Aviroop Paul

Builder / Full Stack + AI. I ship AI the hard way: engineered, evaluated, and tested until it holds up in production.

Right now.

I build backends and AI pipelines, and I spend my spare time finding out what Opus 5.5 can do next.

The day jobs.

Recruit41: Software Development Engineer II

Apr 2026 - Sep 2026, Bengaluru, India. End to end voice pipeline for running 200+ concurrent AI interviews.

  • Resume Screener (1000+ resumes a batch): Shipped in a week: upload a batch, get a ranked shortlist against the recruiter's own criteria.
  • Interview proctoring (<3% false positives): Rebuilt the gaze detector as an audio + video fusion scorer with a Gemini analysis pass.

Think41: Software Development Engineer II

Jul 2024 - Apr 2026, Bengaluru, India. Founding engineer, parachuted into Series A startups to get their AI out of demo mode.

    Atomicwork

    Series A, $25M. Launched Atom, their Universal AI Coworker.

    • Universal AI Coworker (POC → production): Agentic Chrome extension that puts Atom on any web page. Now the platform's moat.
    • Voice and vision: Brought LiveKit voice calls and live screen reading to Atom, so it can see what the user sees, and made both run inside a browser extension.

    SpotDraft

    Series A, $50M. Led a 5-engineer team on contract intelligence and VerifAI.

    • Clause compliance (65% → 85%): Replaced chunked RAG with full-document context.
    • Durable evals (1000+ golden data points): Randomized evals in CI on every prompt or model change.
    • Obligation extraction (100+ page contracts): LangChain extraction over full contracts that unlocked new client deals.

    Nokia: Software Development Intern

    Sept 2023 - Jun 2024, Bengaluru, India. First paycheck. Backend for Nokia's internal platforms.

    • Django REST APIs (2x operational efficiency): Multi-tenant APIs that automated complex clone operations.
    • Feature work: Turned customer asks into shipped features, with Docker and Kubernetes CI/CD.
    • Performance review (Top intern): Their words, not mine.

    Builds

    Weekend builds that got out of hand.

    bleep

    Mac app, 2026. Your agents, unblocked. Your inbox, handled. A little droid that floats on your Mac. It clears the safe asks from your Claude Code and Codex sessions, reads your mail and calendar, and taps you only when something needs you.

    • Approves tests, lint and reads; never a force push
    • Meeting prep, due bills and waiting replies, before you ask
    • Runs on the Claude plan you already pay for, or your own key

    Built with Rust, Tauri, React, Claude Code.

    WhisprCatch

    Mac + Linux app, 2026. You talk. It types. Hold a key, say the thing, let go. Punctuated text lands wherever your cursor is, in any app. The speech model runs on your own laptop, so your voice never leaves it.

    • ~25x faster than real time, on a plain CPU
    • No cloud, no account, no subscription
    • The free, open-source Wispr Flow alternative

    Built with Rust, ONNX Runtime, Parakeet.

    agentconf

    CLI + local web UI, 2026. Your coding agents have 538 settings. You've set 32. One command scans ~/.claude, ~/.codex, ~/.cursor and ~/.gemini, then shows every setting you've set and every one you haven't, each explained and linked to the docs.

    • One Go binary, scans in milliseconds
    • Read-only: never writes a byte, never opens a secret
    • Cross-agent checkups with agentconf doctor

    Built with Go, TypeScript, JSON Schema.

    Travelust

    Travel agent web app, 2026. Find the flight. Plan the days. Go. Tell it the trip in your own words. It finds live fares, asks how you like to travel, and plans your days with maps. Swap anything, then export a PDF to print and share.

    • Live Google Flights fares, read directly with no paid API
    • Day plans from Wikivoyage, Wikipedia and Google Search
    • Every choice stays in the chat: the agent pauses and waits for you

    Built with LangGraph, Gemini, FastAPI, React.

    ChaosLab

    System design playground, 2026. Build a system. Break it. Learn why. Drag servers, caches and databases onto a whiteboard, then crank the load from 10 users to 500 million. Queueing math decides, live, what holds up and what melts.

    • Queueing math, not dice: M/M/1 latency and cascading failures
    • Netflix and Instagram presets, built from their engineering blogs
    • Autoscaling, retries and circuit breakers you can tune

    Built with Next.js, TypeScript, React Flow, Three.js.

    Daily drivers.

    Proof of work.

    2451 contributions since January 2026.

    Say hello.

    Bengaluru, India.