Spec each change first: what it must do and the acceptance test that proves it.
Let coding agents draft the implementation.
Verify against acceptance tests and evaluations before it ships.
| Project | What it is | Stack | Status |
|---|---|---|---|
| Notebird | AI study assistant: summaries, generated quizzes, vocabulary games, flashcards, and read-aloud, organized in Learning Nests. Best UI/UX Hack at CruzHacks 2025. | React, Supabase, REST APIs, LLM integration | Retired; site still up at notebird.netlify.app |
| Pulse | Autonomous agent that tracked geopolitical news, threaded it into narratives, mapped them to a stock portfolio, and graded its own predictions. Ran unattended for three months, then retired to cut hosting costs. | Next.js, TypeScript, Supabase, Groq, Tavily, Resend, Vercel cron | Retired; site still up at getpulse.world |
| PM Club at UCSC site | Recruitment site for UC Santa Cruz's first product management club, with a full application flow: resume upload to Vercel Blob and an Airtable row per applicant. Source is private. | HTML, CSS, JavaScript, Vercel serverless function | Racing; live at productatucsc.org |
| Portfolio site | Personal portfolio: a static React app compiled in the browser, no build step. | React, Babel Standalone, CSS, Vercel | Racing; live at tarunchandra.vercel.app |
| Stale-state data generator | Algoverse: deterministic generator for the 5,000-example stale-state evaluation set, with schema invariants and an invariant test suite. | Python (standard library), pytest | Racing |
| J-Lens stale-state probe | Algoverse: Jacobian-lens probing pipeline for Qwen2.5-7B-Instruct. Finds the correct state encoded internally in 137 of 137 stale-output failures. | Python, PyTorch, transformers, Colab A100 | Racing |
| CX-LLM-Bench | Benchmark behind the first-author paper: open-weights and proprietary LLMs on intent classification and summarization for privacy-constrained contact centers. | Python, paired bootstrap CIs, six model APIs | Racing |
| Stealth Startup filesystem agent | Stealth Startup's first reference agent: OpenAI Agents SDK plus the MCP filesystem server, hardened for sandboxed runs. | Python, OpenAI Agents SDK, MCP, Node | Racing |
| Stealth Startup GitHub triage agent | PydanticAI agent on GitHub's hosted MCP endpoint that proposes closable issues as structured output. | Python, PydanticAI, GitHub MCP, GitHub Models | Racing |
Devpost profile: devpost.com/n-tarun2022
| Project | What it is | Hackathon | Team | Links |
|---|---|---|---|---|
| NoteBird | Study buddy that organizes notes, auto-creates quizzes, simplifies tough texts, and summarizes long readings. Winner, Best UI/UX Hack. | CruzHacks 2025 | 2 | Devpost, site |
| QaBot | Autonomous QA agent that clicks through a site, submits forms, finds breakage, and writes a clean bug report. | AGI Inc x OpenAI x Lovable REAL Agent Challenge, November 2025 | 3 | Devpost, repo |
| Clue_0 | Home security agent on NVIDIA Nemotron that watches every frame, reasons about intent, and alerts only when something matters. | Hack-a-Claw x NVIDIA, May 2026 | solo | Devpost, repo |
| Roost | Budgeting app that tracks receipts and gamifies saving by feeding a virtual dragon on smart financial decisions. | Google AI Edge x Qualcomm Developer Hackathon, May 2026 | 4 | Devpost, pitch site repo |
| ACL-Smith | Drop in a Cisco ACL; Claude Opus 4.7 hypothesizes, traces packets through the rules, and finds the doors left open. | Synthesis Hacks, April 2026 | solo | Devpost |
| CanvasCal | Turns Canvas assignments and syllabi into a calendar that students manage by talking to an AI agent. | CruzHacks 2026 | 4 | Devpost |
| MediLink | Healthcare app built to make care clear, connected, and easy for patients and providers. | September 2025 | 4 | Devpost |
| Clientsight | Edit a website by looking at a UI element and speaking or typing a command: eye tracking plus voice. | October 2025 | 3 | Devpost |
First author, under review at the NeurIPS 2026 Workshop on AI Privacy (InfPriv): Closing the Compliance Gap: Benchmarking Open-Weights and Proprietary Large Language Models for Privacy-Constrained Contact Center Automation.
Co-author, under review at the NeurIPS 2026 Trustworthy AI for Good (AI4GOOD) Workshop: One Carrier Isn't Enough: Redundancy-Aware Concept Suppression in Diffusion Transformers.
