/projects$cat cloudofgoods-insight-agent.html
CloudofGoods · Product Insight Agent
iLabs · 2025 · architected, built & deployed end-to-end
the problem
CloudofGoods is a rental marketplace. Shoppers needed real-time answers — what does this cost for my dates? what else should I rent with it? — and those answers had to be grounded in a product catalog living in the org's data lakehouse, not hallucinated.
what i built
An AI assistant on FastAPI, powered by Google ADK and the Gemini APIs, handling live price calculations and site-wide product recommendations. The main challenge was the data path, not the model.
I engineered an optimized streaming data pipeline from the lakehouse via Trino into a Milvus vector database, so the agent retrieves over fresh, embedded product data instead of round-tripping to the warehouse on every query. That single change cut end-to-end inference latency by ~75%.
making it observable
A model you can't see is a model you can't trust in production. I stood up a Streamlit + Opik dashboard for full-lifecycle LLM observability — traces, latencies, and quality signals — so we could debug and improve the agent with evidence instead of guesses.