I am a pre-final year CS student at LNMIIT Jaipur.
I co-founded Practers, an AI interviewer that runs live voice at sub-second latency and has passed 1000+ users. Before that I built Sloop at Swyclr, a barter marketplace on iPhone and Android, where I wrote the native module and the config plugin myself.
Most of my time goes on the unglamorous half of AI systems. The evals, the guardrails, the thing that stops an agent doing something expensive while nobody is watching. I have shipped an agent that spends real money on its own, and that taught me fairly quickly that a model being right most of the time is not good enough.
I work in TypeScript and Python, React Native on mobile, and I use Claude Code for most of it.
Anyone can get a demo working. The hard part is knowing in ten seconds that an agent has confidently produced garbage.
Here is what I built so that I would know.
- 13 tools over a self-hosted MCP server that an agent uses to provision live cloud infrastructure.
- Parallel judges score every run, disagreement escalates to a human, and the suite runs in CI so a bad prompt change fails the build.
- The spending cap sits inside the payment mandate where the bank enforces it, so a fully compromised agent still cannot exceed it.
- A long-lived watcher that acts on its own, where the hard part is the arming: a capped, time-bound authorization it cannot exceed unattended.
- For proactive systems the failure mode is noise, not error, so the real work is suppression: act on high confidence, escalate on disagreement, stay quiet otherwise.
- Several plans in parallel, routed by which model is genuinely good at the job, with CI guard scripts that fail a build on policy violations.
|
A real-time AI interviewer. It holds a live voice conversation at sub-second latency over WebSockets, and adapts its questions to the candidate's own resume while the interview is happening. Taken 0 to 1000+ users across 1000+ sessions, with live health monitoring. The hard part was never the interview. It was making the score defensible, which is where the judge harness came from.
Stack Next.js 16, React 19, Fastify 5, Prisma, Supabase, PostgreSQL, MongoDB, Redis, BullMQ, Socket.IO, Judge0 Infra Google Cloud Run, Docker, Cloud Build, Cloudflare R2, Razorpay, Resend, MSG91 |
Verified from the private mockr-labs/practers repository insights.
|
|
5 role-bounded LLM agents on a tool-calling harness, transacting with live merchants over MCP and Shopify UCP. Every pick grounded on a Senso RAG context layer. The mandate agent arms a capped, time-bound authorization once, then buys on its own on a price drop. I shipped an LLM-as-judge harness scoring every tool call on real merchant data, to catch confident-wrong actions before checkout.
Stack OpenAI gpt-4o, MCP / Shopify UCP, Prava, Senso (RAG), Next.js, TypeScript, WebSockets |
An agent that provisions live cloud resources, exposed as 13 tools over a self-hosted MCP server. Since the agent is spending and provisioning on its own, guardrails are policy-as-code. A scope, budget and approval engine checks every action before it runs, with a dry-run mode and a replayable audit trace. No agent can ever exceed its granted authority. Built with Claude Code, custom skills and parallel subagents, full stack in days.
Stack Next.js, TypeScript, MCP (HTTP and Stdio), Socket.io, Node.js |
|
Compiles natural-language specs into deployable REST APIs through an LLM tool-calling pipeline. A node-based workflow engine handles data models, auth, validation and business logic, with an execution runtime and one-click deploy. Its self-hosted MCP server exposes every generated endpoint as an agent-callable tool.
|
Real-time collaborative spreadsheet using Operational Transformation over WebSockets. Persistent multi-user state, reconnect-safe synchronisation. Under 100ms recalculation across 2,500+ cells.
|
|
A market where positions are priced by what people actually believe, not by what they say. Building it teaches you one thing fast: a probability only means something when someone has money on it. 
[Live](https://pridiction-market-frontend-nfrj.vercel.app/)
|
AI shopping assistant on AWS Bedrock and Llama 3. Search, recommendations, personalisation, budget flows and generative UI. |
|
Built on SKALE Network. |
Reputation-weighted voting. Live |
|
A carpool app used on my own campus. Code |
|
Jan 2026 to Jun 2026 · Remote Built Sloop, a barter marketplace for iPhone and Android. Swipe to match items, AI item valuation, hybrid cash-plus-item deals, auction rooms, escrow, location discovery and in-app chat. Wrote a custom native Expo module for Truecaller authentication, with its own Android sources, Gradle build and ProGuard rules. Wrote an Expo config plugin to fix a silent Android 11+ package-visibility bug. The payment SDK ships an empty manifest, so UPI apps were invisible to the checkout and the sandbox VPA passed while production quietly failed. Shipped a real-time swipe matching engine where mutual opt-ins instantly create matches, on Supabase row-level security with live Expo push, released through a GitHub Actions CI/CD pipeline with automated tests and error tracking.
|
Apr 2026 to Present Led architecture for an AI-native interview and hiring platform as a Turborepo monorepo. Built resume-aware AI mock interviews with authenticated WebSockets, Groq voice and text sessions, Gemini and OpenAI integrations, stage orchestration and rubric-scored reports. Built secure coding, contest and hiring modules on Judge0 with Redis and BullMQ queues, hidden tests, real-time verdicts, ATS analysis and OA proctoring. Deployed on Google Cloud Run with Docker, Secret Manager, R2-style storage, Razorpay, Resend and MSG91. |
|
Aug 2025 to Dec 2025 Built scalable REST APIs for automated resume parsing using NLP-based extraction. Cut manual screening effort by 35 percent across recruiter workflows. Developed AI recruitment workflows for candidate matching, ranking, scheduling automation and recruiter decision support. |
Mentored students across two semesters, communicating complex ideas and presenting solutions in labs. B.Tech Communication and Computer Engineering CGPA 7.52 · Aug 2024 to Aug 2028 |
| Project | What I did | Why it counts |
|---|---|---|
| Zulip GSoC org | Onboarded into their large-scale production Django codebase and was assigned by a core maintainer to migrate message-editing banners to a modern shared component. Shipped with passing unit tests and a design proposal. | Code other engineers actually use, in a repo I did not control. PR #39811 |
| AOSSIE GSoC org | Diagnosed and fixed a governance-audit bug in OrgExplorer that counted closed issues and PRs as open. | It was inflating dashboard metrics by up to 12x. PR #123 |
Google Big Code Hackathon 2026: top 1% after three national elimination rounds.
AI and agents · MCP · eval harnesses · LLM-as-judge · multi-agent orchestration · RAG · prompt evaluation · tool-calling
