About VerifiQuant

The Problem

Large Language Models are increasingly used for financial computations, but they fail silently. A wrong NPV, a misinterpreted discount rate, or an invented input can propagate through downstream decisions with no warning. Current approaches focus on making the model "more accurate" — but accuracy without verifiability is a liability.

The Approach

VerifiQuant doesn't try to make LLMs better at math. Instead, it builds a diagnostic funnel around the computation: six layers (M/N/F/E/I/C) that classify failure modes, intercept broken inferences, and apply mathematically verified repairs when possible. The system uses Financial Inference Contracts (FICs) — pre-built, auditable computation cards with declared invariants, scale checks, and semantic hints.

Key Ideas

  • Predictable failure — Every failure is classified, not hidden
  • Verifiable transforms — I-class repairs are bound by algebraic invariants
  • No SymPy, no magic — Pure Python AST analysis + numerical verification
  • Normalization is the hard gate — LLM prompts are aspirational; post-processing filters are the real safety net
  • Deterministic layers are the stable part — E/F checks don't vary run-to-run

Tech Stack

Model Gemini 2.5 Flash
Backend Python 3.11 / Flask / SQLAlchemy
FIC Store SQLite + FTS5
Retrieval BM25 + Sentence Transformers
Deploy Google Cloud Run
Verification AST analysis + numerical checks

Author

Built by Vincent Ko aka Datafox. For more projects and writings, visit datafox.tw.