Available for work — Remote or relocation
Multi-agent orchestration, agentic RAG, and the unglamorous reliability work underneath.
Engineer at Ember AI, shipping multi-agent and document-retrieval systems for legal tech and media — currently in production with paying users.
Orchestration
Staged multi-agent pipelines with typed handoffs — LangGraph, Pydantic schema contracts, explicit stage ordering over emergent agent chatter.
Retrieval
Hybrid search over document-heavy corpora — vector plus full-text merged with RRF, structure-aware chunking, evals that keep their failures visible.
Infrastructure
Async Python services, Celery/Redis job pipelines, multi-provider LLM abstraction with failover, and the cost engineering that makes it affordable.
~40 specialized LLM agents composed into staged synthesis pipelines.
Read the breakdown
FastAPI · LangGraph · Claude · Gemini · OpenAI · Pinecone · …
Hybrid retrieval and cross-document inference over venture-financing deal documents.
Read the breakdown
FastAPI · Pinecone · Claude · Gemini · MongoDB · Celery · …
A public RAG system built to show the parts most demos skip.
Read the breakdown
Python · FastAPI · PostgreSQL + pgvector · Next.js · OpenAI
Seven stages, ~40 agents, and a typed contract at every handoff.
A left-to-right pipeline of 7 stages containing approximately 40 agents in total. Stage 1 of 7: FOUNDATION. It hands off a CaseFoundation payload to the next stage. Stage 2 of 7: DISCOVERY. It hands off a DiscoveryPlan payload to the next stage. Stage 3 of 7: MOTION PRACTICE. It hands off a MotionSet payload to the next stage. Stage 4 of 7: HEARING PREP. It hands off a HearingBrief payload to the next stage. Stage 5 of 7: DEPOSITION. It hands off a DepositionPlan payload to the next stage. Stage 6 of 7: EXPERTS. It hands off a ExpertReport payload to the next stage. Stage 7 of 7: FINAL STRATEGY. Each arrow between stages is a typed, schema-validated contract rather than free text.
~40 agents across seven stages
Every handoff is a Pydantic-validated schema with explicit status enums and mandatory evidence citations — which is what makes ~40 agents agree with each other.
Remote worldwide, and open to relocation. If you’re building something where a confident wrong answer is expensive, that’s the work I want.