserve" aria-label="Dark mode"><title>Dark mode</title><path d="M6,17.5C6,16.672,5.328,16,4.5,16h-3C0.672,16,0,16.672,0,17.5 S0.672,19,1.5,19h3C5.328,19,6,18.328,6,17.5z M7.5,26c-0.414,0-0.789,0.168-1.061,0.439l-2,2C4.168,28.711,4,29.086,4,29.5 C4,30.328,4.671,31,5.5,31c0.414,0,0.789-0.168,1.06-0.44l2-2C8.832,28.289,9,27.914,9,27.5C9,26.672,8.329,26,7.5,26z M17.5,6 C18.329,6,19,5.328,19,4.5v-3C19,0.672,18.329,0,17.5,0S16,0.672,16,1.5v3C16,5.328,16.671,6,17.5,6z M27.5,9 c0.414,0,0.789-0.168,1.06-0.439l2-2C30.832,6.289,31,5.914,31,5.5C31,4.672,30.329,4,29.5,4c-0.414,0-0.789,0.168-1.061,0.44 l-2,2C26.168,6.711,26,7.086,26,7.5C26,8.328,26.671,9,27.5,9z M6.439,8.561C6.711,8.832,7.086,9,7.5,9C8.328,9,9,8.328,9,7.5 c0-0.414-0.168-0.789-0.439-1.061l-2-2C6.289,4.168,5.914,4,5.5,4C4.672,4,4,4.672,4,5.5c0,0.414,0.168,0.789,0.439,1.06 L6.439,8.561z M33.5,16h-3c-0.828,0-1.5,0.672-1.5,1.5s0.672,1.5,1.5,1.5h3c0.828,0,1.5-0.672,1.5-1.5S34.328,16,33.5,16z M28.561,26.439C28.289,26.168,27.914,26,27.5,26c-0.828,0-1.5,0.672-1.5,1.5c0,0.414,0.168,0.789,0.439,1.06l2,2 C28.711,30.832,29.086,31,29.5,31c0.828,0,1.5-0.672,1.5-1.5c0-0.414-0.168-0.789-0.439-1.061L28.561,26.439z M17.5,29 c-0.829,0-1.5,0.672-1.5,1.5v3c0,0.828,0.671,1.5,1.5,1.5s1.5-0.672,1.5-1.5v-3C19,29.672,18.329,29,17.5,29z M17.5,7 C11.71,7,7,11.71,7,17.5S11.71,28,17.5,28S28,23.29,28,17.5S23.29,7,17.5,7z M17.5,25c-4.136,0-7.5-3.364-7.5-7.5 c0-4.136,3.364-7.5,7.5-7.5c4.136,0,7.5,3.364,7.5,7.5C25,21.636,21.636,25,17.5,25z"></path></svg><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="nightIcon" x="0px" y="0px" viewBox="0 0 100 100" style="enable-background:new 0 0 100 100" xml:space="preserve" aria-label="Light mode"><title>Light mode</title><path d="M96.76,66.458c-0.853-0.852-2.15-1.064-3.23-0.534c-6.063,2.991-12.858,4.571-19.655,4.571 C62.022,70.495,50.88,65.88,42.5,57.5C29.043,44.043,25.658,23.536,34.076,6.47c0.532-1.08,0.318-2.379-0.534-3.23 c-0.851-0.852-2.15-1.064-3.23-0.534c-4.918,2.427-9.375,5.619-13.246,9.491c-9.447,9.447-14.65,22.008-14.65,35.369 c0,13.36,5.203,25.921,14.65,35.368s22.008,14.65,35.368,14.65c13.361,0,25.921-5.203,35.369-14.65 c3.872-3.871,7.064-8.328,9.491-13.246C97.826,68.608,97.611,67.309,96.76,66.458z"></path></svg></button></div><div style="flex-grow: 0; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><button class="readermode" aria-label="Reader mode"><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="readerIcon" fill="currentColor" stroke="currentColor" stroke-width="0.2" stroke-linecap="round" stroke-linejoin="round" width="64px" height="64px" viewBox="0 0 24 24" aria-label="Reader mode"><title>Reader mode</title><g transform="translate(-1.8, -1.8) scale(1.15, 1.2)"><path d="M8.9891247,2.5 C10.1384702,2.5 11.2209868,2.96705384 12.0049645,3.76669482 C12.7883914,2.96705384 13.8709081,2.5 15.0202536,2.5 L18.7549359,2.5 C19.1691495,2.5 19.5049359,2.83578644 19.5049359,3.25 L19.5046891,4.004 L21.2546891,4.0045
27396 C21.6343849,4.00457396 21.9481801,4.28672784 21.9978425,4.6528034 L22.0046891,4.75457396 L22.0046891,20.25 C22.0046891,20.6296958 21.7225353,20.943491 21.3564597,20.9931534 L21.2546891,21 L2.75468914,21 C2.37499337,21 2.06119817,20.7178461 2.01153575,20.3517706 L2.00468914,20.25 L2.00468914,4.75457396 C2.00468914,4.37487819 2.28684302,4.061083 2.65291858,4.01142057 L2.75468914,4.00457396 L4.50368914,4.004 L4.50444233,3.25 C4.50444233,2.87030423 4.78659621,2.55650904 5.15267177,2.50684662 L5.25444233,2.5 L8.9891247,2.5 Z M4.50368914,5.504 L3.50468914,5.504 L3.50468914,19.5 L10.9478955,19.4998273 C10.4513189,18.9207296 9.73864328,18.5588115 8.96709342,18.5065584 L8.77307039,18.5 L5.25444233,18.5 C4.87474657,18.5 4.56095137,18.2178461 4.51128895,17.8517706 L4.50444233,17.75 L4.50368914,5.504 Z M19.5049359,17.75 C19.5049359,18.1642136 19.1691495,18.5 18.7549359,18.5 L15.2363079,18.5 C14.3910149,18.5 13.5994408,18.8724714 13.0614828,19.4998273 L20.5046891,19.5 L20.5046891,5.504 L19.5046891,5.504 L19.5049359,17.75 Z M18.0059359,3.999 L15.0202536,4 L14.8259077,4.00692283 C13.9889509,4.06666544 13.2254227,4.50975805 12.7549359,5.212 L12.7549359,17.777 L12.7782651,17.7601316 C13.4923805,17.2719483 14.3447024,17 15.2363079,17 L18.0059359,16.999 L18.0056891,4.798 L18.0033792,4.75457396 L18.0056891,4.71 L18.0059359,3.999 Z M8.9891247,4 L6.00368914,3.999 L6.00599909,4.75457396 L6.00599909,4.75457396 L6.00368914,4.783 L6.00368914,16.999 L8.77307039,17 C9.57551536,17 10.3461406,17.2202781 11.0128313,17.6202194 L11.2536891,17.776 L11.2536891,5.211 C10.8200889,4.56369974 10.1361548,4.13636104 9.37521067,4.02745763 L9.18347055,4.00692283 L8.9891247,4 Z"></path></g></svg></button></div></div><div class="explorer nav-files-container" data-behavior="link" data-collapsed="collapsed" data-savestate="true" data-data-fns="{"order":["filter","map","sort"],"sortFn":"(a2, b2) => {\n if (!a2.isFolder && !b2.isFolder || a2.isFolder && b2.isFolder) {\n return (a2.displayName || \"\").localeCompare(b2.displayName || \"\", void 0, {\n numeric: true,\n sensitivity: \"base\"\n });\n }\n if (!a2.isFolder && b2.isFolder) {\n return 1;\n }\n return -1;\n }","filterFn":"(node) => node.slugSegment !== \"tags\"","mapFn":"(node) => {\n return node;\n }"}"><button type="button" class="explorer-toggle mobile-explorer hide-until-loaded" data-mobile="true" aria-controls="explorer-39" aria-label="Explorer"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide-menu"><line x1="4" x2="20" y1="12" y2="12"></line><line x1="4" x2="20" y1="6" y2="6"></line><line x1="4" x2="20" y1="18" y2="18"></line></svg></button><button type="button" class="title-button explorer-toggle desktop-explorer" data-mobile="false" aria-expanded="true"><h2>Explorer</h2><svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="5 8 14 8" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="fold"><polyline points="6 9 12 15 18 9"></polyline></svg></button><div id="explorer-39" class="explorer-content" aria-expanded="false" role="group"><ul class="explorer-ul overflow" id="list-0"><li class="overflow-end"></li></ul></div><template id="template-file"><li><a href="#" class="nav-file-title tree-item-self"></a></li></template><template id="template-folder"><li><div class="folder-container nav-folder-title tree-item-self"><svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="5 8 14 8" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="folder-icon nav-folder-collapse-indicator collapse-icon"><polyline points="6 9 12 15 18 9"></polyline></svg><div><button class="folder-button"><span class="folder-title"></span></button></div></div><div class="folder-outer"><ul class="content tree-item-children"></ul></div></li></template></div></div><div class="center"><div class="page-header"><div class="popover-hint"><nav class="breadcrumb-container" aria-label="breadcrumbs"><div class="breadcrumb-element"><a href="../">Home</a><p> ⯠</p></div><div class="breadcrumb-element"><a href="../2025/">2025</a><p> ⯠</p></div><div class="breadcrumb-element"><a href>Papers2Dataset - AI agent that browses scientific papers and extract structured information</a></div></nav><h1 class="article-title">Papers2Dataset - AI agent that browses scientific papers and extract structured information</h1>
2<details class="note-properties metadata-container" open data-collapsed="false"><summary class="note-properties-header"><span class="note-properties-title">Properties</span><span class="note-properties-count">1</span></summary><table class="note-properties-table"><tbody><tr class="note-properties-row metadata-property"><td class="note-properties-key metadata-property-key">tags</td><td class="note-properties-value metadata-property-value"><span class="note-properties-tags"><a href="../tags/project" class="internal internal-link tag-link">project</a><span class="note-properties-separator">, </span><a href="../tags/weekend-project" class="internal internal-link tag-link">weekend-project</a><span class="note-properties-separator">, </span><a href="../tags/research" class="internal internal-link tag-link">research</a><span class="note-properties-separator">, </span><a href="../tags/ai" class="internal internal-link tag-link">ai</a><span class="note-properties-separator">, </span><a href="../tags/engineering" class="internal internal-link tag-link">engineering</a><span class="note-properties-separator">, </span><a href="../tags/llm" class="internal internal-link tag-link">llm</a><span class="note-properties-separator">, </span><a href="../tags/science" class="internal internal-link tag-link">science</a></span></td></tr></tbody></table></details><p show-comma="true" class="content-meta"><time datetime="2025-12-28T00:00:00.000Z">Dec 28, 2025</time><span>3 min read</span></p></div></div><article class="popover-hint"><div class="markdown-preview-view markdown-rendered"><p><a href="https://github.com/eamag/papers2dataset" class="external external-link">I built a tool<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> that uses AI agents to walk the citation graph and extract structured information. The first version was tailored to extract properties of Cryoprotective Agents, hereâs a shorter X tread:</p> 3<blockquote class="twitter-tweet"><p lang="en" dir="ltr">Listened to a recent episode of <a href="https://twitter.com/owl_posting?ref_src=twsrc%5Etfw" class="external external-link">@owl_posting<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> with <a href="https://twitter.com/huntercoledavis?ref_src=twsrc%5Etfw" class="external external-link">@huntercoledavis<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>, found out there's no dataset with data about cryotoxicity at different temperatures. Decided to build a tool that uses AI agents to walk the citation graph and extract CPAs and their properties 1/N â¬ï¸</p>â Dmitrii Magas (@EamagAI) <a href="https://twitter.com/EamagAI/status/2002822962937041153?ref_src=twsrc%5Etfw" class="external external-link">December 21, 2025<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a></blockquote> 4<h2 id="why">Why<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#why" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 5<p>AlphaFold exists because PDB exists, but most of the time the dataset you need is not there. Some data only exists as text in papers, and it takes too long to manually search and extract one data point at a time. This project helps to automate data extraction from open access papers, using AI agents to walk the citation graph and creating a CSV with data and sources.</p> 6<h2 id="how">How<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#how" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 7<h3 id="manual-testing">Manual testing<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#manual-testing" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 8<p>I always start doing things manually to understand how difficult the task is. Here I started with searching for relevant papers using semantic search tools like <a href="https://platform.edisonscientific.com/" class="external external-link">https://platform.edisonscientific.com/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> and <a href="https://asta.allen.ai" class="external external-link">https://asta.allen.ai<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>. Iâve got a list of papers that I read and tried to find relevant information and jotted it down. The
8n I uploaded these PDF to <a href="https://aistudio.google.com/" class="external external-link">https://aistudio.google.com/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> and spent some time figuring out a correct prompt for an LLM to extract the same structured information for different PDF. That worked quite well, so I just had to automate it</p> 9<h3 id="combining-things-together">Combining things together<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#combining-things-together" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 10<p>I wanted to make it easy to work with different LLM providers, so I chose <a href="https://www.litellm.ai/" class="external external-link">https://www.litellm.ai/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> to query different LLM endpoints, and <a href="https://openrouter.ai/models?max_price=0" class="external external-link">OpenRouter<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> to save some money by using free models. I tried using AI agents like Claude Code to write the most of the code, but I noticed it made too many mistakes in details, so in the end I used <a href="https://antigravity.google/" class="external external-link">https://antigravity.google/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> IDE as a really good autocomplete. I used <a href="https://openalex.org/" class="external external-link">https://openalex.org/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> to fetch papers, their PDF location and citations because I still canât get <a href="https://www.semanticscholar.org/" class="external external-link">https://www.semanticscholar.org/<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> API key :(</p> 11<h3 id="biggest-problems">Biggest problems<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#biggest-problems" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 12<ul> 13<li>
13PDF downloads. Even though many recent papers are in open access with preprints available, sites like BioArxiv block PDF downloads, and I had to spend some time figuring out how to actually download papers for reviews. I tried to built some additional functions to use official APIs to download PDFs, but this is not feasible to do for every location!</li> 14<li>Meta prompts. It was easy to build CPA-specific pipeline, but to extend it to other goals took some time, mostly because the model makes slight mistakes but has no feedback loop to change the result when the pipeline is in progress. Thatâs why I added Agent Skills</li> 15</ul> 16<h3 id="agent-skills">Agent Skills<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#agent-skills" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 17<p>After trying to solve problems above manually for other goals, I realized I was doing the same work, and it should be possible to fix it all automatically. Iâve also read that <a href="https://agentskills.io/home" class="external external-link">https://agentskills.io/home<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> became standardized, and Iâve added the more descriptions so this project can be installed as a skill. Iâve tested it for Claude Code and OpenAI Codex, and it works, but itâs not amazing. The good news: itâs the worst it will ever be, so I can pick this up next year and everything should work better with new models!</p></div></article><hr/><div class="page-footer"></div></div><div class="right sidebar"><div class="graph"><h3>Graph View</h3><div class="graph-outer"><div class="graph-container" data-cfg="{"drag":true,"zoom":true,"depth":1,"scale":1.1,"repelForce":0.5,"centerForce":0.3,"linkDistance":30,"fontSize":0.6,"opacityScale":1,"showTags":true,"removeTags":[],"focusOnHover":false,"enableRadial":false}"></div><button class="global-graph-icon" aria-label="Global Graph"><svg version="1.1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" x="0px" y="0px" viewBox="0 0 55 55" fill="currentColor" xml:space="preserve"><path d="M49,0c-3.309,0-6,2.691-6,6c0,1.035,0.263,2.009,0.726,2.86l-9.829,9.829C32.542,17.634,30.846,17,29,17 18 s-3.542,0.634-4.898,1.688l-7.669-7.669C16.785,10.424,17,9.74,17,9c0-2.206-1.794-4-4-4S9,6.794,9,9s1.794,4,4,4 19 c0.74,0,1.424-0.215,2.019-0.567l7.669,7.669C21.634,21.458,21,23.154,21,25s0.634,3.542,1.688,4.897L10.024,42.562 20 C8.958,41.595,7.549,41,6,41c-3.309,0-6,2.691-6,6s2.691,6,6,6s6-2.691,6-6c0-1.035-0.263-2.009-0.726-2.86l12.829-12.829 21 c1.106,0.86,2.44,1.436,3.898,1.619v10.16c-2.833,0.478-5,2.942-5,5.91c0,3.309,2.691,6,6,6s6-2.691,6-6c0-2.967-2.167-5.431-5-5.91 22 v-10.16c1.458-0.183,2.792-0.759,3.898-1.619l7.669,7.669C41.215,39.576,41,40.26,41,41c0,2.206,1.794,4,4,4s4-1.794,4-4 23 s-1.794-4-4-4c-0.74,0-1.424,0.215-2.019,0.567l-7.669-7.669C36.366,28.542,37,26.846,37,25s-0.634-3.542-1.688-4.897l9.665-9.665 24 C46.042,11.405,47.451,12,49,12c3.309,0,6-2.691,6-6S52.309,0,49,0z M11,9c0-1.103,0.897-2,2-2s2,0.897,2,2s-0.897,2-2,2 25 S11,10.103,11,9z M6,51c-2.206,0-4-1.794-4-4s1.794-4,4-4s4,1.794,4,4S8.206,51,6,51z M33,49c0,2.206-1.794,4-4,4s-4-1.794-4-4 26 s1.794-4,4-4S33,46.794,33,49z M29,31c-3.309,0-6-2.691-6-6s2.691-6,6-6s6,2.691,6,6S32.309,31,29,31z M47,41c0,1.103-0.897,2-2,2 27 s-2-0.897-2-2s0.897-2,2-2S47,39.897,47,41z M49,10c-2.206,0-4-1.794-4-4s1.794-4,4-4s4,1.794,4,4S51.206,10,49,10z"></path></svg></button></div><div class="global-graph-outer"><div class="global-graph-container" data-cfg="{"drag":true,"zoom":true,"depth":-1,"scale":0.9,"repelForce":0.5,"centerForce":0.2,"linkDistance":30,"fontSize":0.6,"opacityScale":1,"showTags":true,"removeTags":[],"focusOnHover":true,"enableRadial":true}"></div></div></div><div class="recent-notes"><h3>Recent Notes</h3><ul class="recent-ul"><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../2026/ai-calibration" class="internal">When an AI Says It's 90% Sure, How Would You Check?</a></h3></div><p class="meta"><time datetime="2026-09-17T00:00:00.000Z">Sep 17, 2026</time></p><ul class="tags"><li><a class="internal tag-link" href="../tags/ai">ai</a></li><li><a class="internal tag-link" href="../tags/llm">llm</a></li><li><a class="internal tag-link" href="../tags/discovery">discovery</a></li><li><a class="internal tag-link" href="../tags/research">research</a></li><li><a class="internal tag-link" href="../tags/ai-safety">ai-safety</a></li><li><a class="internal tag-link" href="../tags/weekend-project">weekend-project</a></li></ul></div></li><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../2026/backgammon-bot" class="internal">
27I Asked Meta Spark 1.3 to Build an AI Backgammon Bot. Why Did Learning Stall?</a></h3></div><p class="meta"><time datetime="2026-09-15T00:00:00.000Z">Sep 15, 2026</time></p><ul class="tags"><li><a class="internal tag-link" href="../tags/ai">ai</a></li><li><a class="internal tag-link" href="../tags/project">project</a></li><li><a class="internal tag-link" href="../tags/weekend-project">weekend-project</a></li><li><a class="internal tag-link" href="../tags/agents">agents</a></li></ul></div></li><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../" class="internal">Home</a></h3></div><p class="meta"><time datetime="2026-09-15T00:00:00.000Z">Sep 15, 2026</time></p><ul class="tags"></ul></div></li></ul></div><div class="toc"><button type="button" class="toc-header" aria-controls="toc-31" aria-expanded="true"><h3>Table of Contents</h3><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="fold"><polyline points="6 9 12 15 18 9"></polyline></svg></button><ul id="list-0" class="toc-content overflow"><li class="depth-0"><a href="#why" data-for="why">Why</a></li><li class="depth-0"><a href="#how" data-for="how">How</a></li><li class="depth-1"><a href="#manual-testing" data-for="manual-testing">Manual testing</a></li><li class="depth-1"><a href="#combining-things-together" data-for="combining-things-together">Combining things together</a></li><li class="depth-1"><a href="#biggest-problems" data-for="biggest-problems">Biggest problems</a></li><li class="depth-1"><a href="#agent-skills" data-for="agent-skills">Agent Skills</a></li><li class="overflow-end"></li></ul></div><div class="backlinks"><h3>Backlinks</h3><ul id="list-0" class="overflow"><li><a href="../2025/links/links-and-a-retrospective-of-2025" class="internal">Links for 2025 (And A Retrospective)</a></li><li><a href="../2026/geometric-deep-learning-in-jax" class="internal">Is JAX A Good Fit For Geometric Deep Learning?</a></li><li><a href="../" class="internal">Home</a></li><li class="overflow-end"></li></ul></div></div><footer class><p>Created with <a href="https://quartz.jzhao.xyz/">Quartz v5.0.0</a> © 2026</p><ul><li><a href="https://github.com/eamag">GitHub</a></li><li><a href="https://x.com/EamagAI">Twitter/X</a></li><li><a href="https://www.linkedin.com/in/eamag/">LinkedIn</a></li><li><a href="https://eamag.me/index.xml">RSS</a></li><li><a href="http://eepurl.com/i35ils">Newsletter</a></li></ul></footer></div></div><!-- Cloudflare Pages Analytics -->
vendor: 99 bytes, line 27
27<script defer src='https://static.cloudflareinsights.com/beacon.min.js' data-cf-beacon='{"token": "
273cc480702b614857b7b14f0e3ea77531
vendor: 61 bytes, line 27
27"}' type="3af50e427fb83ca387164fec-text/javascript"></script>
27<!-- Cloudflare Pages Analytics -->
vendor: 142 bytes, line 27
27<script src="/cdn-cgi/scripts/7d0fa10a/cloudflare-static/rocket-loader.min.js" data-cf-settings="3af50e427fb83ca387164fec-|49" defer></script>
27<script type="module" src="https://static.cloudflareinsights.com/beacon.min.js/v31edd6df95cf4e85bb4c19e7a9bdbcba1788362987495" integrity="sha512-iIg7k2xntmwu6/uSb5tpc/hySgZc4eoL31yB29W6tJFo2akwjPWcEqnCEdJvGexCL0KEQwVYv5BlowfhVz26hg==" data-cf-beacon='{"version":"2024.11.0","token":"57213ca53c61439ab0cebc0fe6ef211a","r":1,"spa":2}' crossorigin="anonymous"></script>
27 28</body>
28<script src="../static/resource-after-e13fffa1.js" type="3af50e427fb83ca387164fec-application/javascript" data-persist="true"></script>
28<script src="../static/resource-after-8b336f9a.js" type="3af50e427fb83ca387164fec-application/javascript" data-persist="true"></script>
28<script src="../static/resource-after-baba652b.js" type="3af50e427fb83ca387164fec-application/javascript" data-persist="true"></script>
28<script src="../static/resource-after-5eb75950.js" type="3af50e427fb83ca387164fec-module" data-persist="true"></script>
28<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/contrib/copy-tex.min.js" type="3af50e427fb83ca387164fec-application/javascript" data-persist="true"></script>
28<script src="../postscript-237007ac.js" type="3af50e427fb83ca387164fec-module" data-persist="true"></script>
28</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.