Projects

Hackathon Project

A security firewall in front of LLM applications that inspects every prompt and response for prompt injection, encoded payloads and data leaks. Eight detectors run in three tiers (unicode and encoding pre-processing, regex and YARA rules, ML classifiers), and a YAML policy engine decides to allow, sanitize, challenge or block in real time. Ships with a live operator console, a feedback loop for true/false positives, and benchmarks reporting precision, recall, F1 and false-positive rate.

• Built a bidirectional LLM security gateway that screens every prompt and response through 8 detectors across 3 tiers (unicode and encoding pre-processing, regex and YARA rules, ML classifiers), mapped to OWASP LLM Top 10 risks.

Public Benefit Project

A cloud-based platform that lets hospitals exchange patient data securely and in real time through a unified API layer, cutting through the fragmentation of siloed hospital record systems. Built with role-based access control and a scalable patient identity system so multiple institutions can coordinate care without compromising privacy.

• Built a cloud-based healthcare interoperability platform enabling secure real-time hospital data exchange using APIs.

Hackathon Project

A hybrid blockchain platform for fighting counterfeit goods across the supply chain: manufacturers mint an NFT for each product, distributors log shipment checkpoints on-chain, and consumers scan a secure QR code to instantly verify a product's authenticity and full chain of custody.

Hackathon Project

Physics-constrained deep learning that reconstructs subsurface ocean temperature and salinity at 18 depths down to 2,000 m from satellite surface data. Trained on 8,340 Argo profiles matched to satellite passes over the Indian Ocean (2021-2023) and tested on a held-out Bay of Bengal region and a held-out monsoon season. Its 80% and 95% uncertainty bands cover 81% and 95% of observations; results are published honestly, including where it only matches the baseline.

Razorpay AI Buildathon

A reconciliation agent that proves each payment-ledger mismatch with a sandboxed, re-executable script hashed into an append-only chain, checks that discrepancies conserve across the whole batch, then fuzzes its own matching logic. On a 60-record test batch it reached precision and recall of 1.0 at about 3,700 records/sec and caught a systematic INR 175.87 drift.

Personal Project

A webcam-driven file manager: pinch to grab an on-screen file icon, drag it onto a folder or the trash zone, and — after a cancellable safety countdown — the real file moves on disk. Gesture recognition is rule-based (pixel-distance state machines over MediaPipe hand landmarks), no classifiers or training involved.

Hackathon Project

A temporal knowledge graph built from thousands of merged pull requests that shows who actually knows a module, why the code is the way it is, and how much unwritten knowledge walks out when a maintainer leaves. Forks the graph to run what-if scenarios, such as a contributor stepping away or a decision being reversed.

Hackathon Project

A phone watches an expert's hands (or rides on the tool), finds the moments no manual explains, asks the expert why, and turns each answer into a card that coaches apprentices at the same point in the work. Hand tracking runs in the browser; the pipeline is craft-agnostic.

Hackathon Project

An event-driven RAG engine for live voice and chat that starts retrieval before the user finishes speaking, splits compound requests into parallel sub-queries, and patches individual claims when a late constraint arrives instead of restarting. Citation fabrication is structurally impossible, and every decision is traced as JSON.

Personal Project

Career history stored as a queryable graph, with resumes generated from it. Code deterministically selects which positions, bullets and skills appear; the model only writes the phrasing, and a skill with no evidence cannot be claimed.

Personal Project

A local-first Android app with no accounts, no cloud and no permissions: timetable, per-course attendance with miss-budget gauges, habits, to-dos, money tracking and learning, all stored on the device.