Nitish Chowdary
ML Systems Engineer.
Overview
MS Software Engineering @San Jose State University
Machine Learning Intern, Ads Economics (Summer 2026) @DoorDash
he/him
Resume
Social links
About
I don't just want models that work. I want models you'd actually trust with a decision!
Five years of studying and building AI systems has taken me across ads marketplaces at DoorDash, financial data at Broadridge, industrial IoT at Harman, fraud detection at Kona AI, and edtech at Daira, which I co-founded. Very different problems, but the same question keeps pulling me back: how do you make something this powerful know its own limits?
So that's what I build! Agents that plan their way across a browser. A bandit platform that models how one decision spills into the next. An agent runtime small enough to run on a microcontroller. A robot dog that recognizes you and holds a conversation. Some of my favorite projects started as a weekend hackathon with friends and never quite stopped.
I'm finishing my Master's in Software Engineering at San Jose State University and looking for full-time roles starting May 2027! I moved here from India in 2025 to do this properly. If you're building something ambitious, I'd love to help. The projects are all here, code included.
Outside work, you'll usually find me on a bike. Long-distance cycling is the thing that clears my head, and the longer the ride the better it works! The rest of my time goes to books and to hunting down music I haven't heard yet, which is a habit I have no plans to fix. I'm also a hopeless animal person, and yes, there's a cat, and no, I don't get a say in anything!
Projects(15)
An incentive experimentation platform for a simulated delivery marketplace where zones interfere with each other! Delivered as a FastAPI and Next.js application with 8 analytical dashboards, pluggable Dasher swarm simulators (built-in, OASIS, MiroFish), and a GPT-4o agent that proposes experiment configuration while a human authorizes every write.
- Extends a DoorDash-style Thompson Sampling incentive baseline with networked linear Thompson Sampling, shape-constrained Bayesian optimization, and a delayed doubly robust (AIPW) reward correction, so interference between neighboring zones is modeled instead of assumed away.
- Across 50,400 simulated incentive decisions in 10 Bay Area zones, cumulative regret fell 67.4% against vanilla Thompson Sampling. That is one run against a synthetic oracle, so what it supports is the ranking between the two methods and not the figure itself.
- A separate 3-day run measured a direct effect of +0.081 on the treated zone against an indirect effect of -0.092 on its neighbors, a 114% spillover ratio. Incentivizing one zone pulls supply away from the zones next to it, which is precisely what an independent per-zone experiment cannot see, and why such an experiment overstates true impact.
- Solo
- Python
- FastAPI
- Next.js
- Causal Inference
- Multi-Armed Bandits
Experience
DoorDash
- Location
- San Francisco, CA
- Location type
- (On-site)
- Shipped Ads Econ Health, a production health-monitoring system for the ads marketplace that computes 90 metrics from 28 queries across Snowflake, Databricks, and Prometheus, writes them idempotently to a daily snapshot table, and delivers a digest to Slack.
- Designed a fail-closed LLM root-cause agent that collapses an alert storm into a single incident, using change-point onset detection and a Beta-Binomial posterior over root causes to localize failures across 6 upstream subsystems.
- Cut root-cause time on real alerts from ~1 hour to minutes by scoring evidence across 20 diagnostic checks (change correlation, blast radius, drift) and emitting only typed claims an executable predicate could verify against a warehouse row.
- Python
- SQL
- Snowflake
- Databricks
- Prometheus
- LLM Agents
- Change-Point Detection
- Bayesian Inference
Broadridge Financial Solutions
- Location
- India
- Location type
- (Hybrid)
- SQL
- Python
- Query Optimization
- Concurrency
- Automated Testing
Harman Connected Services
- Location
- India
- Location type
- (On-site)
- Python
- Streaming Ingestion
- IoT Telemetry
- Time Series
- Plotly
Education
- GPA: 3.9
- Coursework: Machine Learning, Generative AI, Reinforcement Learning, Recommender Systems, Deep Learning, Data Mining, Enterprise Software
- Machine Learning
- Deep Learning
- Reinforcement Learning
- Recommender Systems
- Data Mining