PickWise v2 — AI Product Recommendation Platform
LangGraph ReAct Agent, 4-Stage RAG Pipeline & Next.js 16 Interface
Jun 2026 – Present
FastAPI
LangGraph
LangChain
Google Gemini
pgvector
PostgreSQL
Docker Compose
Next.js 16 (React)
TypeScript
Playwright
- Multi-Step Agentic Reasoning: Built a production ReAct agent with 4 custom tools (semantic search, review evidence / YouTube review retrieval, price calculation, live market lookup) querying PostgreSQL directly, streaming responses over SSE and deciding per turn whether to retrieve or answer from conversation context.
- Calibrated 4-Stage RAG Pipeline: Engineered high-accuracy personalized retrieval (pgvector top-50 recall → reranking → constraint relaxation → relevance gating) with 768-dimension Gemini embeddings on PostgreSQL, calibrated at a 0.53 threshold so low-confidence queries return an interactive clarifying question instead of hallucinated recommendations. Evaluated and logged per request via NDCG@10.
- Bilingual Eval Harness: Implemented an automated evaluation framework combining LLM-as-judge and deterministic rule checks that intercepted agent market price fabrications across conversation turns; established a 3-run variance baseline to separate true regressions from model noise.
- Automated Data Ingestion: Built resilient Playwright scrapers with Gemini-powered extraction and combinatorial variant generation to ingest raw HTML into AI-ready records across 270+ SKUs (Apple, ASUS, Acer), fully containerized via Docker Compose.
- Explainable Scoring Engine: Designed an 8-factor deterministic scoring engine with transparent explainability breakdown per recommendation, diagnosing and correcting a normalization defect that had ranked items against unrepresentative catalog extremes.