I'm an AI engineer and full-stack developer, in my final year of Computer Engineering at FCRIT, Navi Mumbai. Most of what I build these days has an LLM in it: agents that call tools and change real data, inside TypeScript, Next.js and Postgres apps.
I also build with AI: Claude Code and Codex are part of my daily workflow, and I work across Claude, GPT and Gemini models depending on the job.
- 🤖 Model-agnostic agents on the Vercel AI SDK, with the limits that keep them safe: quotas, step caps, human approval
- 🚢 Research intern at IISER Mohali (2025-26): fuel-consumption models for a Kamsarmax bulk carrier on ~83k rows of real telemetry, XGBoost at R² 0.98
- 🦀 Rust services (Axum, Tokio) and Tauri desktop apps
- ⚙️ GitHub Actions release pipelines, signed auto-updates, Docker deploys
- 📬 Co-founded LetterStack with Krish Pinto: open source, with paying customers
- 🧩 Built LYX UI, a shadcn-based React component library
- 🎨 UI/UX design in Figma, as a hobby
- 🎧 Off the keyboard: music and anime
- Agents inside products. LetterStack's email editor has an agent that edits the document through typed tools. Tool calls run in the browser so edits are instant, while the server keeps the API key, per-org quotas, a usage ledger and a hard step ceiling.
- Tool-calling agents with memory. eve-agent-framework: a model-agnostic agent on Vercel's Eve framework where each new skill is one typed tool file, with notes persisted in Neon Postgres, usable from a web UI, an HTTP API or a terminal REPL.
- Durable agent workflows. Career OS (in progress): multi-step agent runs on Workflow DevKit that survive restarts and can be replayed, with an approval gate before anything is sent to the outside world.
- LLM features that ship. AI roadmap generation in Blingo, AI health summaries from uploaded prescriptions in HealthSync, and an AI schema visualiser with Upstash rate limiting.
- Classic ML. At IISER Mohali I modelled a Kamsarmax bulk carrier's fuel use from real telemetry: XGBoost hit R² 0.98, and synthetic data cut its error by 17%. Served over FastAPI as shipfuel-ai. Also PyTorch/Keras demand forecasting explained with SHAP and LIME.
- Agentic dev workflow. Claude Code and Codex every day. Typeset even writes LAYOUT.md index files so agents can find their way around a notes workspace.
| Project | What it is | Stack |
|---|---|---|
| LetterStack | Open-source email campaign SaaS I co-founded (~72k lines, paying customers). I built the AI editing agent and Razorpay invoicing (repo) | Next.js · AI SDK · Neon · AWS SES · Razorpay |
| eve-agent-framework | Extensible, model-agnostic agent with persistent memory | Eve · AI SDK · Neon |
| Typeset | Windows Markdown app; 5 signed, auto-updating releases via GitHub Actions | Tauri · Rust · Next.js · CodeMirror |
| Moss | Overleaf-style LaTeX editor with a Rust compile service on Docker | Next.js · Supabase · Rust (Axum) · Docker |
| LYX UI | shadcn/ui-based component library + docs | Next.js · Fumadocs · Motion |
| Blingo | AI project-roadmap generator for hackathon teams | Next.js · Gemini · Appwrite · React Flow |
| shipfuel-ai | Ship fuel-consumption prediction from the IISER research (R² 0.98) | FastAPI · XGBoost · Next.js |
Languages
Frontend
Backend & systems
Data
Desktop & mobile
AI & ML
CI/CD & infra
-232F3E.svg?logo=false&variant=branded&size=xs){kind=icon}
- Tagged releases → signed desktop builds. In Typeset, pushing a
v*tag runs a GitHub Actions workflow that sets up Node and the Rust toolchain, builds the Tauri app on Windows, signs it for the auto-updater with keys kept in repo secrets, and publishes a GitHub Release that installed copies update from. - Containerised Rust service. Moss runs a LaTeX compile backend written in Rust (Axum + Tokio). It ships as a Docker image on Render as infrastructure-as-code (
render.yaml), with health checks, upload limits and compile timeouts. - Deploy automation. Vercel preview and production deploys on every project, plus a GitHub Actions deploy hook that works around Vercel's single-owner build limit on team repos.
- Quality gates. TypeScript strict mode, ESLint/Biome, Vitest + Testing Library, Zod at API boundaries.
