22 projectsranked 1 to 225 areas
All work
One line-up, strongest first: 9 landmarks, 5 stations and 8 waypoints across research, software, hardware, hackathons, leadership. Every project shows how far it got and what is still unfinished.
- Live3DLive 3D reconstruction from any cameraA server that takes video from a phone, a drone or a CCTV stream and builds a 3D model of the space live in the browser, then a refined model in real meters that can be measured, walked through and exported. It runs on one Apple Silicon workstation, with no cloud GPU.34 of 36benchmark datasets with a live delay under one second
Progress70%
- LawVerseLegal NLP research and a platform for Indian lawA legal research platform for Indian law, and the study behind it. Across 231,908 court judgments, our first pipeline said 22 of 25 legal concepts had changed in meaning. After we found and fixed four silent defects in it, one did.22 → 1of 25 concepts read as changing, before and after the pipeline was corrected
Progress75%
- One Camera BrainTurns any existing camera into a smart cameraOne system that takes video from cameras a site already owns, runs theft, safety and work-insight rules on it, and seals evidence clips so that tampering shows. Tested on recorded public datasets, not yet on real recorder hardware.98% / 98%precision and recall for after-hours intrusion on a shopping-corridor dataset (CAVIAR)
Progress75%
- AI E (AI Employees)Role-specific AI agents whose work is checked before it is certifiedFive AI employees for finance, data analysis, HR, QA and DevOps on one shared runtime. The language model only drafts: numbers and rules come from deterministic engines, and an answer is marked certified only when those checks pass and a reviewer model from a different family agrees.246 / 250analyst questions answered correctly by Vega, the data analyst (98.4%)
Progress70%
- Ego-contingent trajectory predictionWhat-if motion forecasting on the nuScenes driving datasetAn LSTM model trained on the nuScenes driving dataset that predicts how a nearby road user responds to what the ego vehicle plans to do: carry on, brake or accelerate. Written up as an IEEE-format paper, peer-reviewed (three reviews), on which I am a co-author.0.449 m/saverage displacement error after retraining on all 850 scenes
Progress75%
- PAMOLocal guardrails for AI coding agentsA macOS daemon and native app that sits between me and terminal AI coding agents. It lets routine actions through, holds destructive ones for a human, and logs every decision. In three weeks of logged use it made 3,699 decisions and held 258.3,699permission decisions made and logged
Progress75%
- TextegrityAI-text detection, robustness testing and plagiarism analysisA content-integrity pipeline: an ensemble of eight detectors for AI-written text, a back-translation stage that tests how easily detection breaks, and a plagiarism check. Its own evaluation shows 97.86% accuracy on its training distribution and 55.7% on text from newer models.97.86%accuracy of the four-model meta-classifier on 2,099 held-out rows
Progress70%
- VidyaTidNEET exam-prep platform, pausedAn exam-preparation platform for India's NEET entrance exam: a tutor that answers from NCERT textbook material, a cleaned bank of 39,385 past questions in English and Hindi, and timed mock tests. Rebuilt between January and September 2026; not deployed, and paused since.39,385questions in the cleaned, publishable bank
Progress75%
- ARI SyncReal-time, local-first AI coaching engineA coaching engine that turns an exercise video into a live session: it watches you through a webcam, compares your form with the reference, and speaks a correction while you are still moving. Every model runs on one local machine.9joint angles checked on every pose sample
- Auto Image EnhancerBlind ×4 image super-resolution with a web app around itA self-trained ×4 super-resolution model for blurry, noisy or compressed photos, behind a web app with credits, payments and job queues. The first stage is trained and served; the planned diffusion stage was written but never trained.21.38 dBvalidation PSNR after 200 epochs of blind ×4 training (SSIM 0.613)
Progress60%
- Raitha MitraFarming assistant with leaf-disease detectionA web app for farmers, built by a team of four. Photograph a leaf, and a small convolutional network I trained names one of 38 disease classes across 14 crops; the app then writes its advice in Kannada or English. It is live. Its accuracy has never been measured.38disease classes across 14 crops
Progress75%
- CreeperX robot and mainboardQuadruped robot and a custom four-layer controller boardA 12-servo quadruped robot from a university course, and the controller board I designed afterwards to replace its tangle of breakout modules: four layers, 142 components and three independent 5 A power rails. The board design is complete and checked; it has not been fabricated yet.142components on 89 nets, on a four-layer 100 × 100 mm board
Progress55%
- LumeEdge / SafeEdgeIoT security device with a 3D monitoring dashboardAn ESP32 security device with a cloud pipeline for anomaly detection and incident summaries, and a dashboard with a 3D floor plan and an attack simulator. A team project built for Microsoft Imagine Cup 2026. The model was trained on synthetic data, most sensor readings are simulated, and the Azure deployment is unverified.1 of 4sensor readings comes from real hardware: motion
Progress55%
- CitiZen AICivic complaint platformResidents report local problems such as potholes, garbage and water leaks with photos and map pins, and municipal staff work the queue with deadlines, escalation and analytics. The full stack runs; the urgency model that is meant to rank the queue is effectively untrained.15API routers in the FastAPI backend, over 18 services
Progress65%
- FlakyTestXFlaky-test detectorA developer tool that reruns a pytest suite many times, scores how flaky each test is, and charts the result. Built in a day for the Kyiv QA Automation Hackathon 2025; on its own sample suite it flagged the 1 flaky test among 13.1 of 13tests flagged as flaky in the saved run
Progress65%
- MSARSERacing lap analysis in 30-meter micro-sectorsSplits a racing lap into micro-sectors of about 30 meters and compares laps sector by sector to show where a driver loses time. A team hackathon project on Toyota GR Cup data from Circuit of the Americas; the track positions are synthetic and the results are not yet meaningful.184micro-sectors of about 30 meters on the lap, as the app shows
Progress60%
- Smart Mirror Workout AssistantPose-based rep counting for a mirror displayA webcam app that tracks your pose, recognizes five exercises with joint-angle rules, counts repetitions and gives form tips. It was meant to sit behind a two-way mirror on a Raspberry Pi. The app runs on a laptop; the mirror build is not documented, and rep-count accuracy was never measured.5exercises recognized by joint-angle rules: squat, push-up, jumping jack, lunge, bicep curl
Progress60%
- Doc2QA APIDocument to question-and-answer REST APIA REST API that turns PDF, DOCX or plain-text documents into study question-and-answer pairs in four modes, with API keys, plan tiers, usage metering and rate limits. The questions come from rules and heuristics, self-hosted, with no outside AI service.27 of 105tests failing at the last recorded run
Progress70%
- MLX HackathonOrganizer of an online machine-learning hackathonI created and ran MLX (Machine Learning Xtreme), an online machine-learning hackathon for students on Devpost in February 2025, with three co-managers. The public page showed 255 participants in November 2026, and its gallery holds 31 submitted projects.255participants
Progress80%
- OpenFundsCrowdfunding app from an Agile course projectA Streamlit crowdfunding app where creators post campaigns with a Bitcoin testnet address. A small, untested model estimates a campaign's chance of success, and sentence embeddings flag near-duplicates. A four-person course project, later the subject of a university-guided internship report.Not measuredaccuracy of the campaign success predictor
Progress55%
- CyberTriage AICyber-fraud triage as an MCP server and an RL environmentTriage for Indian cyber-fraud complaints: classify the fraud type, score urgency, and route the case to the right authority. Built for two hackathons, once as a config-driven MCP server with 14 tools and once as an OpenEnv reinforcement-learning environment with three graded tasks. Classification is rule-based, and no run was saved.14tools in the MCP server, with 3 resources and 3 prompts
Progress70%
- RC airplane (Boeing aeromodelling)Team entry in a national aeromodelling competitionA team entry in the zonal round of the Boeing National Aeromodelling Competition at Kshitij, IIT Kharagpur, in January 2024. Participation, not a placing; no build or flight records are on file.Participationzonal round, 20 and 21 January 2024
Progress50%