Where I've worked
Wipro Technologies
Aug 2022 – PresentAutomation Engineer
My main project was Citibank's compliance automation framework — a system that tracks and manages high-risk tools and processes across the bank to meet regulatory standards. I designed the workflow systems, approval chains, data structures, and audit trails that made the whole lifecycle run.
The piece I'm most proud of: I noticed the existing process had a lot of unnecessary back-and-forth between systems, so I redesigned the data integration layer to eliminate it.
Projects
Personal Project
LoanIQ — Intelligent Loan Approval Assistant
Loan decisions aren't just about a credit score — they involve income stability, debt ratios, regulatory compliance, and edge cases that break simple rules. LoanIQ automates that entire evaluation with a pipeline of four specialised AI agents, each responsible for one layer.
An Applicant Profile Agent, Financial Risk Agent, Loan Decision Agent, and Compliance Agent work in sequence — each backed by its own MCP server. The output is a structured decision (Approve / Reject / Manual Review) with a risk score, confidence level, and a compliance case ID.
AWS-Wipro Codeathon
Revenue Leakage Finder
Telecom companies lose money every day to billing gaps they don't even know exist — usage that never got billed, pricing rules applied wrong, payments that failed silently. This project builds an agentic system that hunts those anomalies automatically.
I designed the architecture — an S3-first data lake with EMR/Spark for batch processing, Redshift Serverless for analytics, and Aurora PostgreSQL for transactional data, all provisioned with Terraform — and used Amazon Kiro to guide the build. On top of that sits a multi-agent setup on Bedrock where different agents handle reconciliation, reasoning, and reporting.
Personal Project
Food Recommendation System
I wanted to understand how vector search actually works, so I built something I'd use: a food recommendation engine where you can ask questions like "what's nutritionally similar to chicken breast?" or "what gives me more iron?" — and get real answers, not keyword matches.
The search layer uses embeddings and ChromaDB. The Q&A layer is a RAG pipeline over a nutritional knowledge base — so answers come with sources, not hallucinations.
Personal Project
Indian CA RAG
Indian tax and compliance questions have real, citable answers buried across the Income Tax Department, ICAI, and GST Council sites — but finding them means digging through PDFs and circulars. This is a RAG API that answers those questions from a knowledge base you control.
PDFs and web pages get chunked and embedded with Voyage AI into ChromaDB; a FastAPI endpoint retrieves the relevant passages and has Claude answer in a CA persona, with sources attached.
Personal Project
AI Meeting Notes Assistant
Meeting recordings are useless if nobody writes up what was actually decided. This tool takes an audio recording and turns it into structured meeting minutes and a task list — no manual note-taking.
Whisper transcribes the audio, an LLM pass on watsonx.ai cleans up financial and product terminology the transcript gets wrong, and the result is served through a simple Gradio interface.
Personal Project
YouTube Video Summarizer
Long videos, short attention span — this app pulls a YouTube video's transcript and lets you summarize it or ask questions about it instead of watching the whole thing.
Summaries run the full transcript through an IBM watsonx.ai foundation model; Q&A chunks and embeds the transcript, indexes it with FAISS, and retrieves only the relevant passages before answering — all through a Gradio interface.
Just for Fun
Wild Tails
Not every project needs a business case. This one's a living habitat simulation — an Indian wolf, a Bengal tiger, an Asiatic lion and a cobra, each roaming its own habitat across the whole page. A behaviour loop runs them through patrolling, stalking, hunting, feeding and resting on their own, weighted by species.
They're wild, and they behave like it: touch one and it pins its ears, warns you off and leaves. Every animal, plant and habitat is hand-drawn SVG, stride rate is derived from real ground speed so the feet don't skate, and the habitat sound is synthesised with the Web Audio API — no image or audio assets, no libraries, no backend.
Personal Project
Ask Amit — In-Browser RAG Assistant
A small AI assistant on this site that answers questions about me — but it never generates text. It embeds your question with a tiny model running entirely in your browser, matches it against a fixed set of passages taken straight from this site, and shows the closest one verbatim. Off-topic questions get a plain "I don't know," never a guess.
The knowledge base isn't a separate file I have to remember to update — a script parses this site's own HTML and re-embeds it, so editing a project description here is the only maintenance the assistant ever needs. No server, no API key, no LLM generation: same question in, same answer out, every time.
Technical skills
Certifications
RAG for Generative AI Applications Specialization
IBM
Generative AI Learning Plan – Developers (Partner)
Amazon Web Services
Microsoft Certified: Azure Data Fundamentals
Microsoft
Microsoft Certified: Azure Fundamentals
Microsoft
Want the full picture?
Here's my resume.