Writing

Notes

Field notes on what I build and what it teaches me — ML infrastructure, geospatial systems, and shipping software for hard places. Short and practical; more as I write them.

InfrastructureAug 20258 min read

Architecting an async ML inference pipeline on AWS

Detection jobs don't belong in a request/response cycle. Here's how the SORA pipeline works end to end — a presigned S3 upload, an SQS message to decouple submission from compute, GPU workers that long-poll for jobs, and a FastAPI control plane tracking state in Postgres — with notes on idempotency, backpressure, and keeping GPU cost sane.

FastAPISQSEC2S3PostgreSQL

GeospatialJun 202510 min read

Physics-based flood simulation, from DEM to depth map

Flood-Sight turns a rainfall total and a terrain model into a flood-depth map. Inside the engine: SCS Curve Number for runoff, Landlab's OverlandFlow solver for shallow-water flow, and how I wrapped a slow scientific simulation in a Celery + FastAPI backend with a MapLibre GL frontend that stays responsive while jobs run.

PythonLandlabCeleryMapLibre GL

Machine LearningApr 20256 min read

From YOLOv5 to YOLOv8: a 22% accuracy story

The model swap was the easy part. What actually moved the needle was the evaluation harness — class-specific precision and recall, IoU, and mAP that exposed where the detector failed across terrains — and using it to drive targeted retraining. A short case study in measuring before optimizing.

PyTorchYOLOOpenCV

ProductFeb 20257 min read

Shipping M-Pesa payments in a React/Node SaaS

Building Karamu for the East African market meant meeting customers where they are: M-Pesa for payments, SMS and WhatsApp for confirmations, and a real-time floor view over Socket.io. Notes on the Prisma schema, handling mobile-money callbacks idempotently, and designing for intermittent connectivity.

ReactTypeScriptPrismaSocket.io

DataDec 20245 min read

GeoJSON to TIFF: automating geospatial indexing

Manual coordinate review was the bottleneck in our annotation pipeline. I built a converter that indexes GeoJSON annotations into georeferenced TIFFs, cutting review time and lifting annotation accuracy about 25%. A small pipeline with an outsized effect on iteration speed.

PythonGDALRasterio

Back to portfolio