Beana AI.

Key impact

• —Cut AI search latency from 20+ seconds to ~3–4 seconds — profiled the full retrieval and ranking path and fixed the bottlenecks (to 6–7s), then parallelized tool calling to get under 4 seconds.
• —Raised RAG pipeline accuracy from ~30% to 90%+ on QA test suites validating filters and LLM outputs.
• —Diagnosed a core sharp image-library memory issue — freed fragments never returned to the OS — and fixed it via an alternative memory allocator: ~99% of idle memory reclaimed, unlocking unlimited parallel mockup generations (validated with 500 concurrent).
• —Catalog of 100k+ items behind the AI search, in production with real users.

What I built

• —Owned the Chat, Generations, and Accuracy areas: structured the frontend architecture end-to-end and built out parts of the NestJS backend.
• —Built production LLM features on Node.js / NestJS: intent-based AI search with parallel tool calling, multi-turn conversations, and post-search LLM ranking.
• —Worked hands-on with OpenAI and Google Gemini models — prompt engineering, model selection, and structured outputs across search, ranking, and generation features.
• —Implemented RAG retrieval over Elasticsearch + PostgreSQL with Redis caching, plus data-enrichment pipelines on AWS (SQS queues, Lambda functions) feeding the search index.
• —Built eval cases and harnesses so search and ranking quality stays measurable across every model and prompt change.
• —Delivered image-generation and agentic workflows; profiled and optimized heap and request-path performance.

Stack & focus