Founding AI Engineer @ Parcha β Building AI agents for compliance automation
I design and ship production LLM systems for KYC, AML, and adverse media screening. My work sits at the intersection of AI reliability and regulatory requirements β where getting it wrong isn't an option.
Just shipped Grep β an AI-powered business due diligence agent built on Anthropic's Agent SDK.
Give it a business name, get a complete intelligence report in under 3 minutes:
- Verified business profiles across 190+ jurisdictions
- UBO verification and ownership structures
- Sanctions screening (OFAC, UN, EU, UK), PEP checks, adverse media
- Risk assessment with cited sources
Try it free through December with code GREPIT β grep.ai
- π€ Multi-agent orchestration β Claude API, Anthropic Agent SDK, tool use at scale
- π LLM evaluation systems β Statistical frameworks for compliance-critical AI
- π Document understanding β OCR, entity extraction, cross-document verification
- β‘ Production infrastructure β AWS/GCP/Kubernetes, Terraform, Redis/Celery, microservices
Previously ML Team Lead at Carvana and Senior MLE at Augment CXM (5 years). 15+ years shipping data systems across fintech, automotive, and enterprise.
π³ K-Base β Branching Conversation AI
Exploring how conversation structure affects learning and brainstorming. Built a prototype that treats AI chats as trees instead of linear logs.
Key idea: What if you could fork any conversation, explore multiple solution paths simultaneously, and collapse tangents without losing context?
Stack: React + TypeScript, FastAPI, PostgreSQL + pgvector, LiteLLM
Status: Phase 2 complete (branching + streaming working), exploring RAG integration next
Why this matters: Most chat interfaces force linear thinking. K-Base lets you think in parallel β branch to explore "what if" scenarios, maintain multiple hypotheses, and use AI more like a collaborative thought partner than a sequential Q&A bot.
Current features:
- Branch conversations at any point
- Tree visualization with SVG rendering
- Streaming responses with SSE
- Session memory and context management
Next up: Collapsible branches with AI summaries, cross-session RAG, user annotations.
This started as a personal tool for exploring complex technical decisions where I wanted to preserve multiple solution paths. Turns out tree-structured conversations are pretty useful for learning and brainstorming too.
- AI agent evaluation methodologies
- Reliability patterns for LLM applications
- Entity disambiguation and adverse media screening
- Making compliance teams 10x faster
- Conversation structure and knowledge representation