I engineer production AI systems and resilient backend architectures using Python, LLMs, RAG, semantic search and high-throughput document pipelines — from FastAPI services and vector retrieval to multi-stage LLM workflows running over millions of records.
24from fastapi.routing import APIRouter27router = APIRouter()30@router.post("/audit")31async def audit(url: str) -> AuditReport:32 # deterministic pass — 80% of the final score33 page = await fetch(url, timeout=5)34 dom = parse(page.html, strict=False)37 technical = score_technical(dom)38 entities = score_jsonld(dom)41 if llm_enabled:42 sem = await judge.run(dom, fallback=heuristic)45 return AuditReport.merge(technical, entities, sem)
Specialized in production AI systems, vector retrieval pipelines, and backend reliability.
ALL MR ONLINE (NearLaw)
At NearLaw, I engineer production AI pipelines for an enterprise legal search platform managing 18M+ court records. My daily work revolves around document ingestion, OCR, semantic indexing, and multi-stage LLM evaluation workflows.
I take pride in solving the unglamorous problems that make AI reliable in production: messy real-world PDFs, OCR edge cases, structured extraction accuracy, latency optimization, and deterministic safety checks.
LLM-powered systems, agentic workflows, deterministic routing, and tool-using AI applications.
Semantic search, hybrid BM25 + dense retrieval, custom embeddings, and document-grounded generation.
FastAPI services, asynchronous pipelines, task queues, and data-intensive backend architectures.
OCR, PDF structure parsing, semantic chunking, and tabular data extraction pipelines.
Quantized models (GGUF, llama.cpp), low-latency offline voice assistants, and hardware-constrained inference.
Engineering production AI systems, asynchronous pipelines, and high-performance backend platforms.
ALL MR ONLINE (NearLaw)
Utkarsha Interiors
Campus Credentials
Production AI systems, agentic workflows and retrieval pipelines — with the architecture behind each one.
A FastAPI application that audits a website and scores how visible it is to AI answer engines such as ChatGPT Search, Google AI Overviews, Perplexity and Claude — with evidence-backed findings and copy-paste fixes.
A multi-agent healthcare assistant built around patient-data isolation, deterministic emergency handling and tool-based workflows for medical information, patient history and appointments.
An AI support agent that combines retrieval, tool calling and human-in-the-loop workflows to resolve customer issues and handle actions such as refunds and cancellations.
A multi-agent learning platform that creates personalized study plans, retrieves learning resources, generates quizzes and provides performance feedback.
A local AI voice assistant combining speech recognition, LLM inference, text-to-speech and semantic retrieval, designed to run on consumer hardware.
Numbers taken from production work and shipped systems — each one verified and benchmarked.
NearLaw 18M+ pipeline · LangGraph evaluation runs · Utkarsha MVC gains · GTX 1650 4GB local inference
Core technologies used to architect, build, and deploy production systems.
A one-page summary of my experience building production AI systems, backend services and AI products.
Current Role
AI Engineer Intern
NearLaw (ALL MR ONLINE) · Feb 2026 – Present
Education
B.Tech in Computer Engineering
Mumbai University (CGPA: 8.68/10)
Certifications
Oracle Cloud & AI Associate
CS50 Python, LangChain & Vector Databases
Frequently asked questions about how I build, what I'm learning and what I do outside of code — two files, ten answers. Open one to read it.
I enjoy problems where AI has to work reliably inside a real backend system, not just in a demo. I focus on LLM workflows, RAG, semantic search with vector databases, and building resilient API pipelines that handle real-world edge cases.
Anantum is an offline AI voice assistant I built to explore what is possible with local AI on limited hardware. It pushed me to work with quantized GGUF models using llama.cpp, low-latency vector search with FAISS, and speech recognition using Whisper, without relying on external cloud APIs for its core AI workflow.
Working with real-world legal documents at NearLaw taught me that data ingestion and document quality can be harder than the AI itself. Dealing with OCR failures, unusual PDF formats and extraction issues has pushed me to design more resilient processing pipelines with fallback mechanisms instead of assuming every document will be clean.
I am currently exploring two main areas:
I start by understanding the requirements, defining data boundaries, and identifying state transitions and edge cases. I prefer decoupled architectures and consider approaches like Redis caching, asynchronous worker queues and structured schemas where they fit the problem, helping systems handle failures without affecting the entire pipeline.
Outside of software, I enjoy reading and following geopolitics and global economics. I am fascinated by how macroeconomic shifts, trade policies and global supply chains affect technology adoption and emerging markets.
I am curious about how ideas turn into actual products. I like understanding not just how something is built, but why it should exist, who would use it, and what makes it sustainable over time.
I enjoy discussing ideas, challenging assumptions early, and figuring things out together. I like environments where people openly discuss architectural trade-offs, share what they know and learn from different perspectives.
I am working on connecting technical design with product outcomes. I want to understand how engineering decisions, from latency targets to database schemas, affect user experience, customer needs and business goals.
I would probably be building something of my own, perhaps around product strategy or market analysis. I enjoy understanding complex systems, spotting opportunities and turning ideas into something useful, and I would want to do that regardless of the medium.
Have a problem worth building? I am always interested in interesting AI, software and product problems.