Experience and the projects I've built, most involved first.
Experience
AI Center of Excellence, CIT
Oct 2025 – Dec 2025
AI Engineering Trainee
Built retrieval-augmented and ML-powered applications end to end, from OCR and embedding pipelines to production-style APIs and dashboards.
—Developed IntelliLearn, a RAG study assistant answering questions over user-uploaded notes and PDFs using OCR, sentence-transformer embeddings, FAISS vector search, and LangChain, served through FastAPI.
—Worked on PhishShield, a phishing-detection dashboard in React, TypeScript, and Tailwind backed by Python ML models.
—Integrated ML backends with frontends and shipped hackathon prototypes with a team.
Chennai Institute of Technology
2024 – 2028
B.E. Computer Science (AI & ML)
Undergraduate program focused on AI/ML, currently at a 9.08 CGPA.
—Coursework foundation in DSA, OOP, DBMS, operating systems, computer networks, and distributed systems.
—Applied that coursework to real systems — from distributed databases to on-device AI.
A distributed transactional key-value store written from scratch in Rust — no raft-rs, no RocksDB. Hand-rolled LSM storage engine with WAL and SSTable compaction, Raft consensus per range shard, MVCC, and Percolator-style 2PC, with per-table tunable consistency: CP tables get Snapshot Isolation, AP tables take a leaderless W=1 path resolved by hybrid logical clocks. Validated under real failure — leader kills, SIGKILL recovery, and a bank-transfer invariant across 170+ tests with zero acknowledged-write loss.
rust
distributed-systems
raft
lsm-tree
mvcc
Distributed transaction under failure
The client opens a cross-shard transaction and the Timestamp Oracle issues its snapshot timestamp.
An authorization layer that lets an AI agent spend real money without ever being trusted to. The agent proposes; deterministic code decides. Merchant-defined per-transaction limits, vendor allowlists, and velocity caps live in SQL rather than in the prompt, so a compromised or persuasive model has nothing to argue with. A database-enforced idempotency key makes execution exactly-once even under replay, and an append-only audit log guarded by SQL triggers means no entry can be altered or deleted after the fact. A self-built adversarial suite attacking the live API defended 14 of 14 attacks — replay, price tampering, race conditions, compromised-LLM bypass — with zero unintended charges.
agentic-ai
ai-safety
fastapi
payments
sqlite
An agent tries to spend money
The agent asks to buy something. Asking is all it is ever allowed to do.
An LLM-powered interview evaluator that scores communication and technical clarity as two separate dimensions. Speech-to-text and audio analytics run concurrently, delivery metrics (pacing, filler words, structure) are computed alongside the model's technical judgment, and the twin returns targeted coaching feedback on both.
llm
rag
nlp
fastapi
One answer through the pipeline
The candidate answers a system-design question by voice.
A real-time navigation assistant for visually impaired users that runs entirely on-device. Object detection and metric depth estimation run in parallel on each camera frame, objects are placed into a 3D scene graph and tracked across frames, and a small language model turns that spatial state into spoken navigation cues.
computer-vision
yolo
depth-estimation
edge-ai
A frame becomes a voice cue
A new frame arrives from the chest-mounted camera.
An industrial IoT monitoring platform for small workshops that can't afford unplanned downtime. ESP32 nodes publish vibration, temperature, and current telemetry over MQTT; an Isolation Forest flags anomalous windows against each machine's own baseline, and Firebase pushes the alert to the floor manager's phone before a fault becomes a breakdown.
iot
anomaly-detection
mqtt
firebase
From vibration spike to phone alert
An ESP32 on a lathe publishes its telemetry window.
A RAG study assistant that answers strictly from a student's own material. Scanned PDFs are read through OCR, chunked and embedded into FAISS, and retrieved passages ground every LLM answer — so the tool cites the notes instead of hallucinating around them.