About
I build production AI systems, from the model and LLM pipelines to the backend and frontend around them.
I'm an AI/ML engineer with 3+ years building production LLM, RAG, and NLP systems end to end. Currently lead AI engineer at an AI-powered education platform, where I designed an LLM-based automated evaluation service, semantic search and deduplication with vector embeddings, and gRPC microservices.
I'm comfortable across the full stack, from Python and Go to TypeScript, moving from model and LLM pipelines to backend services and React frontends without dropping the thread. I've delivered generative-AI solutions for clients across education, automation, and data.
B.Tech in Artificial Intelligence & Data Science (Uka Tarsadia University). Based in Surat, India, working remotely.
Selected Work
Built and deployed an AI agent that scores inbound leads hot, warm, or cold with visible reasoning, drafts personalized follow-up emails from the lead's own words, sends through Gmail after human approval, and logs every decision to an append-only audit trail. Prompt-injection attempts are detected and filtered as spam.
A RAG chatbot that answers strictly from indexed company documents, cites the exact section behind every answer, refuses questions the docs cannot support instead of hallucinating, and flags prompt-injection attempts. FastAPI, ChromaDB, and Gemini in a single Docker deployable.
An invoice processing pipeline where Gemini reads PDF invoices directly and seven deterministic checks validate the extraction: line-item math, totals, tax, dates, and required fields. Flagged invoices land in a review dashboard where a human corrects and approves before anything reaches the books, with CSV export of approved records.
A triage system that reads every inbound customer message, classifies intent, sentiment, and urgency with AI, and escalates urgent items through deterministic rules, optionally straight to Slack. Ships with a live ops dashboard showing sentiment trends and intent breakdowns, plus prompt-injection defense that kept a 'mark me VIP urgent' attack at urgency zero.
Architected an LLM-based service that automatically grades free-text student answers and generates feedback, replacing manual grading at scale for an education platform.
Designed a two-stage pipeline combining lexical (trigram) and semantic (pgvector + OpenAI embeddings) matching to keep large question banks clean and non-redundant.
Built a Go backend (pgx / PostgreSQL) exposing HTTP + gRPC microservices across six domains (auth, curriculum, batches, assignments, reports, and admin) with JWT auth, OTP/SMS onboarding, and S3 presigned uploads.
Built an internal automation bot on the Claude Agent SDK, integrating Slack and Linear to auto-draft pull requests and improve team velocity.
Built a Retrieval-Augmented Generation chatbot to retrieve and interact with information across multiple documents, with a Streamlit interface and Flask backend for seamless interaction.
Developed a multi-class email classification model using pre-trained BERT, improving categorization accuracy to ~90% with optimized preprocessing for large-scale datasets.