I'm a B.Tech CSE (AI/ML) student at NIAT, Hyderabad (2025–2029). I build full-stack, AI-integrated products end to end — UI, API, data, deploy — and I ship them live.
I'm looking for internships at product companies and early-stage startups (Hyderabad, remote, or San Francisco).
What I care about: products that are honest about what they do. Saves that say "Saved" only when the server confirms. AI output you review before it lands, with undo. Automation with hard safety limits and a paper trail.
Live demo · Code |
Live demo · Code |
 |
 |
AgentGate — live · Code |
DocuMind — live · Code |
🧠 CogniSpace — how it works
- Editor: TipTap block editor with a
/menu, toggles, to-dos and a selection toolbar. - Copilot: Gemini responses streamed over SSE. You review, insert, and can undo in one click. AI output is sanitized before it is rendered.
- Reliability: autosave only reports Saved after the server confirms; offline retries with backoff; version check (optimistic concurrency) with a "load theirs / keep mine" flow.
- Safety nets: version history with one-click restore, undo for deletes and AI inserts, JSON export, account deletion.
- Not yet: real-time multiplayer — two tabs are detected, but there is no live co-editing.
- Stack: React 19, TypeScript, Vite, Tailwind, Framer Motion, Express 5, MongoDB, Gemini. Vercel + Render.
⚡ RecoverFlow — how it works
- Built for the Razorpay AI Buildathon, Track 03 (AI Revenue Recovery).
- Flow: failed payment → evidence-based diagnosis → policy engine → one bounded action → human approval → audit trail. The AI never moves money on its own.
- Policy codes:
NO_CONSENT,HIGH_VALUE_REVIEW,EXCEPTION_QUEUE,LOW_CONFIDENCE,HUMAN_GATE. - Security: HMAC-verified Razorpay webhooks, idempotent processing, hash-chained audit log, strict CSP, only
public/is served. - Quality: zero runtime dependencies, 40+ passing tests, a 21-scenario held-out evaluation suite, CI on every push.
- Honest limits: demo data is synthetic and Razorpay is Test Mode only.
🚦 AgentGate — how it works
- What it is: a gate between an AI agent and its tools. Every tool call is checked against a JSON policy you can read. Default is deny.
- Flow: agent proposes a call → policy allows, holds for a person, or denies → approved calls run once → every step lands in a hash-chained log that shows if anyone edits it.
- Try it: the public playground on the landing page runs real policy decisions with no sign-in. The dashboard (runs, approvals, policy) runs on Render with Postgres.
- Quality: 75 passing tests; strict CSP, no inline scripts.
- Honest limits: the "world" the agent acts on is simulated.
📄 DocuMind — how it works
- What it is: document question answering where every answer is marked at its source. Answers are extractive, so the text comes from your documents rather than being invented.
- Retrieval: BM25 plus LSA hybrid ranking, written from scratch. Zero runtime dependencies.
- Safety: rate limited (20 requests a minute), 29 passing tests, strict CSP.
- Honest limits: it finds and quotes passages; it does not write new answers.
Also built: Nimbus (AI analytics product site, front-end showcase with simulated data) · Balaji Family Dhaba (live, digital menu site)
| When | What |
|---|---|
| Aug–Sep 2024 | IoT & Robotics Trainee, T-Works — AI-based smart street lights |
| Apr–Jul 2024 | Student Intern, Airbac Labs — machine learning for stock prediction |
| Aug 2024 | Volunteer, Rubaroo — STEM awareness for 100+ students |
If you're a founder or engineer who wants someone who ships, reviews their own work honestly, and learns fast: LinkedIn · upparikarthik505@gmail.com