I build whole products: the frontend people use, the backend and database behind it, the deployment that keeps it running, and the guardrails that stop an AI agent from doing something it shouldn't. Every project below is live in production, tested in CI, and published with its numbers and the caveat next to them.
- 🧱 Full stack: 7 apps live today, from Next.js and React frontends to FastAPI and Express backends on PostgreSQL, deployed on Azure and Render
- 🛡️ Guardrails: 0 of 240 prompt-injection attacks moved money through SELLABLE's deterministic gate
- 🔬 Research: first author of Errors Are Not Choke Points, selected for the XERJ Corpus Hub
- 🏆 Winner, CodeX 3.0 Hack Arena · 2x Smart India Hackathon (college round)
- 🎓 B.Tech CSE (AI & ML), KIET Group of Institutions, 2028 · open to full-stack, AI and LLM engineering internships
| Layer | What I build | Where you can see it |
|---|---|---|
| 🎨 Frontend | Next.js 15 and React apps with 3D maps (MapLibre, deck.gl), a voice-driven UI, and ML that runs in the browser (ONNX, MediaPipe WASM) | StrataIQ · ProctorIQ · Abhed |
| ⚙️ Backend | FastAPI and Express services, REST APIs, JWT auth, crash-safe payment state machines, PostgreSQL/PostGIS, Supabase, SQLite, MongoDB | SELLABLE · StrataIQ · ExpenseTracker |
| 🚀 Deployment & ops | Docker Compose on an Azure Ubuntu VM behind Caddy (automatic HTTPS, security headers, zero-error restarts), Render services, GitHub Actions CI, end-to-end and live smoke tests, database backups, load tests | StrataIQ on Azure · SELLABLE CI |
| 🛡️ Guardrails | Deterministic policy gates for agent tool calls, prompt-injection defence, human approval bound to a hash of the exact action, signed tamper-evident ledgers | SELLABLE · Abhed |
| 🤖 AI & agents | Remote MCP servers, tool calling, voice agents, RAG, LLM evaluation, where the LLM proposes and tested code decides | SELLABLE MCP · Research |
A shopping agent that structurally cannot spend your money. The LLM may only propose a SKU and a quantity; a 12-rule deterministic, fail-closed gateway decides whether money moves, at p95 < 5 ms. Remote MCP server for Claude, ChatGPT, Cursor and VS Code.
Live MCP server · Code · PyPI |
One gate for every action an AI agent takes. A rulebook checks each tool call, risky ones go to a person's phone, and every decision lands on a hash-chained, Ed25519-signed ledger. I owned AI security and backend in a team of five.
Live app · Code · MCP server |
Test a shop location before you sign the lease. 10,000 synthetic customers, each with a 615-parameter genome grounded in official census data, choose between you, real competitors and not buying. A voice agent drives the app; the LLM never makes a number.
|
Exam-integrity analytics where video never leaves the device. MediaPipe finds 478 face landmarks in the browser, a 27 KB ONNX model classifies attention in under 15 ms, and the server only sees signed events.
|
Where is an LLM's failed reasoning actually decided? Resampling from every prefix shows a verifier finds 73% of human-labelled errors but only 45% of decisive steps, and plain restart beats repair at matched compute. Released as ChokeBench.
|
From 2,629 real YouTube channels to personalised, human-reviewed outreach: explainable filters, a trained niche classifier, an authenticity score, Gemini-written messages and n8n + Gmail sending with crash-safe dedupe.
|
More projects
| Project | What it is | Links |
|---|---|---|
| Fraud scoring API | XGBoost on a leakage-safe temporal split: PR-AUC 0.793 on unseen future transactions, and an honest note that it doesn't beat logistic regression | Live API · Code |
| abhed-mcp | Zero-dependency MCP server that puts any agent's tool calls behind Abhed's gate | Code |
| RAG pipeline | Medical document RAG with hybrid search, reranking and page-level citations | Code |
| ExpenseTracker | React 19 + Express 5 + MongoDB finance tracker with JWT auth, bcrypt, Helmet and user-scoped CRUD | Code |
| DarshanEase | React + Express + MongoDB temple booking platform with user, organiser and admin flows | Code |
| NLP labs | Self-attention, tokenisation and transformer fine-tuning from scratch | Code |
| Deep learning essentials | CNNs, RNNs, backprop and optimisers implemented in PyTorch | Code |
| Number | What it measures | Caveat |
|---|---|---|
| 0 / 240 | attacks that moved money through SELLABLE, vs 13% with a standard pay tool | three open-weight models, 864 calls, 95% CI 0–2% |
| 578 | automated tests in SELLABLE (invariants, security, chaos, concurrency) | concurrency is still single-node SQLite |
| 23 / 26 | attacks blocked by Abhed's SDK path, 0 of 18 normal actions wrongly blocked | the no-code gateway let 6 of 21 through |
| 45% vs 73% | decisive reasoning steps vs human-labelled errors a verifier finds | ChokeBench v1, 230 failed traces, exploratory |
| 27 KB | ProctorIQ's attention model, running in the browser | 99.2% on 30 acted clips from one webcam |
| 0.793 | fraud PR-AUC on a temporal test split | not better than logistic regression |
Frontend: Next.js · React · TypeScript · Tailwind · MapLibre / deck.gl · ONNX Runtime Web
Backend: FastAPI · Express · REST · JWT auth · PostgreSQL / PostGIS · Supabase · MongoDB
Deploy: Docker · Caddy · Azure VM · Render · GitHub Actions · Playwright end-to-end tests
Guardrails & agents: MCP servers · tool calling · guardrails · prompt-injection defence · human-in-the-loop approval · OpenAI / Gemini / Groq APIs
ML: PyTorch · XGBoost · ONNX · MediaPipe · process reward models · calibration · bootstrap CIs · adversarial red-teaming
- 🥇 Winner, CodeX 3.0 Hack Arena, with Abhed (team of five)
- 🥇 Winner, Smart India Hackathon, college round (2x) · Winner, Innotech (department)
- 📚 Research selected for the XERJ Corpus Hub as a reference for AI agents
- 💳 Built SELLABLE for the Razorpay AI Buildathon · ☁️ AWS Certified Cloud Practitioner
Hiring a full-stack or AI engineer intern? [email protected] · LinkedIn · Portfolio