Service exampleAILive
AgentFlow
Multi-agent platform that turns complex business files into cited, verifiable executive briefs.
The problem
Executives waste hours manually synthesizing complex CSV, Excel, PDF, and DOCX files into actionable summaries.
What I built
Built an orchestrated DAG of five specialized LLM agents (Orchestrator, Ingest, Retrieval, Verifier, Summarizer) using LangGraph and hybrid semantic search plus RAG.
The outcome
Every brief is automatically grounded in source documents, with strict citations and flagged risks.
How it works
- 01
Files land in Ingest, which normalises CSV, Excel, PDF, and DOCX into one chunked, embedded representation.
- 02
The Orchestrator decides which agents run for this request rather than running the full graph every time.
- 03
Retrieval runs hybrid search over the chunks: dense embeddings for meaning, keyword matching for the exact figure or clause.
- 04
The Verifier checks each claim in the draft against the retrieved passage and drops or flags anything it cannot ground.
- 05
The Summarizer writes the brief with citations attached, and risks it could not verify are surfaced rather than smoothed over.
The AI layer
Five agents in a LangGraph DAG with distinct responsibilities, because one prompt asked to ingest, retrieve, verify, and summarise does all four badly. The Verifier is the load-bearing one: it exists so an unsupported sentence never reaches the brief.
The engineering layer
A FastAPI service around it with per-format document parsing, background jobs for anything slow enough to time out a request, and a vector store the retrieval step reads from. Citations are stored as spans back into the source file, not as text the model repeated.
Key technical decisions
A verifier agent instead of a stern prompt.
Telling a model to state only what it can support does not survive a long document. A separate pass that checks each claim against a retrieved passage does, because it is a different question with a checkable answer.
Hybrid retrieval, not embeddings alone.
Dense search is good at meaning and unreliable at exact tokens. Financial briefs turn on exact tokens: a figure, a date, a clause number.
A DAG, not a conversation between agents.
Free-form agent chatter is hard to debug and unbounded in cost. An explicit graph makes every run reproducible and every step inspectable.