AI engineer building the evaluation, guardrails, and identity-verification layers behind an AI legal platform — the code that decides whether an LLM's answer is trustworthy enough to show a lawyer. MS CS @ George Washington University '26 · Springer IEA 2024 published · Patent filed.
What I work on
🔍 Adversarial evaluation & citation verification — an off-request-path agent crew that checks every AI legal answer before a user sees it
🛡️ Guardrails & identity verification — face-liveness detection, OCR, bar-status checks for an attorney marketplace
📚 Retrieval — CourtListener case-law search + BM25 relevance scoring
🤖 Multi-agent systems — CrewAI-based orchestration, in production and in personal projects
Featured work
Project — What it shows
rl-code-reviewer — real PPO training with honest evals, including documented negative results
IntentOps — multi-agent AWS cost investigation, gated behind human approval
-
Avnet
- Arlington, VA
-
03:33
(UTC -04:00) - in/iamkushaldave
Highlights
- Pro
Pinned Loading
-
Currency_Vision
Currency_Vision PublicYOLOv8 (custom SE-attention variant) banknote detection for 6 currencies, exported to ONNX and running fully client-side in the browser — no server.
Jupyter Notebook 1
-
bandit-pricing-lab
bandit-pricing-lab PublicA reinforcement-learning dynamic pricing system using Multi-Armed Bandits, applied to real retail data. Implements Epsilon-Greedy and SW-UCB, with full environment setup, evaluation metrics, regret…
Python
-
IntentOps
IntentOps PublicMulti-agent FinOps: LLM agents investigate AWS cost anomalies — every destructive action gated behind human Slack approval.
Python
-
rl-code-reviewer
rl-code-reviewer PublicRL agent that learns an adaptive prompting policy over a frozen LLM (Mistral 7B) to deliver context-aware code reviews — trained via PPO on human preference data from Microsoft CodeReviewer
Python
-
citation-lens
citation-lens PublicCitation family trees from any paper title — interactive graph UI + MCP server so AI agents can walk citation lineage under a fixed-cost rate-limit budget.
JavaScript
-
StockWise_GeorgeHacks
StockWise_GeorgeHacks PublicForked from bhushan7322/StockWise_GeorgeHacks
George Hacks 2026: marketplace for DC food-desert corner stores — 3-tier fulfillment engine, SNAP/EBT-aware checkout, and a Gemini assistant wired to live transactional data.
JavaScript
If the problem persists, check the GitHub status page or contact support.



