Vesta AI
Autonomous synthesis of SEC filings and earnings call transcripts.
An autonomous research agent designed to parse structured and unstructured financial disclosures. Synthesizes 10-K/10-Q filings, transcripts, and financial balance sheets into traceable, citation-backed analytical briefs.
The Context
Financial analysts spend days reading 100+ page SEC filings, financial footnotes, and earnings call transcripts to extract company guidance and risk factor disclosures. Vesta AI was built to automate the extraction and synthesis workflow while adhering to strict compliance standards requiring verifiable citations for every data point.
Core Constraints & Failure Modes
Standard LLM RAG pipelines fail on financial disclosures because balance sheets and cash flow statements lose spatial table structures when chunked into simple text paragraphs. Furthermore, financial analysis cannot tolerate hallucinations: citing an incorrect revenue metric or EPS guidance can have severe commercial consequences.
Architectural Solution
We engineered a dual-stream multi-agent architecture. Tabular statements are extracted deterministically using table parsers to preserve row-column semantics, while textual risk disclosures are embedded into a vector space with document metadata. An evaluator agent checks the synthesizer agent's output, requiring exact line and page citations from the SEC filings before output is released.
Verified Scope & Lifecycle Boundaries
Strict truth in engineering: we clearly separate what is functional in code, what is bounded by prototype scope, and what represents future architectural roadmap.
- Deterministic HTML/XBRL financial table parser preserving row and column coordinates
- Vector embedding and retrieval pipeline for textual risk factor disclosures
- Multi-agent synthesis graph separating extraction from audit verification
- Citation assertion layer rejecting any claim lacking exact filing chunk citations
- Operates over curated historical 10-K and 10-Q disclosures for targeted evaluation companies
- Vector index stored locally for prototype benchmarking
- Real-time SEC EDGAR live stream ingestion and filing notification webhooks
- Automated cross-statement 3-statement financial model reconciliation
- Fine-tuned domain classification models for non-GAAP reconciliation footnotes
System Topology & Component Decomposition
Multi-agent extraction graph, semantic vector indexing, citation verification pipeline, deterministic tabular parsing.
Tabular Disclosure Parser
Python / BeautifulSoup / PandasPreserves hierarchical financial tables, converting SEC HTML/XBRL disclosures into typed tabular data frames.
Multi-Agent Graph Orchestrator
FastAPI / LangGraph / Vector DBCoordinates specialized agents: an Extractor Agent, a Financial Modeling Agent, and an Auditor Agent with stateful checkpointing.
Citation Verification Harness
Python / Pydantic / RAGDeterministic verification engine that validates every extracted claim against SEC EDGAR ground-truth text chunks.
Critical Architectural Choices
Table Normalization Before Vectorization
Generic text chunking splits balance sheet rows across chunks, destroying numerical context.
Tables are converted to markdown/JSON structures with column headers preserved on every row entry.
Auditor Verification Loop
Single-pass generation occasionally quotes figures from previous quarters or misinterprets guidance ranges.
The secondary Auditor agent cross-checks every synthesized number; unverified statements are flagged and omitted from the final brief.
Complete Technology Stack
Interested in deploying a similar architecture or commissioning a custom digital engineering engagement based on this system pattern?