Three products I designed, built and shipped from scratch, from the database to the interface.
Android · On-device ML
HealthyCropping
In development
An offline-first Android app that detects coffee, cacao and banana diseases from a phone photo, built for farms in Costa Rica where the signal drops.
A
Diagnosis without signal
Each crop has its own MobileNetV3 model running on the phone with TensorFlow Lite. Scans, farms and plots stay in SQLite, so nothing depends on a connection.
B
Trained on real field photos
Public datasets from coffee, cacao and banana plots, deduplicated by perceptual hash and split by plant so no leaf leaks between training and test.
C
It shows where it looks
Grad-CAM heatmaps show which part of the leaf drove each prediction, and every experiment, including the failed ones, is documented.
Small businesses already sell through WhatsApp, one chat at a time. Booth gives each store a single link with its catalog and keeps orders, payments and stock in one place.
A
One link per store
Shoppers open the store link, browse the catalog and build a cart. No app to install and no account to create.
B
Orders, not chats
The cart turns into a clean WhatsApp message, stock is reserved atomically, and the owner moves each order from pending to ready on the dashboard.
C
Receipts checked by AI
For bank transfers, Gemini reads the receipt and tells the owner whether the amount and recipient match before the payment is confirmed.
Upload PDFs and notes, ask in English or Spanish, and get answers that cite the exact passages they use. When the documents do not hold the answer, it refuses instead of guessing.
A
Answers with receipts
Every answer cites numbered passages, and citations to passages the model was never given are removed and reported.
B
Hybrid retrieval, then a gate
pgvector and Postgres full-text search fused with Reciprocal Rank Fusion, an optional Cohere reranker, and a calibrated threshold that stops the model when the evidence is weak.
C
Measured on every pull request
A benchmark of 82 questions in English and Spanish runs in CI and fails the build if retrieval quality drops.
React Native
Expo
TypeScript
TensorFlow Lite
Python
FastAPI
Firebase
SQLite
Gemini
Vite
Tailwind
TanStack Query
NestJS
Prisma
PostgreSQL
Playwright
Vercel
SQLAlchemy
Angular
Cohere
Groq
Docker
pytest
LinkedIn