Bangalore, IN
Sep 2025 – Present
AI-native email and CRM, backed by a16z. Seed-stage, about ten engineers. These four systems are mine, end to end.
Serving ~100 → 1,400 MAU
Every inbound email gets read once and turned into structure: sender type, status, labels, and a reply draft waiting in the composer. Micro's labels sync back into Gmail, so the classification is visible wherever you read mail. The hard part isn't the model — it's that Needs Response is a Gmail label anyone can toggle, so work is armed on every observation and collapsed by a debounced queue rather than run at ingest.
400K
messages / mo through the queue I built
Micro
$1M+
bookings through the flow I owned
Pilot
Software Engineer II
Owned the trip planner and worked on a hotel booking flow that processed $1M+ in bookings. Delivered Smart Import end to end and helped migrate the platform from Firebase to MongoDB.
Product Engineer II
One of three engineers on DrCuro, a HIPAA-compliant EHR — scheduling, consultation records, prescriptions, plus a WhatsApp booking flow. Integrated ABHA, India's national health record network, straight against the official APIs instead of through a reseller. Clinics trialled it and it never launched; all of it argued out before agentic coding existed.
Senior Developer, contract
Built the ECG annotation tool a diagnostics company's technicians worked in daily, and the pipeline that normalized their labelled output for an in-house arrhythmia model. Laid the front end and the data flow the rest of the team built on. Twenty technicians used it every day.
Founding engineer
First technical hire. Built the app, hired the team, and rewrote the product when COVID killed the offline-first version it depended on. The market wasn't ready for the replacement and we shut it down — but I had hired and led engineers before I'd finished my diploma.
Earlier: Roundups (2021) · ShapeAI (2021) · Flabber (first internship)
A conformance benchmark for agent tool-call authorization. It runs adversarial calls through each framework's real approval machinery to test argument scope, delegated authority, and escalation quality. Early design stage.
Pulse watches my Garmin data and nudges me on Telegram when my routine slips; months of using it on myself turned a simple notifier into a health-coaching system, with every false alarm or missed intervention shaping the next version.
Slack-native SEO agent — drafts, rank tracking, GEO citation monitoring. Private beta.
Leaderboards, badges, and XP as an API and SDK, so an app can add progression without designing the schema for it. My first go at shipping something for other developers to build on.
Every first version I've shipped was too confident.
A context builder treated missing data as negative data. The model filled the gaps with a confident explanation and turned two observations into a pattern. I fixed the input, then required every claimed pattern to cite its sample size. The pipeline had erased uncertainty before the prompt.
Workers claim due jobs with a lease. If one crashes, the same query that claims new work also picks up expired jobs. Recovery happens during normal operation, removing the need for a separate reaper process that can fail or fall behind.
Hit me up if you need an engineer to help figure out the system and own what ships, whether you're starting from a blank repo or improving something people already depend on.
devu.nm21@gmail.com