Back to projects
Featured projectIn developmentSince September 2026

SIFA

Autonomous Financial Intelligence System

A self-hosted financial research agent that collects, verifies and contextualizes market information, with a multi-agent investment committee. Personal project under development.

Python
FastAPI
LLM
llama.cpp
NLP
PostgreSQL
TimescaleDB
pgvector
Prefect
Docker
per analysed document
2.4 s
per analysed document
end-to-end latency
47 s
end-to-end latency
of VRAM, fully local
8 GB
of VRAM, fully local

Overview

SIFA starts from a simple observation: an individual investor cannot follow on their own the thousands of daily publications (regulatory filings, press releases, news, earnings calls), most of which are noise or syndicated copies. The system continuously collects structured data (financial statements, prices) and unstructured data (news, regulatory filings, transcripts), verifies each claim, extracts typed events anchored in their source, then relates them to the company's history, its sector and market expectations. An investment committee of AI agents then debates decisions and logs them. Its guiding principle: it does not predict prices, it lowers the cost of research and makes every decision traceable. It runs entirely locally, on a single workstation with an RTX 3070 Ti (8 GB of VRAM).

Key features

  • Continuous collection: regulatory filings (EDGAR 8-K, Form 4), press releases, news, prices and financial statements
  • Filtering cascade: small CPU models (fastText, DeBERTa, FinBERT, GLiNER) before the GPU language model
  • Event extraction as constrained JSON by a quantized 8B model, always anchored in the source document
  • Calibrated confidence score, detection of contradictions and of copies of the same press release
  • Per-company investment view on five independent axes, recomputed from dated atomic signals
  • Multi-agent investment committee with anti-sycophancy mechanisms and an immutable decision log
  • Near real-time alerts weighted by confidence and portfolio exposure

Context

Personal project under development. I wrote a complete specification (vision, constraints, architecture, model choices, data, delivery plan) before writing the first line of code. Phase 0, dedicated to benchmarking the target hardware, is complete: the throughput figures in the specification are now measured rather than estimated.

What I did

  • Designed a two-speed architecture for 8 GB of VRAM: a single 8B model resident on the GPU for extraction, and a ~20B MoE model loaded on demand in RAM for the committee's reasoning.
  • Phase 0 benchmark: 89 to 95 tokens/s generation, 2.4 s per extracted document, about 47 s end-to-end latency against a 90 s target.
  • Economy principle: each task goes to the smallest model that handles it correctly, so the GPU only works about seven minutes a day.
  • Four requirements on every conclusion: sourced, scored, falsifiable and logged.
  • Planned stack: Python 3.12, FastAPI, llama.cpp, PostgreSQL + TimescaleDB + pgvector, Prefect, Docker Compose, then a Next.js interface.