Vijaykumar Saravanan

↳ manual effort − 60% · Rock Star Award

A hybrid-search + LLM system that classifies products at scale, with a human-in-the-loop UI for cluster sampling and index training.

problem ·

Categorizing huge product catalogs by hand was slow and inconsistent.

approach ·

Hybrid search + LLM classification on Vertex AI, FastAPI and PostgreSQL, plus a review tool that keeps humans in the loop.

result ·

Cut manual categorization effort by 60%; earned a company Rock Star Award.

• Vertex AI
• LLMs
• FastAPI
• PostgreSQL

↳ document-grounded · multi-tenant · zero-install

A branded AI support assistant that turns a business's own documents into a grounded chatbot on Telegram and WhatsApp — no app, no sign-up, just message the business.

problem ·

Generic support bots hallucinate, and answers stay buried in PDFs, manuals and policy docs.

approach ·

Multi-tenant RAG on Google Cloud with per-business Pinecone namespaces and Gemini, answering only from curated docs and refusing when coverage is missing.

result ·

Self-serve support grounded in each company's own content, with isolated vector namespaces and role-based admin curation.

• Google Cloud
• Pinecone
• Gemini
• Telegram
• WhatsApp

↳ hybrid beats lexical & semantic alone

A production-grade product search engine that fuses lexical and semantic retrieval, built on a medallion data pipeline and evaluated with an LLM judge.

problem ·

Keyword search misses meaning; pure embeddings return irrelevant hits — each mode alone under-serves real queries.

approach ·

Bronze/Silver/Gold layers in Databricks (Spark + Delta) feed offline BM25 and FAISS indexes, fused 65/35 semantic-to-lexical and served over FastAPI.

result ·

Gemini-as-judge win-rate scoring shows hybrid retrieval consistently beats single-method BM25 or FAISS.

• Python
• FastAPI
• FAISS
• BM25
• Databricks
• Spark

A smart fleet and transport management system, built to make running a fleet simple. Still in build — the product site is up at fleetezy.in.