Mohnish Bangaru — AI Engineer

About This Human

I write software because I like it. The job pays for the hardware, but most of what is on this desktop started as a question I wanted answered on a weekend and then got out of hand.

Lately that means agents: things that look at a screen, work out what to do, and then go do it. Before that it was making models small enough to run where nobody expected them to run at all.

I am open to what comes next. If you are building something in that direction, the mailbox at the bottom of this page works.

What I reach for

Python PyTorch LangGraph LangChain Vision Language Models Quantization Parameter Efficient Fine Tuning SwiftUI Docker AWS GCP SQL

Boot Sequence

Aug 2025 – Aug 2026

Founding Engineer

Drizz (New York, NY)

Built Fathom, an agent that uses a phone the way a person does. You tell it what to test in plain English and it works through the app on iOS and Android, planning each step, looking at what is on screen, and checking its own work as it goes. It runs on a LangGraph state machine.
Wrote the 71 tools it calls to read files and drive the device, and got the loop tight enough that each step lands in under four seconds.
Put a layer of smaller agents underneath it so a long test does not run past the context window and lose the thread.
Built the harness that grades it on whether a test finished and whether the agent stayed on task, then ran more than 200 combinations of prompts and tool layouts through it to find the one that scored best.

Vision Language Models LangChain Python LangGraph PyTorch GCP

Sep 2023 – May 2025

Graduate Student Analyst

New York University (New York, NY)

Fine tuned RoBERTa to read open ended survey comments and pull out what people were actually complaining about. It found the right problem 80% more often than what it replaced.
Kept the university data warehouse loading cleanly, about five million rows of it, which mostly meant catching malformed records before they landed and staying inside the governance rules.
Built more than 15 Tableau dashboards so department heads could watch their own numbers instead of emailing me for them.

Microsoft Excel Tableau Qualtrics Docker Oracle DB scikit learn Python

Apr 2022 – May 2023

Data Scientist

KPMG (Bangalore, IN)

Wrote the Python that reads financial filings and drops the numbers straight into the report. It saved the team more than 264 hours a year of copying figures by hand.
Trained a BERT model to find the names, dates, and amounts buried in those filings. It picks the right one 99 times out of 100.

Jan 2021 – Mar 2022

Analyst

Rewrote the slow SQL behind the live dashboards and took 30% off how long they took to load.
Wrote and documented the REST endpoints that served the forecasting models running on AWS.

Alteryx API Development AWS API Documentation Python

Applications

Clearbook

A money app for iPhone you can ask about your own spending. Apple Intelligence runs the answers on the phone itself, so nothing about your finances ever leaves it. The model never does the arithmetic either — real code adds everything up, and the model only explains what it found.

Apple Intelligence Runs On Your Phone Privacy by Design SwiftUI

Fine Tuning Large Models on a Small Budget

Retraining a big model normally costs more than a student has. I found a way around it that trained about ten times faster for a little over half the money.

Parameter Efficient Fine Tuning Quantized Low Rank Adapters Residual Connections Efficient Inference

Deep Learning Small Enough for Everyday Hardware

Shrank a model that recognizes images down to half its size so it runs on ordinary devices. It came out about ten times faster and needs no server behind it.

Parameter Reduction Quantization Efficient DNN Training

A Model That Reads Chest Radiographs

Taught a model to look at a chest radiograph and describe it the way a radiologist would. It names the right condition 96 times out of 100.

Multimodal Fine Tuning LLMOps Vision Language Models

Fine Tuning LLMs with QLoRA

Spent a semester teaching language models to follow instructions on a shoestring, running one setting after another to find out what actually worked.

Fine Tuning LLMs Experimentation Deep Learning

MacinTalk

Jul 2026

Your Agent Has Amnesia: Why Mobile AI Agents Need Memory to Be Useful

droidcon USA (Orlando, FL)

Watch the Talk

Training Data

Sep 2023 – May 2025

Master of Science in Computer Engineering

New York University

Jul 2017 – May 2021

B.Tech in Computer Science and Engineering

SRM Institute of Science and Technology

Ping Me!

Say hi, pitch me something fun to build, or ask about anything on this desktop. I actually reply.