FinRet

Inspiration

Financial research is full of signal - but buried in filings, earnings transcripts, and reports that are too large and inconsistent for a single AI context window. We built Finance Research Harness to make AI-assisted research evidence-first: every answer should be traceable to the exact source material behind it.

What it does

The harness ingests financial research from MCP data source, structures the corpus into searchable evidence, and answers research questions with verified citations. It is designed to handle noisy or conflicting data honestly rather than confidently inventing an answer.

How we built it

We built a Python service with FastAPI, PostgreSQL, and pgvector, plus a React research chat interface. Our pipeline:

Snapshots raw documents immutably and creates versioned datasets
Extracts documents, source spans, chunks, tables, facts, and findings
Combines lexical and semantic retrieval with structured fact lookup
Uses a bounded OpenAI tool loop and deterministic calculations
Records every answer, citation, tool call, and dataset version for auditability

The system can refresh a complete MCP corpus, reuse unchanged versions by hash, and immediately ingest a new dataset - making it suitable for the surprise-data challenge.

What we learned

The hard part of reliable AI research is not generating polished prose; it is building the evidence infrastructure around the model. We learned that source-level citations, versioned data, read-only evidence access, evaluation suites, and explicit conflict detection are what turn an LLM response into a research result someone can trust.

Challenges we faced

The biggest challenge was designing for ambiguity. Financial documents can restate values, use inconsistent labels or currencies, and contradict one another. We addressed this with structured extraction, validation findings, citation verification, and an audit trace that lets users inspect how each answer was produced.

Built With

agent
ai
harness
openai
python
react

Submitted to

Hack the North 2026
- Winner RBC: Signal in the Noise

Created by

I built FinRet's research engine and ingest pipeline, which answer every question with citations from the source documents. On ingest, the pipeline parses financial reports it hasn't seen before into structured, versioned evidence. That includes heading paths, table chunks split by row, embeddings and typed facts such as periods, estimates, ranks, counts, trends and events. It also pulls out the companies and other entities each document is about, with their aliases. Validators then check the extracted facts for consistency.

On the research side, I built a hybrid search that combines keyword and embedding matching. It matches entities by name, re-scores candidate answers, and declines to answer when the evidence doesn't support one. I also hardened the API: connections are shared, concurrency is capped, and the engine streams each step live, so users can watch the search, candidates and verification happen. I set up the MCP data-source integration and the database role grants, so the engine can read evidence and can only write its own audit records.

To prove it works, I built a scoring system for our RBC question set. It checks answers against the evidence in the answer key, and a single command runs the whole check from a clean install and produces a scorecard. FinRet scored 29/30 on a corpus it had never seen, with no changes to the code, prompts or schema. I also redesigned the Research Desk frontend (the citation panel, conversation history and the Sources workflow), wrote the setup documentation and put together our demo script.

Ridaan Gupta
- Designed an ingestion pipeline for MCP data sources. It preserves immutable, hash-addressed source snapshots and creates versioned datasets, so each answer can be traced back to the exact material that was ingested.

- Built document processing that turns reports into searchable evidence: it extracts entities, aliases, reporting periods, table-row chunks, and typed facts, while retaining the section and source span for context.
Implemented hybrid retrieval combining lexical search with vector search, then connected retrieved evidence to an answer engine with verification and citation checks.

- Added validation for conflicting claims, counts, rankings, trends, and mixed currencies to help surface data quality issues instead of silently treating every extracted value as reliable.

- Built an agent loop that searches evidence, uses tools for calculations, and verifies answers and citations before responding.

Pravin P.
Wasi Rehman