Projects
A picked list, not an archive. Case studies or build logs where the work earned one, and the repo for every project.
Featured
One drone pass in, a georeferenced, textured 3D model out, on one laptop GPU. A synthetic ten-minute clip ran in 548 s of a 900 s budget, at 0.344 m median error with 98.5% within 1 m of ground truth.
Featured
Four-drone swarm, simulated in PX4 and Gazebo, that surveys a disaster zone beyond line of sight. Kill the relay mid-mission and the swarm elects a new one: route home back in 38.9 s, 120 of 120 cells covered, 5,917 of 5,917 packets delivered. Built for IIT Bombay Techfest's PUSHPAK Grand Challenge and selected for Stage 2, one of 14 teams from 448 registrations.
Featured
Hindi and English voice question answering over 950K passages. Hybrid dense and BM25 retrieval, under 200 ms P90 from transcript to answer in both languages, and speculative decoding cut English generation p50 from 84.7 ms to 43.5 ms.
Featured
Evaluation-first support call intelligence: ASR, diarization, and event detection benchmarked on 60 calls with exact ground-truth labels.
Vision-language model trained from scratch that reaches 50.3% image-to-text recall@1 on COCO, beating ViT-Tiny CLIP baselines.
Turns a brain-dump of avoided admin chores into ready-to-send messages that escalate on their own when ignored. The model drafts, plain code validates, nothing sends without a tap. Built for OpenAI Build Week 2026.
Benchmark for Indian-language document RAG across Hindi, Marathi, Tamil, and Gujarati, designed to run from the scanned page to grounded answers. The first retrieval baseline is built and measured; hybrid retrieval, OCR and generation come next.
Continual-learning research. Per-task LoRA on a 4-bit Qwen3-14B kept task 1 at 75.3% after training task 2; v3 tests whether a gate on incoming facts keeps poisoned edits out of a model.
Deepfake detection and cryptographic media provenance pipeline for sports media, deployed on Google Cloud.
Civic issue reporting app that uses Gemini Vision to detect and categorize problems from a single photo.
Virtual stock trading platform that replays 14 trading days of synthetic NSE prices one 30-minute candle at a time, with limit orders, real brokerage charges and 103 tests.
RAG tax assistant for Indian CAs, grounded in the Income Tax Act 2025 with a rule engine for regime and HRA math.
Claude Code plugin that runs a project plan one week per tick and only accepts exit codes and greps as proof, never the agent's own report.
Mental health risk detection you can read, trained on a 20,000-post sample of a 1.85M-post Reddit dataset: 76.83% accuracy, 85.15% ROC-AUC, SHAP and LIME behind every flag.
Quant study linking Hyperliquid trader performance to the Bitcoin Fear and Greed index. Extreme Greed was this cohort's best regime, at an 89.2% win rate.
Adaptive testing engine that uses Item Response Theory to fit each question to a candidate's live ability estimate.