┌─────────────────────────────────────────────────────────────┐
│ GAURAV DUBEY · Senior Engineer │
│ Expedia Group · gurgaon, india · ↑ since 2022 │
└─────────────────────────────────────────────────────────────┘
🟢 building agentic systems that catches fraud — LLMs orchestrated through deterministic graphs, grounded in policy, and observable enough to put on call.
11 years in. Java| Python & distributed systems → now full-time on agent infrastructure.
| role | Senior SDE· Expedia · Fraud & Risk platform |
| building | Multi-agent fraud reasoning over LangGraph + MCP |
| shipped | Privacy-preserving LLM gateway · three-tier tokenisation · eval-gated in CI |
| writing | A field guide to deterministic agents — wip |
| open to | collaborating on agent infra · eval tooling · MCP servers |
AI LangGraph · LangChain · MCP SDK · FastAPI · Pydantic · Presidio · spaCy
Models Bedrock · SageMaker · Claude · OpenAI · Groq · Ollama · OpenRouter
Data PostgreSQL · DynamoDB · Redis · Elasticsearch · Kafka · Airflow
Core Java · Python · Golang · TypeScript · AWS · Docker · K8s · Vault
Crypto HKDF · HMAC-SHA256 · CRC8 · per-conversation key derivation
Eval LangSmith · OpenTelemetry · Grafana · property-based · golden sets
|
📘 diffprompt Calibrated LLM-as-judge with structural rubrics. CI exit codes that block regressions before they merge — the thing I always wished I had. 🟦 Python |
📘 fraud-graph Multi-agent LangGraph runtime for transactional fraud — policy nodes, vector recall, a verifier loop, and deterministic replay. 🟦 Python |
|
📘 mcp-policy-server An MCP server that exposes risk policies as typed tools. Plugs into Claude / Bedrock; ships with golden cases. 🟦 TypeScript |
📘 agent-eval Property-based evaluation harness for agent traces. Diffs two runs, surfaces drift, blocks deploys on regressions. 🟦 Python |
last synced · 2026-04-29
┌─ impact ──────────────────────────────────────────────────────────────┐
│ 5 merged · 10 open · 935k⭐ reached
│ ▅█ █▂▂▂▅▅ prs over last 30 days
└───────────────────────────────────────────────────────────────────────┘
| repository | ⭐ | merged | open | latest title |
|---|---|---|---|---|
| affaan-m/everything-claude-code | 169k | #1490 | #1567 | fix(hooks): bypass gateguard friction gates inside subagent invocat... |
| huggingface/transformers | 160k | #45611 | — | Raise clear error for problem_type="single_label_classification" wi... |
| google-gemini/gemini-cli | 102k | #25822 | #26074 #26011 | fix(core): handle ENAMETOOLONG/ENOTDIR in robustRealpath (#26010) |
| apache/airflow | 45k | #65685 | — | Honor AUTH_ROLE_PUBLIC in FastAPI API server |
| danny-avila/LibreChat | 36k | #12825 | — | 🧹 fix: Graceful MCP OAuth Revoke Cleanup When Tokens Are Missing |
| openclaw/openclaw | 365k | — | #70287 | fix(msteams): drop unsupported $search on msteams:search (AI-assisted) |
| Lightning-AI/pytorch-lightning | 31k | — | #21686 | Fix torch.compile breaking toggle_optimizer / untoggle_optimizer |
| mastra-ai/mastra | 23k | — | #15769 | fix(core): reject plain functions as individual tool entries in Age... |
| langchain-ai/langchain-google | 370 | — | #1708 | fix(genai): wrap batch texts in Content objects for correct embeddi... |
| LambdaTest/agent-skills | 240 | — | #5 #4 #3 | feat(api): standardize frontmatter, update validation, register 24 ... |
Sorted by upstream star count. Closed-unmerged PRs skipped.
- › Benchmarks over vibes — datastores picked with IOPS +
$/hr; LLM routes picked with eval numbers - › Security is the first tenet, not the last — name the threats, name the controls
- › Entity thinking — if you're debugging individual events, you're already behind the fraud ring
