Work

Airbnb Superhost-recognised monthly rentals business in Bangkok. Sole engineer responsible for the full technology platform.

• Designed and built an event-driven ERP platform from scratch on GCP, owning architecture, implementation, and production operations end to end.
• Identified reliability issues with synchronous external integrations — redesigned the processing layer with an event-driven architecture, improving fault isolation and responsiveness under load.
• Eliminated request timeouts under heavy compute load by offloading long-running processing to dedicated Go workers.
• Diagnosed duplicate processing from at-least-once external delivery — designed idempotent pipelines with retry and exponential backoff, eliminating manual error recovery.
• Initiated systematic load testing, identifying and resolving throughput bottlenecks; platform sustains 50+ req/s at P95 < 1.5s.
• Built CI/CD pipelines and a full observability stack from scratch — neither existed; shortened time to detect and respond to production incidents.
• Secured the platform with least-privilege cloud IAM per service, managed secrets, and application-level rate limiting.
• Built an AI-assisted messaging platform (TypeScript, Go, OpenAI) to automate customer communication triage: intent classification, tool calling, and structured response generation.
• Designed RAG pipelines with human review checkpoints, flagging low-confidence responses before delivery and maintaining operational control over automated communications.

Stack: TypeScript, Go, PostgreSQL, Redis, OpenAI, GCP

Mon is a UNESCO-classified vulnerable language. No OCR toolchain or dataset existed before this project. Built an end-to-end system to digitise printed Mon materials and make them searchable. Two models: a lightweight on-device version (MobileNetV3 + BiLSTM + CTC, ~11.4M params) and a heavier server-side version (~54M params). Both export to ONNX; the mobile model also supports CoreML.

• Deployed across Web (WASM), Android (NNAPI), and iOS (CoreML).
• Human feedback loop to collect validated samples and improve model accuracy over time.

Stack: Python, PyTorch, ONNX, CoreML, SvelteKit, Go

Geospatial platform used by organizations across Africa and Asia for community mapping and land-use planning.

• Handled data contributions from organizations across multiple countries — each with different formats, update cadences, and access requirements.

Stack: React, PostGIS, Mapbox GL, MongoDB, Redis, GraphQL, GitLab CI/CD, Linode

Migrated a legacy system to a simpler Next.js + Baserow architecture, cutting infrastructure cost by 33%.

• Rebuilt data pipelines to reduce failures and manual cleanup.
• Added backup and recovery — neither existed in the previous setup.

Stack: Next.js, Baserow, PostgreSQL, Geoserver, Vercel

Built in partnership with Kontinentalist and the Global Initiative Against Transnational Organized Crime. Designed and built a global monitoring interface aggregating wildlife trafficking intelligence from multiple sources, handling sensitive data under strict security and compliance constraints.

• Engineered the data ingestion layer to normalize incident data across sources with different formats and update cadences.

Stack: SvelteKit, PostgreSQL, Vercel