Technical architecture

How the AI research pipeline works.

From an official company release through AI extraction to a predictive model and an independent final test. Collection, language interpretation and evaluation remain separate; each step retains auditable intermediate files.

01 · Sources

Datasets

SEC 8-K filings, daily and minute-level prices, sector benchmarks and publicly disclosed Form 4 transactions.

Raw source files
02 · Processing

Python + pandas

Downloading, cleaning, aligning timestamps, selecting events and building reproducible tables.

Chronological events
03 · Language

Ollama + Qwen3-4B LLM

Local semantic extraction of financial developments, guidance, risks, capital allocation and evidence excerpts.

Structured features
04 · Models

scikit-learn

Baselines, Ridge regression, text vectors and boosting are compared using the same chronological splits.

Frozen predictions
05 · Evaluation

Holdout pipeline

Directional accuracy, correlation, uncertainty intervals, transaction costs and independent replication.

Audited final test
06 · Publication

Public report

Results, limitations, example records, documentation and live progress are published without selectively filtering the findings.

Website + downloads

Two roles for AI

AI supports both development and research.

AI-assisted development · vibe coding

AI assistants help develop the website, prototypes and research software. Generated code is tested; software development is not evidence of a financial relationship.

AI in the research pipeline

Qwen3-4B LLM extracts features and evidence excerpts from selected financial passages. A separate model then examines their relationship with subsequent price reactions. Statistical analyses and simple baselines remain part of the comparison.

Read how Feronix is developed

Separation of responsibilities

Why the components are not combined into one black box.

DATA

Source before model

Raw documents and prices are retained before features or predictions are created.

LLM

Extraction, not knowledge of future prices

Qwen3-4B LLM receives financial text, but no future price reactions are included in its prompt.

MODEL

Decisions before the holdout

Model settings and benchmarks are fixed before the final test set is evaluated.

REPORT

Findings and limitations

A null result, an uncertain interval or a failed replication remains part of the published record.

Technology stack

Local and reproducible by design.

Python Orchestration, data processing and reporting
pandas Time series, events and tables
Ollama Local model server without external text processing
Qwen3-4B LLM Semantic extraction from earnings releases
scikit-learn Baselines, pipelines and model comparison
Vercel Public research website and progress updates