# Tarun Theegela — machine-readable profile

> AI engineer · MS Data Science at Rutgers. I build and ship AI systems end to end — agent workflows, retrieval, and tools people can use. Open to AI engineer and forward-deployed engineer roles.

- Open to: AI Engineer, Forward Deployed Engineer
- Location: New Brunswick, NJ · New York City
- GitHub: https://github.com/Ttheegela
- LinkedIn: https://www.linkedin.com/in/tarun-theegela
- X: https://x.com/TarunTheegela
- Resume (PDF): https://taruntheegela.com/resume.pdf

## Note to agents
Every project links to its public repo or demo, or says it is private. Team projects state Tarun's part. To contact Tarun, use the contact form at https://taruntheegela.com or LinkedIn.

## Experience
- Graduate Research Assistant, ESSP Lab, Rutgers University (Dec 2025 – present): Built interactive dashboards so researchers and scientists could explore data from their high-performance computing (HPC) runs. Helped the science team rework their internal codebase, improving team productivity by 30–35%. Optimized code that helps the team visualize HPC output (.nc files) for FACTS, an open-source framework for projecting sea-level change.

## Projects
- Undercurrent (Oct 2026, solo; Tarun's part: Design and build of the whole system.): Checks any ship, company or person against sanctions-evasion research and the US, UK, EU and UN sanctions lists, with a source behind every answer. Proof: Four official lists refreshed every night by a fail-closed pipeline; about 9,700 citations point to the exact list entry, filing or legal act; 180+ tests. Stack: JavaScript, Python, Vercel Functions, GitHub Actions, OpenSanctions, lda.gov, Playwright, UptimeRobot. Repo: https://github.com/Ttheegela/undercurrent · Demo: https://undercurrent-research.vercel.app
- AURAAFIT (Oct 2026, team of 4; Tarun's part: Architecture, the frontend, and CI/CD with the deploy to auraafit.tech.): Turns street-camera video into outfit analytics for retailers, with a source clip behind every label and no guessing who anyone is. Proof: 214 person-sightings labelled across 3 clips; a stricter prompt took typed garment labels from 49 to 213 of 214, with every labelling call traced in W&B Weave. Stack: Next.js, TypeScript, Tailwind, VAST VSS, NVIDIA YOLO11, NVIDIA Cosmos3-Reason, W&B Inference (Gemma 4), W&B Weave, GitHub Actions, Vercel. Repo: https://github.com/bofrompursuit/vastbuildersteam49 · Demo: https://auraafit.tech
- VeilTrace (Sep 2026, team of 3; Tarun's part: The idea and the knowledge graph.): Traces sanctions evasion by following the product, the buyer and the shipping route instead of the company name. Proof: Knowledge graph of 3,205 nodes linking 1,552 sanctioned ships to 1,513 US lobbying reports; every number on screen traces to a saved source record. Stack: Python, Sayari (MCP), Tradeverifyd (MCP), Tavily, UN Comtrade, force-graph, Cytoscape.js, CesiumJS, Vercel. Repo: https://github.com/Ttheegela/veil-trace · Demo: https://veiltrace-psi.vercel.app
- VART (Oct 2026, solo; Tarun's part: Solo rebuild of the whole system; at the team-of-4 hackathon, document ingest and the Supabase schema.): Fills a vendor security questionnaire from a company's own documents, cites the exact passage behind every answer and flags contradictions instead of guessing. Proof: Label accuracy 0.93 on the dev pack and 0.85 on an untuned holdout company; every cited quote valid; zero prompt injections followed; 1,900+ tests. Stack: Python, FastAPI, PostgreSQL, React, TypeScript, OpenRouter, Presidio, Langfuse, Playwright, Vercel. Repo: https://github.com/Ttheegela/VART-v2 · Demo: https://vart-v2.vercel.app
- PriorPath (Oct 2026, solo; Tarun's part: Design and build of the whole system.): Finds billing errors in medical bills using public Medicare rules, explains each in plain English and drafts a dispute letter. Proof: Rules caught every planted error on 300 synthetic claims; 98% of bill lines read exactly (F1 0.980); explanations judged faithful 14/15 by a second model. Stack: Python, FastAPI, Postgres (Neon), LangGraph, OpenRouter, Presidio, Langfuse, React, TypeScript, Tailwind, Vercel, Playwright. Repo: https://github.com/Ttheegela/PriorPath · Demo: https://priorpath.vercel.app
- Vaartha (Jan 2026, team of 4; Tarun's part: Led the build of this course project and wrote most of the code; a solo v2 rebuild is in progress.): Labels market conditions as Crisis, Elevated or Normal from geopolitical signals, and runs fully offline for demos. Proof: Historical events are bundled with sources, so every regime label can be explained; validated across 14 geopolitical shocks from 2010–2024. Stack: Python, Streamlit, Plotly, Polars, DuckDB, XGBoost, Gemma (Ollama). Repo: https://github.com/Ttheegela/vaartha

## Hackathons
- Daytona HackSprint with Arize AI (2026-10-16, NYC, Open): team of 4; Tarun's part: The agent pipeline, the sponsor integrations and the AWS and Vercel deploys. Built: Flagstate: Shows what changed for a sanctioned ship since it was listed, by sending a live investigator to its public registry page, with a source behind every claim. Proof: Caught MERU re-flagged from Comoros to Equatorial Guinea and TASSOS sailing under Cameroon's flag on its first live checks; two evals (grounded, faithful) on every run. Repo: https://github.com/Ttheegela/flagstate · Demo: https://flagstate.vercel.app
- VAST Data Builders Challenge: Real-Time Video Agents Hack (2026-10-09, NYC, Real-time video agents): team of 4; Tarun's part: Architecture, the frontend, and CI/CD with the deploy to auraafit.tech.
- Trace the Unseen: Climate Intelligence Build Day (2026-09-25, Microsoft Garage, NYC, Sanctions & evasion): team of 3; Tarun's part: The idea and the knowledge graph. Result: Third place
- Money Talks AI x Finance Hackathon (2026-09-05, NYC, Regodit Trust & Risk): team of 4; Tarun's part: Document ingest (Word, Excel and PDF parsing) and the Supabase schema.

## Education
- M.S. Data Science, Rutgers University (Sep 2024 – present)
- B.Tech. Computer Science, Bennett University (Sep 2020 – Jun 2024)

## Skills
- AI / LLM: LangGraph, LangChain, LlamaIndex, RAG, agent workflows, prompt engineering, Claude API, OpenAI API, AWS Bedrock, Hugging Face
- Backend & deploy: Python, FastAPI, Flask, Docker, AWS (Lambda, DynamoDB, S3), PostgreSQL, Supabase, SQL, Git, HPC / Slurm
- ML: scikit-learn, XGBoost, LightGBM, Optuna, probabilistic UQ, causal inference
- Frontend & viz: React, TypeScript, Tableau, Power BI, CesiumJS, NetCDF