AI/ML Engineer/Fullstack Developer
I bridge the gap between intelligent systems and production-grade software, turning state-of-the-art machine learning models into scalable, elegant web applications that millions can rely on.
Projects
Multi-tenant AI assistant over Telegram & WhatsApp. Each user brings their own LLM keys and Google account; the backend orchestrates up to 5 tool-calling rounds against NVIDIA NIM (Llama 3.1) for Calendar & Drive. Keys and OAuth tokens are encrypted with AES-256-GCM, voice notes transcribed via Groq Whisper, and a prompt-injection sanitizer filters instruction-override patterns before any LLM call.
Next.js
Fastify
TypeScript
Supabase
NVIDIA NIM
Docker
Dengue early-warning platform predicting risk up to 4 weeks ahead for Colombia's 32 departments. Six model families were benchmarked on 2022–2023 held-out data — SARIMAX was deployed over the slightly better SARIMA for one reason: it attributes each forecast to specific climate variables, making predictions interpretable for public-health decision-makers. Integrated SIVIGILA surveillance, NASA POWER meteorology and NOAA's ONI; weighted RMSE of 25.28 across four horizons.
Next.jsTypeScript
Tailwind CSSSupabase
React
Hugging Face
Full-stack project management app on MERN + end-to-end TypeScript. Beyond the Kanban UI, the real work is in the auth surface: refresh-token rotation with reuse detection (a stolen token revokes the entire chain), httpOnly + SameSite=None cookies tuned for the Vercel/Render cross-site split, and 4 rate-limit buckets backed by Upstash Redis. Covered by Playwright E2E + Vitest with mongodb-memory-server, and instrumented with Sentry & Better Stack across both surfaces.
ReactTypeScript
MongoDB
Express
Redis
SentryDocker
Foliar disease segmentation for sugarcane. A YOLO26m-seg model trained on 6,000+ field images (5 disease classes) achieving mAP@0.5 of 91.7%, exported to LiteRT and served entirely in the browser via WebGPU — no server roundtrip for inference, p99 of 47ms client-side. The FastAPI backend exists only to collect latency from every visitor into a ring buffer — observable MLOps, with a live p50/p99 dashboard and automatic WASM fallback on unsupported browsers.
ReactTypeScript
WebGPU
PyTorch
FastAPIAbout Me
I'm an AI/ML engineer and fullstack developer working at the intersection of applied ML and production software. I led software & firmware R&D at the GIRSI research group — shipping deep-learning models for automated container-code recognition in maritime logistics — and fine-tuned and evaluated LLMs as a Latam Coder at Outlier.
What I care about is the whole pipeline: dataset curation, model training, inference optimization, the API that serves it, and the web app that puts it in someone's hands. YOLO segmentation running in the browser via WebGPU at 47ms p99, an epidemiological forecast pinned to interpretable climate variables, a Telegram assistant orchestrating five tool-calling rounds against NVIDIA NIM — that's where I do my best work.
Experience
Lead the software & ML workstream for the maritime-logistics line: deep-learning models for automated container-code recognition, plus the end-to-end software and embedded firmware stack for the “Interactive Forests” project. Mentor student researchers across ML fundamentals, software architecture and R&D execution.
Contributed to the training, fine-tuning and evaluation of advanced AI models through full-stack application development. Integrated Model Context Protocol (MCP) and conversational-AI architectures to improve context handling and response accuracy, and maintained quality with test suites, bug triage and curated multimodal datasets via image annotation and technical transcription.
Tutor for 200+ students across Circuits I & II, Electronics I, Electricity & Magnetism, Calculus I & III, Classical Mechanics and Linear Algebra for the Departments of Physics & Geosciences, Electrical & Electronics Engineering, and Basic Sciences. CRLA Level I certified tutor (ITTPC).
Skills
Python
PyTorch
YOLO
OpenCV
Hugging Face
OpenAI
LangChain
NVIDIA NIM
TypeScript
React
Next.js
Fastify
Express
FastAPI
Supabase
MongoDB
PostgreSQL
Tailwind CSS
Vite
Redis
Docker
GitHub Actions
Vercel
Render
Sentry
Prometheus
WebGPU
AWS
Huawei Cloud
Cisco