Applied AI Engineer | Voice Agents | Customer Delivery
Applied AI engineer and founder with 3 years across data and AI roles, delivering customer-facing voice agents, Python/FastAPI tool integrations and production data pipelines.
I came back to India to build for Indian businesses, and founded Brokai Labs, an AI and automation agency.
Brokai's first paying customer was a US salon. An AI voice receptionist for a US salon that answers calls, books appointments and handles rescheduling.
The salon handled about 300 calls and 240 confirmed bookings a month through it (client-reported figures).
Jarvis: A voice assistant that remembers context and gets work done across Slack, Linear, Gmail, Calendar and Notion, so you can ask once instead of switching between five apps.
Résumé ranking: led delivery of a bulk résumé-ingestion and candidate-ranking workflow used to screen 300+ applicants.
Hapi: One API that lets an AI agent take actions in Gmail, Sheets, Slack and PostgreSQL, instead of a separate integration for every app.
Founding AI/ML Engineer, Oaklet Berkeley, CA
Sep–Oct 2025
Oaklet's first AI engineer, in my last few months in the US.
Built a healthcare appointment assistant that booked appointments, reserved calendar slots and sent emails.
Implemented a patient-intake workflow with seven question types in React, Express/TypeScript and DynamoDB, with encryption, audit logging and retry handling.
AI Engineer, Centific Bellevue
May 2024–Aug 2025
Bellevue was my first full-time job after my master's. I moved across three AI teams and learned how these systems are built, tested and shipped. It was a steep period of technical learning.
Audio pipeline on Azure Blob Storage and Label Studio: 6,000+ hours a week across 6+ locales, processing time down 70%, failures down 85%, 99.5% uptime.
Human-in-the-loop speaker-label validation for noisy audio: 95%+ labeling accuracy.
Near-duplicate detection across 10+ languages, including right-to-left and character-based scripts: associated costs down 30%.
Data Scientist (Co-op), 7-Eleven Irving, TX
Jun–Dec 2022
Rebuilt ETL pipelines in PySpark across 10+ sources, improving data accuracy 18% and cutting KPI runtimes 23.7%; automated unsupervised clustering of 1,000+ promotional offers at 98.2% accuracy, saving 1,000+ manual hours a year.
Every one of those promotions ends up as a sign on a counter like this. Grouping them automatically gave people back hours they used to spend sorting offers by hand.
Founder, ANA Packaging & Trading Bengaluru
Jun 2020–Jul 2021
June 2020. I had just finished engineering college, the country was in lockdown, and I started a paper-cup factory. Each cup sold for less than ₹1, so every wasted cup was money gone.
Then the rims started cracking. I adjusted the curling stage, where the cracks showed, iteration after iteration. Nothing worked. The real fault was earlier: the paper was slightly misaligned during folding, long before the damage was visible.
That machine taught me the lesson I still use in every AI system I build: trace the fault through the whole line before you touch the part where it shows.
I also ran inventory and buyer-risk decisions on R/SQL analysis and A/B tests:
Stockouts fell 17%
Payment defaults fell 70%
Order conversion rose 16.2%
Projects
Jarvis
Voice agent · Brokai Labs · a team project I started
A voice assistant that remembers context and gets work done across Slack, Linear, Gmail, Calendar and Notion, so you can ask once instead of switching between five apps.
What I did
I started Jarvis and wrote most of the first voice agent on LiveKit, then brought in collaborators and led the team as it grew. I evaluated it with 30–40 test cases per round, scoring task completion, tool routing and answer relevance.
How it works
You speak; LiveKit streams the voice session
Jarvis recalls what it knows about you and the task
It picks the right workplace tool
It takes the action or answers back by voice
Every run is traced and scored
Proof
Built on LiveKit with persistent memory, tool handlers for Slack, Linear, Gmail, Calendar and Notion, and tracing on every session. Each evaluation round used 30–40 cases.
About the demo
The demo here is a small rule-based Slack-to-Linear example, so you can see the flow without connecting real accounts.
The full system was built with a team, and its code is private.
Hapi
Prototype · my own project
One API that lets an AI agent take actions in Gmail, Sheets, Slack and PostgreSQL, instead of a separate integration for every app.
What I did
I designed and built Hapi, with my team supporting the work. It puts four connector families and 28 actions behind a single execution endpoint.
How it works
The agent sends one authenticated request
Hapi finds the action in its registry
The right connector runs it with stored, encrypted credentials
The reply is trimmed to what the agent needs, and the run is logged
Proof
FastAPI backend with OAuth and API-key sign-in, encrypted credentials, database migrations, configurable response fields, execution logs and a React management screen.
About the demo
Hapi is a prototype, not a product in production.
This site does not connect to any real account.
Salon voice receptionist
Client work · Brokai Labs · first paying customer
An AI voice receptionist for a US salon that answers calls, books appointments and handles rescheduling.
What I did
I took it from the first customer conversation to launch: scoping the calls, building the agent on Retell and rolling it out.
How it works
A client calls the salon
The agent understands the request and collects the details
It checks availability and runs the booking tools
It confirms the booking or change on the call
Proof
The salon handled about 300 calls and 240 confirmed bookings a month through it (client-reported figures).
About the demo
The call on this site uses your browser's voice and a pretend calendar; no real booking is made.
A multi-agent framework that writes legal-style documents with six kinds of planted contradictions, to test whether AI systems can catch them. The demo here is a simplified lease example, not the full framework.
A human-in-the-loop framework for labelling personal data across 13 locales and about 336 PII types, producing high-quality data for fine-tuning LLMs. The detector here is a simple rule-based illustration.
Education
Northeastern University, Boston
Master's in Engineering Management, specialisation in Data Science
PES University, Bangalore
Bachelor's in mechanical engineering, 2016–2020
Background
The lane I grew up on Nadiad
Home
I grew up here.
Santram Mandir Nadiad
The town
It is the town's most famous temple, and I have been going there since I was a child.
The family farm Nadiad
Family
My dad owns a dairy farm here with close to 100 cattle. It is where he is happily spending the next stage of his life.
My first apartment Boston
On my own
In Boston, for the first time, there was no one at home to ask. I was in another country, starting a master's at Northeastern, and suddenly every decision was mine: rent, groceries, what to save and what to spend.
I learned to balance classes with work, stretch a budget, and keep applying until the 7-Eleven co-op came through. It was the year I started running my own life.
The badminton court Seattle & Bellevue
Off hours
Badminton, whenever I get the chance.
24 hours at Y Combinator San Francisco
The place I had dreamed about
For years, Y Combinator was a name I only read about. In my last months in the US, I spent 24 hours inside its building for a hackathon, building through the night in the place I had always dreamed of seeing.
Around it was the Bay Area: Oaklet, where I joined as the first AI engineer, and conversations with founders and investors that changed the scale of my ambitions. I left with a habit I still keep: hold the big goal, and break it into steps I can ship this week.