Skip to content
View luizmlo's full-sized avatar
👾
Researching a ton
👾
Researching a ton

Block or report luizmlo

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. 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
luizmlo/README.md

Luiz Felipe Melo

Security engineer and ML researcher based in Tubarão, Brazil. I'm currently working on AI Safety and Offensive Security Capabilities. I have about 8 years of experience in software development and 4 in cybersecurity.

ORCID · LinkedIn

Highlights

  • Research (2026): Agentic Weight-Space Programming as an Alternative to Fine-Tuning in Small Language Models. Claude agents repair a thinking loop in a small language model by editing its weights directly. Pre-paper · code, data and lab notebook
  • Security: built and led a vulnerability-management service monitoring 15,000+ assets for 25+ corporate clients.
  • CTF: 1st place in the RNP CTF at Campus Party Brasília 2019.

Research

Agentic weight editing (2026) · pre-paper, DOI 10.5281/zenodo.24208535 · repository

  • Question: can an AI agent repair a defect in a language model by editing its weights directly, without training?
  • What I did: Agents read the weights of Qwen3.5-0.8B and wrote at most four rank-1 edits that stop its endless reasoning loops, on a single gaming GPU. On locked GSM8K test problems, accuracy went from 15.5% to 61.5–68%, and six agent runs of the same protocol all repaired the defect.
  • Compared against: the edit that GPT-6 Astra wrote in MIT CSAIL's Sparks of Weight-Space Programming, which reaches the same accuracy only with about twice the tokens; LoRA fine-tuning, which falls well short; and simple decoding baselines, which come close. Switching thinking off is cheaper still.
  • Method: every prediction was pre-registered in a public lab notebook, two test sets stayed locked until the end, and the predictions that failed are reported. The same method failed on a diffuse defect, and the paper says why.

Interests: mechanistic interpretability, weight-space editing, LLM agents, AI security.

Security

  • Vulnerability management at scale. At OSTEC Business Security I built and led a continuous vulnerability-management service from scratch. It monitored 15,000+ assets daily for 25+ corporate clients in industry, retail and critical infrastructure (Tenable Nessus and in-house Python automation).
  • Penetration testing. I have tested web applications, APIs and networks for large companies in retail and construction, including a publicly listed multinational, from kick-off to executive report. Clients are under NDA.
  • Attack-surface mapping. I built Scandara, an event-driven scanning platform (FastAPI, Arq, PostgreSQL, Redis) that processes about 50,000 URLs a day.
  • CTF. At Campus Party Brasília 2019 I placed 1st in the RNP CTF and 2nd in the Decred CTF.

Talks

All talks were given at Universidade do Sul de Santa Catarina (Unisul), Tubarão.

Year Talk
2026 End of the World by AI: Where Are We on the Timeline?
2026 Real Bitcoin: Myths, Blockchain and Decentralization
2025 Practical AI: Self-hosting LLMs (workshop)
2023 Exploits in Web Applications and APIs (live red-team demo, OWASP Top 10)

Now

  • Scale Tech: software developer and head of support. I contribute to Scale One, a CRM in which AI agents run WhatsApp sales conversations, and I build the support team's tooling, including an episode classifier using a small decision model (laya)
  • Codara: security and platform engineer.
  • DATUS (Unisul's IT student association): events director. So far I have organised two hackathons, an AI workshop and technical talks.
  • Unisul: BSc in Computer Science, expected 2027.

Selected open source

Project Description
agentic-weight-editing Claude agents edit the weights of Qwen3.5-0.8B to fix its thinking loops: code, lab notebook, every edit (Python, PyTorch, llama.cpp)
a3_botnet Botnet research prototype in Python (asyncio, WebSockets)
dit HTML page, form and field classifier (logistic regression + CRF), in Go
EasyLSTM Data shaping, training and inference for LSTM networks in Keras
aim-ml Training neural networks from screenshots
aic8800d80-kernel7 AIC8800D80 USB Wi-Fi driver patched for Linux kernel 7.0

Tools

ML        PyTorch · transformers · llama.cpp · Keras · scikit-learn
Security  Burp Suite · Nmap · Metasploit · Tenable Nessus · Wazuh
Code      Python · Go · C/C++ · JavaScript · SQL
Systems   Linux · Docker · PostgreSQL · Redis · FastAPI

Languages: Portuguese (native), English (fluent), Spanish (basic).

Pinned Loading

  1. agentic-weight-editing agentic-weight-editing Public

    Claude agents repair a thinking loop in Qwen3.5-0.8B by editing its weights directly (rank-1 edits, no training). Pre-paper, code, lab notebook and data.

    Jupyter Notebook 1

  2. aim-ml aim-ml Public

    Train Neural Networks with screenshots

    Python 2

  3. AlbionTrading AlbionTrading Public

    Forked from hrichharms/AlbiPy

    Framework to analyze Market prices by sniffing network packets, fork of AlbiPy

    Python

  4. PimentelM/mystera-reborn PimentelM/mystera-reborn Public

    A mod for the game Mystera Legacy which added an API to interact with every aspect of the game. I Then used this to create full automation algorithms in a State Machine Based Architecture.

    JavaScript 7 4