Johan Alexis Rendon Granada

Problem

Bring data from external catalogs, web search and language models into structured product reports, and search inside them while seeing where each fragment comes from.

What I did

• I built the Full Report pipeline: 10 research and tagging lenses over one report, with server-side PDF export.
• I designed the semantic search (RAG): chunks split by content boundaries, local embeddings, hybrid retrieval (dense and lexical) and the provenance of each chunk (product, category, source and date).
• I implemented login with httpOnly cookies and CSRF protection, and the admin and developer roles, with an atomic admin transfer so the system never loses its last admin.
• I built the notifications, the role administration and the semantic search administration.

Technical decision: failover between LLM providers

I designed a routing layer between LLM providers with automatic failover and per-provider concurrency control. It also stores the token usage of each call, with an endpoint and an admin dashboard.

Result

Deployed and running in production in the company's private environment. It was handed to an internal team for daily maintenance.

Stack

• Python
• FastAPI
• PostgreSQL
• pgvector
• React
• TypeScript
• Gemini
• Groq

Problem

Turn loose posts from forums and communities into metrics a team can query: sentiment, churn risk, NPS and alerts.

What I did

• I designed 11 REST endpoints with FastAPI to list posts and query their metrics. Filters and pagination run in SQL, not in memory.
• I built the per-product ingestion endpoints: they search, analyze with the LLM and store the results in PostgreSQL.
• I built the live feed in React, with pagination, filters, search and statistics.
• I wrote the tests: 25 backend (pytest) and 29 frontend (Vitest, React Testing Library and MSW). I built the frontend test setup from scratch.
• I documented the schema, the frontend integration and the tests so the team could maintain them.

Technical decision: structured output from Gemini

Gemini rejected Pydantic JSON Schemas with references ($ref, $defs), unions (anyOf) and constants (const), so the answers were not valid JSON. I wrote a resolver that flattens the schema to a format the model accepts.

Result

54 passing tests (25 backend and 29 frontend) and documentation of the schema, the integration and the tests.

Stack

• Python
• FastAPI
• PostgreSQL
• SQLAlchemy
• Pydantic
• Gemini
• React
• TypeScript

Problem

The document management area of SENA needed to archive documents like in a drive and manage PQRSD (petitions, complaints, claims, suggestions and reports), following the Colombian General Archives Law.

What I did

• I did full stack QA: I tested both the frontend and the backend.
• I created the mockups and designed several React interfaces.
• I led the test suite: PHPUnit on the backend and Vitest on the frontend.
• I deployed the system to production on DigitalOcean, in containers with Dokploy.

Challenge: unclear requirements at the start

A communication mistake with the client doubled the work: we built features for methods the client did not want. The team held a clarification meeting and fixing it took a couple of days. Since then we ask the client constant questions and write down every agreement.

Result

In production, with about 60 real users. Team of 9 people.

Stack

• Laravel
• React
• MySQL
• PHPUnit
• Vitest
• DigitalOcean
• Dokploy