Built an evaluated AI support agent for @SpotifyCares as the Hiver SDE Intern take-home: one structured Gemini call classifies a tweet into 12 intents, drafts a ≤280-character reply grounded in six BM25-retrieved threads from 27,627 real conversations, and decides whether to auto-handle or escalate — with deterministic money/security/legal/churn rules that can force an escalation the model cannot undo, and a release veto that replaces any draft carrying an uncited claim or unsupported link with a holding reply.

Treated the proof as the product: a 250-tweet golden set labelled twice against a written guide and adjudicated (κ 0.96 intent / 0.95 escalation), four baselines, paired-bootstrap 95% CIs, and a blind two-order LLM judge on a different model — the agent's replies beat the nearest historical reply by +1.42/5 (CI 1.15–1.67) across 200 messages, with the judge's 73.5% order consistency and its blind spots published alongside.

Froze a fresh 200-tweet holdout behind SHA-256 locks — sample, labels, threshold and every source file hashed before inference, and a runner that refuses to execute at any other code or label revision — then measured 0.938 escalation recall, 0.789 macro-F1 and 35% auto-handling; replaying the same 800 cached receipts through repaired guards lifted recall to 97.5% with zero new model calls, reported as a retrospective regression rather than fresh evidence.

Made every number reproducible without an API key: all 1,771 Gemini calls live in a committed SQLite replay cache keyed by model, prompts, schema and temperature, so one command recomputes the headline metrics in about 1.4 seconds on Linux and Windows CI — while the client survived the free tier with pooled-key rotation, per-key sliding-window rate limits, 429 cooldowns parsed from server retry hints and deadline-bounded retries.

Deployed on Vercel as a slim Python serverless function (FastAPI, BM25 index rebuilt from the committed corpus at cold start, two-slot concurrency, sanitized errors, no visitor prompts persisted) behind a React dashboard with a live playground, bring-your-own-Gemini-key support held only in sessionStorage, and HMAC-signed HttpOnly admin sessions gating the internal evaluation, golden-set and failure-mode pages.

Built Parakh (परख), a serverless AI exam checker: upload a question paper and a student's handwritten answer sheet (plus an optional marking scheme) and it extracts every question, finds and highlights each answer's exact ink region on the sheet, grades it, and writes per-question feedback with a teacher summary — no auth, no database, nothing stored beyond the browser session, and a one-click sample exam to try the whole flow.

Kept the backend stateless by rendering documents in the browser with pdf.js — up to 15 pages as 1200px JPEGs with the text layer extracted alongside — so Vercel functions only ever receive compact images; digital papers take a text-only fast path while scanned ones fall back to page-image vision.

Replaced the original local pipeline (DeepSeek-OCR-2 grounding on an RTX 4060 plus GPT-5.6 via the Codex CLI, behind a FastAPI worker, polling job store and Cloudflare tunnel) with two Gemini structured-output calls: one extracts questions with sub-parts split, unprinted marks estimated and OR-choice groups tagged; one multimodal call reads every answer page at once and returns per-answer bounding boxes on a 0–1000 grid, scores and feedback.

Wrote a raw-fetch Gemini client with free-tier key rotation: any number of keys round-robin, advancing on 429/500/503 and stepping down a model ladder before failing loudly, so several free quotas pool into one.

Handled messy scripts in post-processing — boxes clamped and merged when vertically adjacent, scores capped at max marks, skipped OR alternatives shown as "OR — skipped" with each choice-set counted once, stray writing surfaced as "Unmatched answers" — and drew percentage-positioned overlays that track the ink at 50–200% zoom, with invisible hitboxes so clicking an answer on the sheet selects its question.

Built a fully local video anomaly detector in a one-day hackathon that runs on a single 8 GB RTX 4060 laptop GPU: a trained head on frozen CLIP ViT-B/32 features scores footage at 2 FPS, a LoRA-adapted Qwen3-VL-2B verifies scene context on a bounded cadence, and per-class temporal state machines turn repeated evidence into timestamped events — zero hosted model calls at runtime.

Placed #1 on the AHC Visual Intelligence Hackathon live leaderboard with 67.8 points, 2.3 clear of second place, processing 47.3 minutes of footage in 9.55 minutes — 4.96× real time including model loading, inference, refinement and explanation.

Trained a 1.61M-parameter rank-8 LoRA on Qwen3-VL-2B in 240 updates, supervising only the 12 answer-letter logits so inference is a single constrained-choice forward pass — no JSON generation and zero parse errors across 932 VLM calls — lifting balanced-holdout macro F1 from 29.5% to 40.3%.

Recovered the hardest long-video class without touching the weights: YOLO11n vehicle proposals, ORB/RANSAC camera-motion compensation and occlusion-tolerant background-relative tracking flagged a truck stationary across 22 samples, which the frozen VLM confirmed as a highway stop — lifting the long-context score from 11.3 to 20.5 with no new false alarms.

Shipped a React 19 + Vite review dashboard over a range-request Python server with click-to-seek event timelines and per-class score histories, plus a leak-proof pipeline — content-hash-grouped holdout, offline Hugging Face cache, and an exporter that rejects inconsistent runtime metadata — covered by 36 passing tests.

Built a WebGPU studio that grows any link or message into a 3D voxel tree whose canopy encodes a working QR code — click the tree and it flattens into a top-down view, captured straight off the GPU canvas as a PNG that ordinary phone cameras can scan.

Fixed dark-mode scannability without redesigning the code: the dark theme now decodes reliably on real phone cameras while keeping the tree-derived artwork intact — no black-and-white fallback, no painted frame, and the QR payload, matrix, and finder geometry left untouched.

Shipped four seasonal themes with custom leaf palettes, manually selected light/dark modes with gradual synchronized transitions, seasonal ambient audio, and a responsive desktop/mobile UI — with the chosen leaf color held visually consistent across both themes.

Added QR image decoding by upload, paste, or drag-and-drop with jsQR, so codes can be read back into the studio as well as generated from it.

Tamed a fragile build: the upstream WebGPU engine is pinned to a commit and patched by an ordered chain of source-rewriting scripts, with clean dependency reinstalls enforced on Vercel so cached patched modules can never leak between deployments.

Built a multi-user task tracker with a Next.js 16 (React 19) frontend and an Express + TypeScript REST API on Neon PostgreSQL via Sequelize: owner-scoped workspaces and tasks (title, description, status, priority, due date, position) shown in Board, List, Table and Timeline views.

Implemented secure auth: 6-digit email OTP verification (10-minute validity) before the first sign-in, bcrypt password hashing, 7-day JWT sessions and password reset via a one-hour emailed link — plus Zod request validation, Helmet headers and per-IP rate limiting (100 requests per 15 minutes).

Built the drag-and-drop Kanban with dnd-kit and persisted every move through one batch endpoint that reorders tasks inside a single database transaction; a freshly verified account is seeded with a Getting Started workspace and sample tasks.

Deployed as a single Vercel project: the compiled Express app is exported as one Vercel Function behind an /api rewrite, with idempotent database initialisation shared by the local server and the function, and Jest + Supertest tests run against a Postgres test database with email delivery mocked.

Built an AI-powered Chrome Extension and Node.js backend that learns a user's unique writing style to generate hyper-personalized, context-aware email drafts using Google Gemini and LanceDB.

Integrated Google Gemini's LLM via the API to analyze sent emails, dynamically extracting communication patterns and building distinct writing personas (professional, casual, etc.).

Implemented RAG using LanceDB vector database for semantic search over past email threads, enabling the AI to understand ongoing conversation context for accurate draft generation.

Integrated Gmail API with secure OAuth 2.0 flows within a privacy-first architecture where data processing stays local, and built an injected Gmail UI layer with keyboard shortcuts and a multi-persona editor.

Built a low-cost, AI-powered digital telepathology solution that detects sickle cells from blood smear images using a fine-tuned ResNet50 deep learning model, achieving >92% accuracy.

Reduced early screening costs to approximately ₹75 per test — a 95% cost reduction compared to traditional laboratory methods.

Developed an intuitive Streamlit web application providing actionable health guidelines, precautions, and emergency helpline contacts based on AI-generated results.

Enabled telepathology and telemedicine support for frontline health workers in rural areas on low-power edge devices. Secured 3rd place ($1,000) at University of Miami Horizon AI Hackathon 2025 and recognized among Top 100 Startups at IIT Delhi Youth Ideathon.

Developing software for automated road extraction using CNNs on ISRO's Resourcesat images from the Boonidhi portal.

Built a GUI for specifying areas of interest and generating geographically referenced shapefiles, with email alerts for road changes based on image comparisons.

Optimized for efficient processing of large satellite datasets.

Developing ERP Exerciser, a Brain-Computer Interface (BCI) application that leverages Event-Related Potentials (ERPs) to deliver real-time cognitive training, enhancing memory, attention, and executive functions in patients with cognitive impairments.

Developed an extremely efficient machine-learning application for upscaling and colorizing images, with a model size of just 260 MB and achieving 89% color accuracy in standardized tests (using CNNs and DinkNet).

Potential use cases include military applications, night-vision enhancement, and restoring old photographs.