Hola!

Projects

AgentGate

Action-layer firewall for AI agents: intercepts tool calls in real time and blocks any dangerous or malicious actions

PythonPyPINext.JSLangChainOpenAIEvals

Published to PyPI as agentgate-py, installable with a single pip install
99.5% recall and 97.7% precision across 370 adversarial eval scenarios
Two-tier architecture: deterministic keyword firewall + LLM judge for ambiguous cases
Supabase-backed audit trail generates regulator-ready compliance evidence per agent action

HireSignal

Agentic LangGraph pipeline that decodes AI lab hiring patterns into ranked intelligence: who's scaling, what they're building, and what's coming next.

LangGraphFastAPITavilyStreamlitLangSmith

7-node LangGraph StateGraph: validate, coordinate, fan-out research, extract, analyze, narrate, synthesize
Produces explainable 0-100 AI intent scores with full signal breakdown per company
Ingests job postings, funding news, and arXiv papers per company
Week-over-week hiring velocity deltas computed against persisted JSON snapshots
Full LangSmith tracing on every node — latency and LLM call visibility across the entire pipeline

Customer Service RAG Chatbot

Production-grade RAG service with automated RAGAS evaluation and full LangSmith observability on every pipeline step.

RAGLangChainChromaDBRAGASDocker

RAGAS baseline: faithfulness 0.83, answer relevancy 0.81, context precision 0.82, context recall 0.89
Automated eval harness generates 40 Q&A pairs and scores all four RAGAS metrics per run
LangSmith tracing on every retrieval and generation step
Dockerized with health endpoint and ChromaDB volume mount — docker-compose up deploys the full stack
Multi-turn conversation support with session-scoped context rewriting for follow-up question handling

DiscoverAI

AI-powered ideation platform that turns a company name into signal-grounded project pitches in under 5 minutes, replacing 2-3 hours of manual BD research.

Next.jsFastAPILangChainGPT-4oSSESupabase

Parallel research pipeline via asyncio.gather — fans out 6 Tavily queries, homepage scraping, and logo resolution concurrently
Signal-first prompting enforces named, dated company signals before any idea is generated — measurably reduces generic output
Streams generated idea nodes over SSE with first card rendering in ~3 seconds while the full batch continues loading
React Flow canvas with custom tree layout algorithm positions idea cards in real-time as they arrive from the server
Hot-reloadable guidelines.md prompt config — non-engineers can tune ideation quality without touching code or restarting the server
Multi-provider LangChain abstraction supports GPT-4o and Claude Sonnet in the same request flow
Now the default first step in every Applied Engineering partnership engagement at SJSU

AI Infrastructure Deployment Planner (Ongoing)

Working with Bloom Energy: Decision-mapping tool for enterprise leaders translates a business goal and workload into a full infrastructure dependency chain across compute, power, cooling, and permitting.

Next.jsFastAPILangChainReact FlowClaude

User inputs a business goal, workload type, and deployment timeline — the tool resolves the full chain of infrastructure implications automatically
Dependency graph engine propagates constraints across 8 dimensions: compute, location, latency, cooling, power, utility interconnect, onsite generation, and permitting
AI reasoning layer explains each bottleneck in plain language — not just what the constraint is, but what has to be true for it to be resolved
Highlights feasible vs. infeasible deployment paths given the input parameters, with tradeoff explanations at each fork
Built for Bloom Energy to surface scenarios where utility timeline bottlenecks make onsite power generation the only viable path forward