Current work spans Trainium/Neuron evaluation, multi-agent oversight, adversarial robustness, PEFT/LoRA experimentation, and graph-based traces for agentic and robotic systems.
Build the systems, polish the details, and make the product feel inevitable.
I build software that earns trust under pressure. I’m a computer science student at San Francisco Bay University exploring the systems underneath useful products: AI infrastructure, graph-based security, developer tooling, and research for robust autonomous decision-making.
Professional experience
I focus on small, defensible scopes: a clear data model, safe defaults, observable behavior, and an interface that makes the important decision easy to understand.
This work sharpened the habit that carries into every build: measure what is happening, document the path back, and make the next person's job easier.
Proof of work
GitHub feature imageDOCKET
A source-grounded neighborhood information platform that brings local records, news, incidents, places, posts, and issue voting into one useful civic workspace.
GitHub feature imageBlast Radius
Dependency intelligence for npm supply-chain incidents: trace exposure through transitive dependencies and turn findings into remediation paths.
GitHub feature imageAirlock
A change-control system for AI agents that tests proposed production changes against a safe shadow environment before approval.
GitHub feature imageSemantic Output Cache
A provider-agnostic cache that recognizes related requests while retaining precise, verifiable identities for storage and invalidation.
Notes from the workbench
The approval button should be the last thing you build
Engineering · 8 min read
A dependency graph is a question-answering machine
Security · 6 min read
Cache the meaning, keep the identity
AI infrastructure · 5 min read
Currently building
GitHub contributions · @Rohit-ATS
View profile ↗Technologies I use to solve real problems
C++Systems & algorithms
PyPython
⚡FastAPI
AWSTrainium · Bedrock
◈Docker
PGPostgreSQL
Selected engineering work · 2026
⌘DOCKET: a civic AI system with grounded answersDesigned a source-first retrieval workflow that joins local public information with citations people can inspect.2026↗
⌘Blast Radius: dependency exposure as a traversable graphModeled relationships across direct and transitive npm dependencies to support incident response questions.2026↗
⌘Airlock: change proof for agentic systemsBuilt constraints around proposals, reversible test runs, checksums, and reviewable evidence.2026↗
If you’ve read this far, you might be interested in what I do.
Start a conversation ↗