Separate the evidence
Filing passages become individually addressable chunks in a vector index.
FinSight AI
AGENTIC AI / FINANCIAL RESEARCHFull-stack research system
Every answer should leave a paper trail.
THE CONTEXT
Turn lengthy SEC filings into answers that can be traced back to a source.
THE KEY DECISION
Separate retrieval, deterministic arithmetic and a bounded review loop.
THE RESULT
Citations and structured metrics carried from the reasoning pipeline into the interface.
FinSightSOURCE → CITATION
MECHANISM STUDY / 02SCHEMATIC
TAKE A CLOSER LOOKSELECT A STAGE ↓
SCROLL TO TRACE. SELECT TO EXPLORE.
Follow the connections as you scroll, or select a stage to see what happens there.
Reason in a loop
Guard, retrieve, calculate, draft and judge; retry within a bounded loop.
Trace every claim
The structured answer reconnects to its supporting filing excerpts.
A financial due-diligence workflow built around a simple requirement: the reader should be able to follow an answer back to the filing that supports it.
PROJECT TYPE
Full-stack & retrieval engineering
FOCUS
AI / ML + Full stack + Backend + Data / analytics
STATUS
Full-stack research system
/01
The question.
Financial documents mix long narrative passages with figures that need precise arithmetic. Retrieval, calculation and drafting need different responsibilities, and unsupported answers need a path back to the evidence.
/02
The approach.
SEC filings are chunked, embedded and indexed in PostgreSQL with pgvector/HNSW. LangGraph coordinates input checks, retrieval, drafting with deterministic math tools, and a faithfulness judge. A bounded retry widens retrieval when the draft fails the configured check.
/03
What came out of it.
The system carries citations and structured metrics through shared JSON contracts into a Next.js research interface. It includes ingestion/cache workflows, failure states and evaluation infrastructure. The portfolio presents the implementation without claiming a verified public production service.
01
Arithmetic has an owner
Financial operations run through deterministic Python tools. Drafting and calculation stay separate responsibilities.
02
One contract across languages
JSON Schema is the source for generated Zod/TypeScript and Pydantic models, reducing drift across the API boundary.
03
Failure is visible
Unavailable providers and blocked queries produce explicit states. A local mock interface is distinguishable from a live provider-backed response.
Under the surface.
- Product
- Next.js · React · TypeScript · Zod
- Reasoning
- FastAPI · LangGraph · Python math tools · TensorZero routing
- Evidence & operations
- PostgreSQL · pgvector · n8n · PyRIT
SCOPE & LIMITS
- Provider-backed execution requires configured services and credentials.
- Mock mode is available for local interface development; it is not live financial evidence.
- Judge scores are evaluation signals, not a guarantee of financial correctness.
NEXT PROJECT / 03
Have an idea of your own? Let’s talk