Work

First engineering hire at NOSBAAN — designed the platform architecture from scratch. In production. • Event-driven pipeline: instant acknowledgment, durable queues, async Cloud Run workers (no timeouts, no data loss) • Go microservices behind a Redis job queue — exponential backoff, idempotency guarantees • Load-tested on base infrastructure: 50+ req/s, 25,000+ ops/hour, avg <200ms, P95 <1.5s, 0.00% error rate • Stack: TypeScript, Go, Python, PostgreSQL, Redis, Cloudflare, Google Cloud, Vercel

Mon (mnw) script has no usable training data and no existing OCR pipeline. Scanned physical archives are inaccessible to any existing toolchain. Built a 6.6M-parameter MobileNetV3 + BiLSTM-384 + CTC model (~25 MB, FP32) covering a 315-character Mon charset. Deployed across three runtimes with the same weights: ONNX + Wasm on web (SvelteKit PWA, Cloudflare), ONNX + NNAPI on Android (Jetpack Compose), CoreML + ANE on iOS (SwiftUI, SwiftData). An opt-in Go ingestion API collects user-submitted corrections back into training data — the primary mechanism for building the first Mon language dataset. Also published as an NPM package and a PyPI CLI. • Model: MobileNetV3 + BiLSTM-384 + CTC, ~6.6M params, ~25 MB, 315-char charset • Runtimes: ONNX+Wasm (web), ONNX+NNAPI (Android), CoreML+ANE (iOS) • Platforms: Web (live), Android (Play Store), iOS (App Store review pending) • Stack: Python, PyTorch, ONNX, CoreML/ANE, NNAPI, SvelteKit, Jetpack Compose, SwiftUI, Go

Geospatial data platform serving local governments and civil society organizations across Africa, South Asia, and Southeast Asia. Partners used it for community mapping, land-use planning, and evidence-based policy — contexts where commercial GIS tools were too expensive or too complex. • Problem: No accessible, maintainable GIS tool for small teams in low-resource contexts • Approach: Full-stack platform with CI/CD, adapted for different partners over 4 years • Impact: In active use across multiple countries • Stack: React, PostGIS, Mapbox GL, MongoDB, Redis, GraphQL, GitLab CI/CD, Linode

National biodiversity portal on a multi-service stack (React CSR + Geoserver + CKAN + Airtable + Express + Linode VPS) that required dedicated ops to maintain. Not sustainable for a small research team. Replaced with Next.js + Baserow on Vercel. Infrastructure costs dropped 33%, Airtable seat costs eliminated, load times improved. • Before: React CSR + Geoserver + CKAN + Airtable + Express + Linode VPS • After: Next.js (SSG/ISR) + Geoserver + Baserow + Vercel • Impact: In production — 33% cost reduction, faster loads, no API rate limits • Stack: Next.js, Baserow, PostgreSQL, Geoserver, Vercel

Wildlife trafficking data is fragmented across dozens of sources, jurisdiction-sensitive, and operationally sensitive. Built a centralized tracking platform in partnership with Kontinentalist and the Global Initiative Against Transnational Organized Crime. • Challenge: Fragmented sources, strict compliance requirements, no second chance to get security wrong • Outcome: Live intelligence platform used for transnational crime prevention • Stack: SvelteKit, PostgreSQL, Vercel