Low Kai Hon

Github: https://www.twine.net/signin
● Shipped a conversational CSV analytics web app (Next.js 15, Vercel AI SDK 5, Supabase/Postgres, Vega-Lite) that plans analyses, runs SQL, generates charts, and outputs editable/exportable Markdown reports.
● Built an agentic loop with two modes—Focused Q&A and Deep-Dive—that executes a multi-step, tool-calling plan with prompt customisation; returns a concise summary plus expandable detailed analysis.
● Designed a reference-based tool chain (executeSQLQuery → queryId → createChart) to minimise tokens and keep results traceable; a Charts tab stores visuals linked to the exact SQL, with side-by-side chart/SQL inspection.
● Added reliability & safety: SELECT-only with auto-LIMIT, timeouts, retries/fix suggestions, clear error surfacing; CSV→Postgres pipeline with type inference, parameterised queries, and transaction-wrapped batch loads.

Github: https://www.twine.net/signin
● Built an adaptive RAG system that auto-selects retrieval strategies, self-corrects, and detects hallucinations—achieving significant retrieval accuracy gains using only budget models (GPT-4o-mini).
● Agentic architecture: 7-node StateGraph with distributed routing; self-correction loops for retrieval (query rewrite, early strategy switch) and generation (HHEM hallucination verification + LLM-as-judge quality check).
● Retrieval & reranking: semantic/keyword/hybrid with two-stage rerank (CrossEncoder→top-10 → LLM-as-judge→top-4) and RRF multi-query fusion; research-backed patterns (CRAG, PreQRAG, RAG-Fusion, vRAG-Eval).
● Evaluation Framework: 4-tier architecture comparison (Basic → Multi-Agent) with curated golden datasets; comprehensive metrics (F1@K, MRR, nDCG) measuring retrieval gains from architectural improvements.