# Brayan Pedraza > Engineer in Bogotá, Colombia, and co-founder of ZAVORA SAS. Builds AI software that real operations run on: a hospital network's maintenance platform (ETHER), AI agents that sell over WhatsApp (ZAVORA), and privacy-first document tools for AI agents (Sırdaş). Facts below are dated; last updated 2026-09-25. ## Facts - ETHER: AI maintenance and help-desk platform in daily use at a multi-site public university hospital network in Colombia. QR and voice reporting, WhatsApp operations for technicians, four scheduled AI agents (dispatcher, deadline watcher, closure auditor, weekly report). Seven major versions between 2026-02-22 and 2026-09-16; 145 commits. Built solo. - ZAVORA (co-founder, with Camilo Abella): AI agents that take, price, charge and record orders over WhatsApp. August 2026 controlled test, the same conversations run through both engines (86 turns each), published with its limits: 16.3% of conversations closed as sales vs 11.6% for the rules bot it replaced. Winner of Innovalab batch 3 (Cámara de Comercio de Bogotá, among 20 startups); Future Giants showcase at GoFest 2026. Brayan leads integrations and the platform (public API, Python SDK, MCP server, payments). - Sırdaş (founder): privacy-first PDF tools for AI agents; web app, CLI, library, MCP server, n8n node, Python SDK. Published on npm and PyPI on 2026-09-17; 223 npm downloads 2026-09-17 to 2026-09-24. Public launch 2026-10-20. - KRONOS.AI: time-and-motion studies with computer vision (MediaPipe) and quasi-Monte Carlo simulation. - 879 commits authored in 2026 (Jan 1 to Sep 24) across his repositories. ## Field Notes - [Seven Marks: how a keyword-matching prototype became a hospital's AI platform](https://brayanpedraza.com/notes/seven-marks/) - [16.3% vs 11.6%: how we measured whether our AI agent actually sells more](https://brayanpedraza.com/notes/measuring-ai-sales/) - [One mind, an army of agents: how I ship like a team](https://brayanpedraza.com/notes/army-of-agents/) ## Languages The site is available in English, 中文, हिन्दी, Español, العربية, Français, বাংলা, Português, Русский and Bahasa Indonesia at https://brayanpedraza.com/{en|zh|hi|es|ar|fr|bn|pt|ru|id}/ (English at the root). ## Links - [Full profile for LLMs: every section and all Field Notes in one file](https://brayanpedraza.com/llms-full.txt) - [Portfolio](https://brayanpedraza.com/) - [LinkedIn](https://www.linkedin.com/in/brayan-andres-pedraza-c) - [GitHub](https://github.com/Brayan15p) - [Substack](https://brayannpedrazza.substack.com) - [Instagram](https://www.instagram.com/andresspedraza/) - [ZAVORA team page](https://zavorai.com/nosotros) - [ETHER](https://ether-ia.com) - [Sırdaş on npm](https://www.npmjs.com/package/sirdas) - [Sırdaş on PyPI](https://pypi.org/project/sirdas/) - [KRONOS.AI](https://kronos-ia.vercel.app) --- # Full profile (generated from https://brayanpedraza.com; English source) ## Manifesto — https://brayanpedraza.com/#manifesto ### The Manifesto What I believe about life and about building. Each belief is paired with the work that shows I live by it. We get one life. I refuse to spend it as a spectator. The future is not something that happens to us. It is something we build, with our own hands, one version at a time. I want to own my days, chase the dreams that scare me, and leave a dent in the universe that outlasts me. From Bogotá, with AI as my team, I am building it now. **Build what should exist.** Not what is easy. Not what is trending. The things the world is quietly waiting for. ETHER runs every day inside a public hospital network. **Ship it. Then make it insanely great.** The first version is a promise. Every version after is how you keep it. ETHER Mark I used keyword matching, and said so. Mark VII runs four AI agents. **Great technology disappears.** The best software never asks people to change. It meets them exactly where they already are. Technicians close tickets and buyers place orders, all from WhatsApp. **Speed is a form of respect.** Every day a problem stays unsolved, someone pays for it. Moving fast is how I honor their time. 879 commits in 2026. Seven versions of ETHER in seven months. **Truth over applause.** Measure against the old way. Publish what stalled. Hype expires. Credibility compounds. 16.3% vs 11.6% on the same conversations, published with the limits of the sample. The ledger lists every project, including the paused ones. **Trust is the product.** People hand us their health, their money, their names. Guard it as if it were your family's. Nightly tests that sign in as every role. Tools that never upload a file. **Focus is saying no.** Every yes costs a thousand others. Knowing what to stop is part of building what matters. LimpioGO became MANO. MANO is paused. Both are on the record. **One mind. An army of agents.** AI is the most powerful bicycle for the mind ever built. I set the vision, agents multiply the hands, and I own every result. Specs, reviews and tests in every repository. **Where you start is your advantage.** Building from Latin America means seeing up close the problems others only read about, then solving them for the world. Customers in Colombia and Argentina. Packages on npm, PyPI and the MCP Registry. **Sign your work.** Care about the parts no one will ever see. Craft is how you tell the universe you were here. You are reading one. ## Works — https://brayanpedraza.com/#works ### Selected Works Shipped, in use, measured, and the ones that taught me the most. Open any case for the problem, what I built and the result. **ETHER** The AI maintenance platform a multi-site public university hospital network runs on every day. - 145 commits, April to September 2026 - 4 AI agents: dispatcher, deadline watcher, closure auditor, weekly report - 7 major versions in seven months **Problem** Hospital engineering teams handle maintenance requests across several sites, each with a deadline, a technician and a sign-off. Phone calls and chat do not scale. **What I built** Deadlines in business days, no-login reporting by QR code or voice, WhatsApp operations for technicians, tool lending, a contractor portal, and an assistant that acts on tickets through tool calls. Four departments on one platform. **Why me** I work in the hospital's engineering department. I build next to the people who use it. Next.js · TypeScript · Supabase (Postgres RLS, Realtime, pg_cron) · Groq · Twilio WhatsApp · Vercel **ZAVORA** AI agents that take, price, charge and record orders over WhatsApp, from text, photos or voice notes. - 16.3% of conversations closed as sales, vs 11.6% for the rules bot (Aug 2026) - 1 of 20 Innovalab batch 3 winner, Bogotá Chamber of Commerce - 183 of my commits in September: public API, SDK, MCP server, payments **Problem** At GoFest 2026, the businesses that stopped at our stand kept describing the same scene: a WhatsApp inbox overflowing with orders. **My part** Co-founded with Camilo Abella, who leads product and the conversational agent. I lead integrations and the platform: Shopify two-way inventory, Mercado Pago, Wompi, Alegra e-invoicing, Kommo CRM, a public API with idempotency, rate limits and webhooks, a Python SDK, an MCP server, and OIDC deploys with rollback. Python · FastAPI · Postgres (Neon, RLS) · Azure · Twilio · React · 3,700+ tests **Sırdaş** Privacy-first PDF tools for AI agents. Markdown, anonymization and datasets, without uploading a single file. - 6 ways to use it: web, CLI, library, MCP server, n8n node, Python - 223 npm downloads in the first eight days - Oct 20 public launch, 2026 **Problem** Teams want to give contracts, medical records and invoices to AI, but those files carry personal data that cannot go to someone else's server. **What I built** One engine that runs entirely on the user's machine, with OCR in WebAssembly. Four npm packages and a PyPI package published Sep 17, 2026, plus an MCP Registry entry. TypeScript · pdf.js · Tesseract OCR · qpdf (WASM) · Web Workers · MCP · Python **KRONOS.AI** Time-and-motion studies with computer vision, a Monte Carlo engine, and an LLM that explains the results. - Vision pose and hand tracking in the browser with MediaPipe - Math Latin hypercube, Sobol and Halton sampling, bootstrap intervals, in a Web Worker **MANO** A two-sided mobile marketplace that connects companies with vetted cleaning, maintenance and logistics crews. - 79 commits in eight weeks, May to July 2026 - ~20 serverless functions: live map, QR and GPS check-in, escrow payments, an AI assistant - 5 days to build LimpioGO, its B2C predecessor, before pivoting to B2B **Problem** Companies need reliable crews on short notice, and independent workers need steady jobs and to get paid on time. **What I built** An Expo app with a live map, QR and GPS check-in to prove the work happened, escrow payments through Mercado Pago, training mini-courses, and "Maia", an AI assistant. It started as LimpioGO, a home-cleaning app with photo-to-quote pricing via Claude vision, and pivoted to serve companies. Expo SDK 54 · React Native · Supabase (13 migrations, edge functions) · Groq Llama 3.3 70B · Mercado Pago · EAS **LOK** Credit offers built from consented Open Finance data, so people get a fair answer before they even ask. - 4 agents: reader, scorer, originator and guardian - MFA required for financial data, with AES-256 column encryption and an append-only audit log - 1 self-correction that mattered: dropped email and calendar signals for policy reasons and moved to Open Finance **Problem** Many people in Latin America are invisible to traditional credit scoring, even when their finances are healthy. **What I built** A consent contract, bank connection through an Open Finance provider, a digital-footprint check, a rules classifier with a machine-learning fallback, a personalized offer engine, and a Habeas Data erasure routine. Hardened through red-team and blue-team passes. Next.js 15 · React 19 · Supabase (13 migrations, RLS) · Tailwind v4 · three.js landing **ATLAS** The paperwork engine for a hospital's engineering department: contracts, invoices, procurement and insurer billing objections. - 166 commits and 42 pull requests - 5 scheduled agents on a single hourly heartbeat - 0 bytes of contract data uploaded: OCR runs in the browser, under Colombia's Ley 1581 **Problem** Hospital engineering teams lose days to contract lifecycles, contractor reports and billing objections with legal deadlines. **What I built** Contracts read from PDF or Word with deterministic rules and in-browser OCR, generated invoices and supervision reports, procurement-plan and public-procurement lookups, contractor background checks, and an email poller for insurer objections. Built on the same foundation as ETHER. Next.js 16 · Supabase · Groq · pdf.js · Tesseract · docxtemplater · Python docgen · Vitest **World Cup 2026** A real-money pool for friends and a statistical engine to predict every match of the tournament. - 1,096 bets settled across all 104 matches, results cross-checked between two data APIs - 15,000 Monte Carlo runs on a bivariate Poisson model with Dixon–Coles adjustments - 4 days to build the model, backtested on the 2014, 2018 and 2022 World Cups ## Ledger — https://brayanpedraza.com/#ledger ### The Ledger Everything I have built, including what stalled. Honest status, real dates. Showing 16 of 16 - When Build What it is Status Link - Sep 2026 Sırdaş Privacy-first PDF tools for AI agents. Web, CLI, MCP, n8n, Python. Published npm ↗ - Apr–Sep 2026 ETHER AI maintenance platform in daily use at a public hospital network. In production Visit ↗ - Apr–Sep 2026 ATLAS Automates hospital-engineering paperwork: contracts, contractor invoices, procurement plan, insurer billing objections. 166 commits, 42 PRs. Pilot Private - May–Sep 2026 ZAVORA AI agents that sell over WhatsApp. Co-founder; integrations and platform. Live Visit ↗ - Jul 2026 ZAVORA CRM The founders' sales CRM with pipeline, hypothesis log and calendar sync. Built in two days. Internal Private - Jul–Aug 2026 LOK Credit offers from consented Open Finance data. MVP with MFA, encryption and an audit log. Co-founded. Demo Visit ↗ - Jun–Aug 2026 Polla Mundialista Real-money World Cup pool: 1,096 bets over 104 matches, two data sources cross-checked. Completed Visit ↗ - Jun 2026 World Cup model Bivariate Poisson with Dixon–Coles and a 15,000-run Monte Carlo, backtested. Built in four days. Prototype Code ↗ - Jun–Jul 2026 AZ Ecosistema Recycling platform for a municipality: citizens, informal recyclers, operators, admins. Multi-tenant monorepo. MVP Code ↗ - Jun 2026 Aula Virtual ICES Set up and documented a school's virtual classroom: 11 manuals, 31 automated checks. Delivered On request - May–Jul 2026 MANO Two-sided app for cleaning and maintenance crews: live map, QR and GPS check-in, escrow. Store-ready. Paused Private - May 2026 LimpioGO Home-cleaning marketplace with photo-to-quote via Claude vision. Built in five days, then pivoted into MANO. Pivoted Code ↗ - Mar–Jun 2026 KRONOS.AI Time-and-motion studies with computer vision and Monte Carlo. Live Visit ↗ - Feb 2026 PIGI ETHER's Mark I: QR incident reporting, three roles, row-level security. Superseded Code ↗ - Nov 2025 Agronecta Desktop app connecting farmers, restaurants and couriers. 75 Python modules. Coursework Code ↗ - Mar–May 2025 SQLi detector Team project: an ML model that flags SQL injection. I built the backend and model serving. Coursework Code ↗ ## Facts — https://brayanpedraza.com/#facts ### Quick Facts Short, dated answers for people and for the AI assistants that summarize them. **Who is Brayan Pedraza?** An engineer in Bogotá, Colombia, and co-founder of ZAVORA SAS. He built ETHER, an AI maintenance platform used daily by a public hospital network, and founded Sırdaş, privacy-first document tools for AI agents. **What is ETHER?** An AI maintenance and help-desk platform for hospitals with QR and voice reporting, WhatsApp operations and four scheduled AI agents. Seven major versions from Feb 22 to Sep 16, 2026. **What has ZAVORA measured?** In August 2026, running the same test conversations through both engines (86 turns each), its AI agent closed 16.3% as sales versus 11.6% for the rules bot it replaced. Winner of Innovalab batch 3, Bogotá Chamber of Commerce. ## Contact — https://brayanpedraza.com/#contact Contact ### The future belongs to those who ship. I'm looking for people building AI for real-world operations, and investors who back technical founders from Latin America. --- ## Field Note — https://brayanpedraza.com/notes/seven-marks/ ### Seven Marks: how a keyword-matching prototype became a hospital's AI platform The version-by-version story of ETHER, from a February prototype whose AI was keyword matching to four autonomous agents in daily use at a public hospital network. On February 22, 2026, I pushed a prototype called PIGI. It let anyone report a broken piece of infrastructure by scanning a QR code on the wall. It had three roles, reporters, technicians and admins, and Postgres row-level security keeping each one in its lane. It also had a feature called "AI categorization". It was keyword matching. The README said so. That line matters more to me than any feature in the product, because every honest version after it was built on top of that first admission. Seven months later the same idea, rebuilt and renamed ETHER, runs every day inside a multi-site public university hospital network. This is how it got there, one Mark at a time. ### The seven Marks - I Feb 22 · PIGI. QR reporting with no login, role workspaces, row-level security. "AI" by keyword, disclosed. - II Apr 19 · ETHER v1.0. Rebuilt from scratch as an AI help desk for the hospital's engineering team. - III Apr 26 · Agentic. Agentic AI, "ETHER Brain", plus a security pass. One week after v1.0. - IV May 17 · WhatsApp. Technicians create, update and close tickets from WhatsApp. The same day, four medium findings from an OWASP ZAP scan were fixed. - V May 24 · Multi-department. Housekeeping, occupational safety and patient safety joined, each with its own flow. Nine issues from a full audit closed. - VI Aug 15 · Proof of isolation. A test suite that signs in as every role and checks what each one can read and write. It runs every night. - VII Sep 16 · Autonomous agents. A morning dispatcher, a deadline watcher, a closure auditor and a weekly management report. A network view across every hospital. ### Rebuild when the foundation lies Mark II was not an upgrade. It was a new codebase. PIGI proved people would scan a QR code instead of calling someone, which was the only thing it needed to prove. Its internals were built for a demo, not for a hospital. Carrying them forward would have meant carrying their shortcuts forward too. The prototype's job is to answer one question cheaply. Once it has answered, its code has done its work. ### Go where people already are The change I am proudest of did not come from a smarter model. It came from Mark IV, when technicians could run their tickets from WhatsApp. They were never going to open another app between two repairs. The software had to come to them. Great technology disappears into the day of the person using it. ### Security is a test, not a promise A hospital network trusts its maintenance platform with who reported what, where, and when. So from Mark VI on, the permission rules are not something I believe are correct. They are something a test proves every night by impersonating each role. If a change leaks data, the suite fails before anyone else notices. ### Agents that work, not agents that chat By Mark VII the AI stopped being a chat window. Four agents run on a schedule: one assigns the morning's work, one watches deadlines, one checks that closed tickets were actually signed off, and one writes the weekly report for management. The interesting part is not that they use a language model. It is that each one does a job someone used to do by hand, on time, every time. ### What seven months taught me - Ship the embarrassing version. Mark I was small and honest. That is why there was a Mark II. - Version numbers are a promise. Each Mark had to be usable on its own, not a step toward some future release. - Build next to the people who use it. I work in the hospital's engineering department. The feedback loop was a hallway, not a survey. ETHER is in production at a public hospital network in Colombia. The dates above come from the repository history. Explore the seven Marks on the home page. --- ## Field Note — https://brayanpedraza.com/notes/measuring-ai-sales/ ### 16.3% vs 11.6%: how we measured whether our AI agent actually sells more How ZAVORA compared its WhatsApp sales agent against the rules bot it replaced: same conversations, a clear baseline, and the limits of the sample published next to the number. Every AI startup says its agent is better. At ZAVORA we wanted a number we could defend in front of a skeptical customer, and we wanted to publish it with its limits attached. This is how we measured it, and what the number does and does not mean. ### Beat the baseline, not a straw man The comparison was not "AI versus nothing". It was our agent versus the rules-based bot it replaced, the menus and keywords that many businesses on WhatsApp already use. A number only means something next to the best alternative a customer actually has. ### Same conversations, same catalog In August 2026 we ran the same set of conversations through both engines, 86 turns each, against the same product catalog. Same questions, same products, same prices. The only thing that changed was the engine answering. - The AI agent closed a sale in 16.3% of cases. The rules bot closed 11.6%. That is about 40% more, relative. - The agent gave a useful recommendation 73% of the time, against 43% for the bot. ### Publish the limits with the number Here is the part most companies leave out. This was a controlled test, not a market study. Eighty-six turns per engine is enough to see a clear difference, not enough to promise any specific business a specific lift. So that is exactly what ZAVORA's public guide says, right next to the figures, with the date of the measurement. Hype expires. Credibility compounds. ### An agent that admits what it does not know Part of why the agent sells more is that it does not invent. When a product or a price is not in the catalog, it says so and asks. A rules bot fails silently; a careless AI fails confidently. Neither closes the sale. We also built a check into our own content pipeline that blocks publishing any capability the product does not actually have. ### What I would tell any founder - Pick the real alternative as your baseline, not the easiest one to beat. - Hold everything constant except the thing you are testing. - Put the date and the limits next to the number. The people you want as customers will trust you more for it. ZAVORA was co-founded by Camilo Abella and me. Camilo leads the product and the conversational agent; I lead integrations and the platform. The measurement and its limits are published in ZAVORA's guide (in Spanish). --- ## Field Note — https://brayanpedraza.com/notes/army-of-agents/ ### One mind, an army of agents: how I ship like a team How I build with AI coding agents: specs, specialist agents, reviews and tests. 879 commits in 2026, and the rules that keep the work mine. In 2026 I authored 879 commits across my repositories. I shipped a hospital maintenance platform, the integrations layer of a startup, a set of developer tools on npm and PyPI, and a real-money World Cup pool with more than a thousand bets. I did not do it by typing faster. I did it by changing what my job is. I set the vision, agents multiply the hands, and I own every result. ### The spec is the product An agent is only as good as the instructions it receives. So most of my time goes into the part no model can do for me: deciding exactly what should exist. What the user does, what the edge cases are, what must never happen. A clear spec turns an agent from a guesser into a builder. ### Specialists, not one generalist In ATLAS, the paperwork engine I built for a hospital's engineering department, the repository includes dedicated agent definitions for different roles: interface, user experience, product management and agent architecture. Each one gets the context for its job and nothing else. It works for the same reason teams do: focus. ### Every line is reviewed Agents write a lot of code. Some of it is wrong. The commit history shows AI coding agents as co-authors across my projects, and every one of those commits went through review before it shipped. An agent's output is a draft until someone who understands the system accepts it. ### Tests are the contract When agents move fast, tests are what keep speed from turning into chaos. In ETHER, a permission suite signs in as every role every night and proves that nobody can see what they should not. In ZAVORA, the codebase carries thousands of automated tests. The rule is simple: if it is not tested, it is not done. One mind. An army of agents. Full ownership. ### What this changes - The bottleneck moves from typing to thinking. The scarce resource is judgment about what to build. - Small teams can take on big problems. A hospital platform, a startup's platform and a developer toolkit, in parallel, from Bogotá. - Ownership does not delegate. Agents multiply the work. The responsibility stays with me. Numbers from the git history of my repositories, January 1 to September 24, 2026. See the month-by-month momentum on the home page.