Skip to main content
AI Engineer

Hi, I'm Kunj Patel

I build AI systems that actually work in production — from RAG pipelines that answer legal questions with citation-level accuracy to autonomous agents that diagnose and fix infrastructure failures without human intervention.

See My Work Get In Touch
3+ Systems Shipped
80% OCR Accuracy
31% RAG Gain (IEEE)
Kunj Patel — AI Engineer

About Me

I'm an AI engineer focused on the hard parts of deploying language models: making retrieval actually relevant, making agents actually reliable, and making the whole system actually scale.

RAG & Retrieval Systems

Semantic search, FAISS, vector databases, embedding optimization, HyDE, knowledge grounding, RAG pipeline design

Cloud & Infrastructure

GCP Cloud Run, Vertex AI, AWS SageMaker, AWS Bedrock, Docker, CI/CD pipelines, serverless deployment

Experience & Background

Where I've applied my craft

AI Engineer

Vypar TaxOne

Chartered accountants deal with thousands of pages of GST circulars, notifications, and legal amendments — documents that are dense, frequently updated, and genuinely hard to navigate quickly. I was brought in to build an AI system that could answer compliance questions directly from this corpus.

  • Built the GST Compliance Assistant: a domain-specific RAG pipeline with long-term memory via MongoDB, deployed serverlessly on Google Cloud Run with auto-scaling for unpredictable query loads.
  • Rebuilt the OCR-to-LLM pipeline from scratch — scanned GST documents had poor OCR quality, which made chunking and embedding unreliable. Increased extraction accuracy from 40% to 80%.
  • Benchmarked four next-gen RAG architectures — HyDE, CAG, KAG, and Agentic RAG — against our baseline to identify which retrieval strategies held up on legal text specifically.
RAG LLM Google Cloud Run MongoDB

Research Author

IEEE ICCES 2024 — RAG-Enhanced LLM for Web-Based Assistance

Built a hybrid retrieval pipeline combining LLaMA 3, Gemini, and FAISS to study how RAG improves LLM accuracy on web-grounded question answering. The system achieved a 31% improvement in contextual accuracy over non-RAG baselines. Published at IEEE ICCES 2024.

LLaMA 3 FAISS IEEE Published

B.E. Computer Science & Design

A.D. Patel Institute of Technology — CGPA: 8.86 / 10

T&P Coordinator Google Cloud Certified AWS GenAI Certified

Technical Skills

Organized by capability, not just keywords

Agentic AI & Orchestration

Google ADK · LangGraph · LangChain · CrewAI · DSPy · MCP · A2A Protocols · Multi-Step Planning

Retrieval & Search

FAISS · Vector Databases · Semantic Search · Embedding Optimization · HyDE · CAG · Knowledge Grounding

Cloud & Infrastructure

GCP: Cloud Run, Vertex AI · AWS: SageMaker, Bedrock · Docker · CI/CD Pipelines · Serverless

Core Stack

Python · FastAPI · MongoDB · LLM APIs (Gemini, GPT, LLaMA) · OCR Pipelines · Git

Featured Projects

Context: problem → solution → stack

Featured Project

Sentinel — Autonomous SRE Agent

Production incidents usually follow a predictable pattern: something breaks, an on-call engineer gets paged, they SSH into the server, read logs, form a hypothesis, and take action. Sentinel automates that loop. It connects to Docker via MCP for sandboxed access, pulls logs, and passes them to a reasoning layer built with DSPy — not raw prompting, but compiled few-shot examples tuned specifically for root-cause analysis. The DSPy compilation step is what makes this reliable enough for production use.

Google ADK DSPy Gemini 2.5 MCP FastAPI Docker SDK

Multi-Agent Code Auditor

Manual code reviews are slow and inconsistent under deadline pressure. Three agents work in sequence: an Orchestrator routes files by risk, an Auditor scans for vulnerabilities via custom A2A protocols, and a Planner synthesizes a prioritized remediation report. File access runs through MCP, keeping agents sandboxed.

Google ADK LangGraph Gemini 2.5 MCP

Sales Transcript Analysis Agent

Analyses renewal call transcripts and produces structured reports — intent scores, key takeaways, actionable recommendations — using LangChain Agents and Google Gemini. Built for sales team to process Hindi/Hinglish transcripts.

LangChain Gemini 2.5 Sales Intelligence
Available for new projects

Get in touch

Let's work on something hard.