Jhoan Sebastian Duque Zapata's Work | ContraWork by Jhoan Sebastian Duque Zapata
Jhoan Sebastian Duque Zapata

Jhoan Sebastian Duque Zapata

AI Engineer building reliable LLM agents, RAG pipelines, and

    6
    Followers
Followed by Benjamin C, Muhammad , and Muhammad Adeen

Bogota, Colombia

WorkPostsAbout

Bogota, Colombia

As part of my transition from QA into AI Engineering, I built AskMyCar, a practice project to go deeper into LLMs and RAG. The idea is simple: you pick your car (brand, model, year) and ask questions about its owner's manual in plain language, like "how often should I change the oil?" or "what tire pressure do I need?". It answers using only the manual and cites the sources, instead of making things up. How it works: a Playwright scraper pulls the official manual, it's stored in AWS S3, and a RAG pipeline built with LangChain and Chroma retrieves the relevant parts to answer. It runs on Claude when an API key is set, or a local Ollama model as fallback. Exposed through a FastAPI backend with auth and rate limiting, and a React + TypeScript frontend deployed on Vercel. The whole pipeline is traced with LangSmith. Next step is automated evaluation with LangSmith to measure groundedness and correctness, where my QA background becomes a real advantage. Demo in the video. Feedback is welcome.
0
8
It started with a simple request. A client told us their team was still typing invoices into a spreadsheet by hand, one by one, and asked if there was a better way. At DZ Labs we like that kind of problem: small, real, and something people deal with every single day. So I built Invoice Register. You give it a PDF invoice, and a few seconds later the data is sitting in a Google Sheet: supplier, dates, totals, and every line item in its own row. It works with digital invoices and with scanned ones too. In the video you can see the whole trip: I upload a PDF, the API answers with the structured data, the rows show up in the spreadsheet, and the service is running live in the cloud. Here is what I used to build it: Python 3.12 and uv to manage the project and its dependencies FastAPI and Uvicorn for the API Pydantic to define and validate what an invoice looks like pdfplumber to read text from digital PDFs pdf2image, Poppler and Tesseract to read scanned invoices with OCR LangChain and Claude to turn raw text into structured data gspread and google-auth to write into Google Sheets LangSmith to trace every run and see what the model did pytest for the test suite And here is what I used to deploy it: Docker to package the app with its system dependencies Cloud Build and Artifact Registry to build and store the image Google Cloud Run to run it, scaling down to zero when nobody is using it Secret Manager to keep the API keys out of the code I will be honest: writing the code was the comfortable part. Getting it to production is where I learned the most. When the data belongs to someone else, security stops being optional. The service runs without any key files, the API is private, and every secret lives outside the code. What I enjoyed most was not the technology. It was knowing that someone will get back hours of their week. Next steps: a simple frontend for uploads, background processing, and automated deployments. If your team is still doing something by hand that a computer should be doing, I would love to hear about it. #Python #FastAPI #LangChain #GoogleCloud #CloudRun #AI #DZLabs
2
0
29
I built Tests Wise: paste a URL, get a test plan and watch AI run the tests live. Most small teams ship without a written test plan. Not because they don't care because writing one takes a day nobody has. And running it takes another. Tests Wise does both: Paste your site's URL (works with sites behind a login). AI agents explore up to 8 pages in a real browser and figure out what each one does. You get prioritized test scenarios steps, expected results, priority as an Excel your QA team or client can use right away. In English or Spanish. Then an AI agent executes those scenarios live in a real browser, driven through Playwright MCP, and reports what passed, what failed and why, with a screenshot of each step. Everything is automated software: no scripts, no selectors, no CI to wire up. Under the hood: FastAPI + PostgreSQL, Redis/RQ workers, Playwright for the crawl, Claude for scenario design and for driving the browser via Playwright MCP, LangGraph for the pipeline. Deterministic exploration, LLM only where it adds value — a full run on an 8-page site costs a few cents. It's free to try, no card needed: https://app.tests-wise.com/ Demo below 👇 I'd love to hear which site you'd throw at it first.
4
3
112