Duch Panhathun

Ailsa HQ Limited

A six-service platform that helps businesses discover grant funding, match against it, and manage their applications — with AI integrated throughout.

• Built five funder-specific scrapers for major UK and EU funding bodies, each using a different extraction strategy: undocumented JSON API reverse-engineering, HTML parsing, hybrid headless-browser fallback, and sitemap-less crawling — with dedicated automated test coverage for each.
• Designed the normalisation layer mapping five heterogeneous funder sources onto a single grant model, plus a compatibility adapter bridging a new schema onto the platform's legacy one.
• Built a provider-agnostic LLM classification and structured-extraction layer across OpenAI and Anthropic, filtering scraped pages down to relevant application data.
• Bootstrapped and served as primary engineer for the customer-facing SaaS app from an empty repository — authoring the founding commit and ~62% of all commits.
• Designed a department- and section-level access-control system, deriving section ownership transitively from user–department membership rather than requiring a schema migration.
• Contributed to the AI matching engine, pairing funding opportunities to company profiles using embedding-based vector search.

• Python
• FastAPI
• Next.js
• MongoDB
• AWS SQS
• Auth0
• OpenAI
• Anthropic
• Docker

Save the Children Cambodia

Deep-learning models that identify diseases in cucumber and cauliflower crops from a photo, delivered to farmers through a Telegram chatbot that also recommends treatment.

• Trained and deployed deep-learning models across two crop families — ConvNeXt-Large and EfficientNet-B5 for cucumber (98.06% test accuracy, 94.82% macro-F1 across 7 classes) and EfficientNetB3 for cauliflower (99.74% test accuracy, 99.68% macro-F1 across 5 classes), each measured on a held-out test split.
• Built complete training pipelines in both PyTorch and TensorFlow/Keras: focal loss for class imbalance, mixed-precision training, cosine learning-rate scheduling, stratified splitting and augmentation, resolving non-trivial Keras 3.x / HuggingFace compatibility issues along the way.
• Built a weather advisory microservice combining government meteorological data with the Windy API, localised into Khmer and deployed as both a scheduled scraper and a REST API.
• Built and deployed the bilingual admin dashboard, including automated weather-alert cron jobs.

• PyTorch
• TensorFlow / Keras
• Python
• Telegram Bot API
• Next.js
• Docker

Save the Children Cambodia

An end-to-end system for running a coffee business: a point-of-sale terminal, an ERP back office, a customer ordering website, and a Telegram Mini App.

• POS and ERP dashboard (Angular + NestJS): a cashier terminal with dual-currency USD/KHR denomination-level cash handling, plus back-office payroll, purchasing, profit-and-loss reporting, and a recipe-costing engine computing live food-cost margins from ingredient stock.
• Customer ordering website (Next.js + NestJS): menu ordering, a stored-value wallet, and a gamified loyalty system of points, rank tiers, missions and collectible stamps.
• Telegram Mini App (TypeScript): conversational ordering, account linking and order notifications.
• Integrated Cambodian QR payment rails (Bakong/KHQR and ABA PayWay) — QR intent generation, polling-based settlement and bank-app deep-linking.
• Integrated OpenAI structured-output generation to assign personalised loyalty badges, validated against a closed catalogue with a deterministic fallback so an AI failure can never break a transaction.
• Implemented an idempotency-key interceptor making payment-mutating endpoints safe against client retries.

• Angular
• NestJS
• Next.js
• TypeScript
• PostgreSQL
• Telegram Mini Apps
• Bakong KHQR
• OpenAI

Save the Children Cambodia

An internal system for managing partnership records and organisational relationships, built around Cambodia's real administrative geography.

• Modelled a four-level Cambodian administrative geography hierarchy — province → district → commune → village — with cascading dependent address selection on the frontend and eager-loaded hierarchical queries on the backend.
• Built multi-file document management, a report submission and review workflow, and role-based access control.
• Containerised the full stack for both development and production.

• Laravel
• Next.js
• PostgreSQL
• Docker