Skip to content
View RAFAELDCOELHO's full-sized avatar

Block or report RAFAELDCOELHO

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
RAFAELDCOELHO/README.md

Rafael D. Coelho

Independent researcher. From-scratch machine learning, and open tooling for the Brazilian financial system.

Python PyTorch NumPy FastAPI On-device

Every repository below is public, runs offline where it can, and states what it does not prove.

PersonaCore teach-then-recall demo: the same question asked with the persona adapter on and off; the memory ON turn answers with the taught dog name, the memory OFF turn does not, and the token-id panel shows an identical prompt

PersonaCore — one question, twice. The prompt token ids are identical both times; only the 331,776 adapter parameters were switched off in between. Memory in the weights, not in a prompt or a database.

Research — machine learning built by hand

Project What it is Proof you can run
PersonaCore A from-scratch 13.9M-parameter GPT — tokenizer, decoder, training loop, LoRA and EWC all hand-written in PyTorch — where personalization lives in the weights, not a database. Includes an adversarial privacy audit of what LoRA memory reveals under black-box attack, and a selective-erasure attempt whose failure is published in full. make demo → streaming story demo on a laptop CPU. Slim checkpoint: m1-demo-v1
tensorforge A deep learning framework from scratch: scalar autograd, NumPy tensors, Transformer primitives, full training loop. A real, already-trained GPT checkpoint was retrained on it and the numbers match PyTorch forward and backward — not a reimplementation validated against itself. Parity tests against PyTorch
personacore-np Inference engine for the same 13.9M GPT in pure NumPy, with optional MLX on Apple GPU. No PyTorch at serve time. make infer → downloads the public checkpoint and runs the parity suite

Brazilian finance — SDKs and infrastructure

Project What it is Status
brazilfi Unified Python SDK for Brazilian financial-market APIs: Bacen, IBGE and Tesouro Direto behind one library, one API. PyPI MIT
rfbcalc Typed Python client for Receita Federal's official CBS/IBS consumption-tax calculator. Wraps the government engine; invents no tax rules. PyPI MIT
pix-billing PIX-native billing API for LatAm SaaS: subscriptions, dunning, signed webhooks, valid BR-Code QR payloads. Async, type-safe, FastAPI. Sandbox mode · MIT
autonomous-hedge-fund Multi-agent LLM trading research on emerging markets, extending TradingAgents with a Macro Economist agent. Ships BrazilBench: offline classical baselines on frozen B3 splits. make bench — no API key, no LLM · Apache-2.0

Also

  • atena-lang — a teaching language with no punctuation noise that transpiles to real Python 3, for people who have never programmed. atena-lang.org

Zero-budget, on-device, reproducible: fixed seeds, pre-registered gates, and negative results published alongside the positive ones.

Pinned Loading

  1. autonomous-hedge-fund autonomous-hedge-fund Public

    Multi-agent LLM trading system with Kronos, FinGPT and TradingAgents

    Python 2

  2. brazilfi brazilfi Public

    Unified Python SDK for Brazilian financial market APIs

    Python

  3. PersonaCore PersonaCore Public

    From-scratch, on-device GPT with weight-based personalization — an adversarial privacy audit measuring what LoRA memory reveals under black-box attack and what selective erasure can (and cannot) un…

    Python