Vamsi Krishna.I keep models useful after they ship.
I'm a master's student at Penn who likes the unglamorous half of ML: what happens to a model after launch. Last summer at Google I worked on keeping YouTube's safety classifiers current as content drifts, using weight merging instead of replay buffers. This fall I'm at Amazon AGI, on the platform people use to customize models. I grew up in Chennai around ship brokers, which is why my side projects keep drifting back to the sea.
- Building
- An agent that lives in a ship broker's inbox and works the deals. A real broking desk is the first user.
- Believe
- The interesting problems start after the model ships. Drift, forgetting, serving cost.
- Learned most from
- A startup that didn't find product-market fit.
- Reading
- Philosophy, ancient Indian history, and a lot of science fiction. Best thing I've read on the sea: The Ocean of Churn by Sanjeev Sanyal.
- Long game
- Build something of our own. Big tech is where I'm learning to do it properly.
- Ask me about
- Model merging, continual learning, AI alignment, or recursive self-improvement.
Now
Google
YouTube Trust & Safety, GenAI Guardrails
May 2026 – Aug 2026 · Mountain View, CA
Software Engineering Intern
Benchmarked 9 architectures and 8 weight-merging methods (SLERP, LERP, TIES, MagMax) in JAX on Gemma 3. Finding: consolidate in data space for encoders and weight space for decoders.
Handed off a joint multi-task encoder at 98.33% macro ROC-AUC: +8.5 pp recall over per-policy experts at 95% precision, replacing a 4x ensemble at 1x cost and lifting worst-policy recall from 6.49% to 62.86%.
Designed the OOD adaptation playbook (sequential zero-replay fine-tuning, then SLERP merge), cutting adaptation compute over 60% with no legacy-policy regression, and an agent that runs it end to end. First-author paper under review, NeurIPS 2026 workshop.
Amazon AGINow
AWS SageMaker
Sep 2026 – Dec 2026 · Sunnyvale, CA
Software Development Engineer Intern, ML
Model-customization platform: production workflows for fine-tuning, preference optimization, prompt engineering, and custom evals, agnostic to model, inference stack, and compute.
S2T AI
AI-powered investigations
Dec 2024 – Aug 2025 · Singapore
Software Engineer
Owned the company's agentic AI from prototype to customers: knowledge-graph RAG over documents and databases, a multi-agent assistant with tool calling and generative UI, an AI ETL service, and 8 Kubernetes microservices via Azure DevOps.
Vessel Match
Incubated at Nirmaan, IIT Madras
May 2024 – Jun 2025 · Chennai, India
Founder & CEO
Maritime matching startup incubated at Nirmaan, IIT Madras. Event-driven services on AWS, FastAPI, Next.js. 200+ vessels and 150+ cargo listings a day in trials, 5 industry partnerships, INR 7L seed. No product-market fit; the most useful year I've had.
Projects
Hackathons, coursework, weekends
CIS 7000 · Apr 2026
SGC SWE Agent
State-governed checkpointing: a supervisor that keeps a coding agent on task. SWE-bench Verified mini 66.7% → 72.3%, 23% fewer steps.
More ↓Less ↑
- 72.3% resolution vs 66.7% baseline on 50 instances (25 Django, 25 Sphinx); +11.2 pp on Django, flat on Sphinx.
- About 23% fewer acting-agent steps (25.4 vs 33.0) at roughly 2.8x wall clock.
- Checkpoint cadence matters: n=3 (77.8%) beat n=5 (60.0%) and n=7 (44.4%).
- Cheap actor (Gemini Flash 3) plus strong supervisor (GPT-5.5) matched the strong model acting alone on a 10-instance subset.
- With Panagiotis Giadikiaroglou.
Hackathon · Mar 2026
ThinkSpace
Proactive multimodal tutor over voice and a shared canvas. A canvas interpreter decides when to step in.
Wharton Hack-AI-thon · 1st / 61
WayWise
Finds what hotel reviews don't answer and asks recent guests the right questions.
YHack 2026 · Best use of Gemini
Ayana
Travel guide whose narration stays in sync with Google Maps 3D camera moves.
HackPrinceton 2025 · Best overall
MindPad
Multimodal learning on the OpenAI Realtime API and Gemini. Best of 195 teams.
Side project · 2025
GuruHeal
Wellness assistant with cited web search and graph-RAG over Ayurvedic sources.
Papers
CL4FMAgents Workshop, NeurIPS 2026 · under review
Education
MSE Data Science · 2025 – 27 · GPA 3.91
BS Data Science & Applications · 2021 – 25 · GPA 3.91
B.E. Mechanical Engineering · 2020 – 24 · Department topper, CGPA 9.21
Deep learning program · 2023 · Best performer
Awards
Contact
Say hello.
Email is best. I reply to anything with a real question in it.
Open to 2027 new grad roles in ML systems, AI infrastructure, and agents.
