Work
Where I have been employed, and what I was responsible for.
01Jun 2026 —
Present
Present
Artificial Intelligence Engineer
Greenboard- Refactored the ingestion and review backend into a modular, provider-agnostic architecture, so new text-parsing, transcription and LLM providers plug in without touching the core pipeline.
- Designed and deployed asynchronous document processing on AWS ECS and SQS, scaling across file types and formats.
- Built an end-to-end benchmark suite over the whole pipeline, catching regressions automatically across provider swaps and code changes.
- Redesigned the customer-facing frontend in React and TypeScript so reviewers could check cited violations against the source document and accept or reject LLM-suggested edits.
- Instrumented the services with CloudWatch for centralised logging, error alerting and traceability across the pipeline, cutting the time to diagnose a failure.
- Built an experimental service to run automated compliance checks on company websites through the help of Firecrawl
StackReactTypeScriptAWS ECSSQSMySQLPostgreSQLCloudWatch
[ Full writeup → ]02Jun 2025 —
Aug 2025
Aug 2025
Research Assistant
Johns Hopkins University- Ran experiments testing whether single agent algorithms such as PPO, TRPO and REINFORCE converge to useful policies in a multi-agent tank battle environment.
- Refactored and documented a distributed training pipeline originally written by the Johns Hopkins Applied Physics Laboratory, and adapted it to new experiment and visualization needs.
- Reimplemented training algorithms in PyTorch alongside the existing TensorFlow code, and presented experimental results.
StackPythonPyTorchTensorFlowMPI
[ Full writeup → ]03Jun 2024 —
Aug 2024
Aug 2024
Innovation Intern
Supernova Technologies- Designed and implemented a retrieval augmented generation pipeline from scratch that answers legal due diligence and compliance questionnaires from the company's own documents, with cited reasoning for a human reviewer.
- Replaced an expensive general purpose parsing service for W2 and similar forms with a layout based approach: standardized bounding boxes per document type, localized extraction inside each box, and structured JSON out.
- Built a custom benchmark to compare strategies and model versions on proprietary documents.
StackPythonOpenAI APIOpenCV
[ Full writeup → ]