Experience

Amazon Lab126

Software Engineer Intern

Jun–Sep 2026
Sunnyvale, CA
  • ·Shipped on-device speaker identification to a production vehicle fleet, personalizing the voice assistant without added latency
  • ·Designed the statistical evaluation governing whether a retrained model is promoted over the one in production

Amazon Leo

ML Project Lead (contract)

Jan–May 2026
College Park, MD
  • ·Forecast the direction of international space policy a year ahead from current academic, government and industry publishing
  • ·Built the ingestion and filtering pipeline that made those signals reliable enough to forecast against

How it works

academicpolicyindustrygateNLI relevance≈24,000keptrejects retained as an audit traildebristrackingdisposaltrafficBGE embeddings · topic registry · biweekly bucketsnowforecastattention per topic, biweeklyhorizon
Ten scrapers pull six years of academic, government and industry writing. Every document clears a relevance gate before it counts.What survives is embedded and matched into a registry: the concerns the field is actually arguing about.Each topic becomes a series. A sequence model extrapolates it past the present, with the cone widening as the horizon grows.

Time-Series Forecasting of Space Policy Attention

Orbit is getting crowded, and the arguments about it start in writing long before they reach policy. This reads six years of what the field published and forecasts which concerns are gathering and which are dying out, so a team sees a debate forming instead of reacting to one.

time-series forecasting · natural-language inference · topic modeling · signal extraction

Private repository.

General Dynamics Information Technology

Software Engineer Intern

Jun–Aug 2025
Annapolis Junction, MD
  • ·Delivered an edge vision system from scratch that identifies objects in live video and streams their positions to operators in the field
  • ·Hardened the models against deliberately deceptive inputs and brought the full pipeline onto the handset, with no network dependency

How it works

three models, on the deviceno networkwhat isthat?seconds, disconnected, in the fieldL1fast triageresolved hereL2transfer stageharder framesL3fine-grainedremainderε hardenedTensorFlow Lite · quantized · thread-safe over live videoclassifiedpushed to TAKthree models, one answerno round trip, no server dependency
A handset in the field, carrying three models and no connection to anything.The three unfold into a cascade. Each stage sees only what the one before it could not settle, so the budget goes to the hard frames.The three agree on one label. It drops as a map pin, pushed to the team the moment a server is reachable.

Adversarially Hardened On-Device Vision Cascade

Someone in the field points a camera at something and needs to know what it is now, with no signal and no server. Three models answer in turn, each taking only what the one before it could not settle, and all three are trained to fail predictably when the input is meant to deceive.

cascaded inference · adversarial robustness · on-device vision · quantization

Private repository.

General Dynamics Information Technology

ML Project Lead (contract)

Jan–May 2025
College Park, MD
  • ·Led an engineering team building a retrieval assistant that let analysts query an agency record corpus directly
  • ·Reduced analyst discovery time from hours to minutes while improving retrieval accuracy

Booz Allen Hamilton

Engineering Project Manager

Sep–Dec 2024
McLean, VA
  • ·Automated the review of medical policy documents, with human judgment retained at each decision point
  • ·Replaced a manual audit process with one that surfaces only the cases requiring review

Internal Revenue Service

Software Engineer Intern

Jan–Dec 2024
Washington, DC
  • ·Delivered the data pipeline underlying fraud detection across the national filing system
  • ·Modernized the division's build and delivery system, removing a long-standing bottleneck for its engineers
ScholarLinkedInGitHub
© Nishad Wajge 2026