Projects

A retrieval-augmented chatbot for a community of Singaporeans abroad (the Singapore Global Network), answering local questions from 115,000+ WhatsApp messages across seven group chats spanning 2020 to 2026. The pipeline parses and chunks the chat history, embeds it with Voyage AI into Supabase pgvector, retrieves the nearest passages by cosine similarity, and has Claude Haiku answer with attribution. A three-tier confidence system decides when the bot answers, adds a caveat, or defers to the group, favoring accuracy over coverage to avoid hallucinations. Built to weigh a custom RAG stack against an off-the-shelf tool, and to learn where the effort in RAG goes: chunking, embeddings, indexing, and confidence calibration.

A production-deployed 3-agent system for automated B2B lead generation. Agent 1 discovers ICP-matching companies using Exa's semantic search (findSimilar). Agent 2 runs a 4-phase enrichment pipeline: company profiling, contact discovery, industry engagement mapping, and qualification scoring against a weighted rubric. Agent 3 drafts personalized outreach per decision-maker. Results are presented in a Next.js dashboard with HubSpot-style UI: filterable lead table, sliding detail panel, inline CRM editing. Deployed on Vercel with Supabase backend. Demonstrates multi-agent orchestration, search strategy design, qualification as explainability (not filtering), and production infrastructure decisions (Streamlit → Next.js, JSON → Supabase).

An automated morning news digest that uses the Claude Agent SDK to orchestrate content fetching, curation, and delivery. The agent reads from 6+ sources (RSS, IMAP, web search), makes editorial decisions about relevance and theme, and sends a curated email every morning at 7am. Includes a deterministic fallback pipeline for production reliability.

The earlier iterations of the daily digest project. v1 is a simple fetch-summarize-send pipeline using free LLMs. v1.5 adds a curation layer with relevance scoring and a user profile config. Built to understand the progression from deterministic to agentic architectures.

A production-grade evaluation system for AI agents using LLM-as-a-judge methodology. Features an 8-dimension rubric, Pearson correlation calibration to align judge scores with human taste (r=0.72), automated scoring pipeline via GitHub Actions, and a Streamlit dashboard for score tracking. Demonstrates task-specific eval design, rubric calibration techniques, and why evaluation costs more than generation.