Systems & Industrial Engineering student. I build production systems: an ERP that a paint factory runs its accounting on, and a multi-sport betting engine that only ships a pick when it can prove an edge against the closing line.

Most of what I enjoy sits where data meets a real decision — pricing, inventory, forecasting, risk.

A web/PWA ERP that replaced the paper notebooks of Pantón Color, a Colombian paint factory. Ten modules covering accounting, color formulas, raw-material inventory, quotations, and production costing.

The parts worth looking at:

• Colombian tax retentions (ReteFuente / ReteIVA / ReteICA), monthly periods with locked closings, and an append-only audit log.
• Money-safe by construction — business logic lives in Postgres SECURITY DEFINER RPCs behind Row-Level Security, so amounts are computed server-side and cannot be tampered with from the client.
• Atomic production runs — starting a run freezes an ingredient snapshot and deducts inventory in a single transaction.
• Weighted-average costing on material purchases.

Next.js 16 · React 19 · TypeScript · Supabase / PostgreSQL · RLS · Tailwind · 248 tests

Detects statistical edge with a gradient-boosting ensemble, has an LLM validate the pick, and delivers it over Telegram. A human places the bet; nothing is automated end to end.

The core thesis is a discipline, not a feature: ship a pick only when CLV beats the Pinnacle closing line by more than +3%. If there is no edge, the system sends nothing. Measuring against the closing line is the honest way to tell signal from variance.

• Ensemble of XGBoost / CatBoost / LightGBM, with Dixon-Coles and bivariate Poisson goal models.
• Sport-agnostic core with a plugin layer per sport.
• Every secret loads from the environment through Pydantic settings — no credentials in code.

Python 3.12 · Polars · XGBoost / CatBoost / LightGBM · Pydantic v2 · Supabase · uv · 874 tests in CI

Given an Academic Development Plan as a PDF, the agent checks it against 179 institutional guidelines and returns a report with cited evidence and prescriptive corrections. Includes a head-to-head benchmark of a rule-driven dispatcher against semantic RAG.

Python · FastAPI · Qwen 2.5 14B · Docling · RAG · Next.js · IEEE technical report

Spanish-language grocery PWA with real-time macronutrient tracking, built around WCAG accessibility and UX research.

Live: macroly.vercel.app

Next.js 16 · TypeScript · Zustand · Supabase · PWA · WCAG