DWG NO. AS-100 · COVER SHEET
GENERAL NOTES
G1ML engineer at a stealth EdTech startup and applied ML lead at Virginia Tech's NLP Research Lab, currently at Cornell.
G2I build multi-agent systems, NLP pipelines, and ML infra that ships to production.
KEY PLAN · SHEET LAYOUT AS-2026
9 SHEETS TOTALFLAGGED · 05 OPENSIGHT
ABHIJIT SOMALA
I build ML systems, from research pipelines to production agents.
I've worked across the ML stack, from unsupervised clustering on data with no labels to inference pipelines running in production for tens of thousands of users. Right now I'm building at Focara, an EdTech startup, and leading applied ML work at Virginia Tech's NLP Research Lab. Before that I built embedding visualization infrastructure at Nomic AI.
Cornell University · Ithaca, NY.
Currently at Cornell. Transferred in from Virginia Tech.
A real-time multi-agent voice assistant that replaces screen readers with active task execution for blind and low-vision users. I built it because screen readers describe a page, they don't act on it.
Top 10 of 2,500+ teams, Google Developer Groups Solution Challenge 2026, North America. Multi-step web task in under 60s.
Stack: FastAPI, WebSockets, Gemini on Vertex AI, Deepgram, Google Cloud TTS. Validated with the Blind Institute of Technology and the Virginia Department for the Blind and Vision Impaired.
Wildfire spread prediction. A deep learning pipeline trained on 100K+ NASA environmental datapoints from VIIRS and MODIS.
Containerized with Kubernetes and Helm.
Harmful algal bloom detection. Ground truth comes from EPA measured microcystin concentrations, not a heuristic proxy.
Leak-free cross-validation splits, so reported performance isn't inflated by data leakage. Cross-validated ROC-AUC 0.58–0.69 and still improving. Active work in progress.
| REV | DATE | DESCRIPTION |
|---|