J-Space Deliberation Engine
A structured latent reasoning architecture exploring whether internal model representations can be actively organized into a deliberative workspace before generation.
Explore research →
Building at the intersection of
intelligence and software
I build AI systems from the model weights to the interface.
I research and engineer language models, latent reasoning architectures, retrieval systems, developer tools and experimental software. My work moves between fundamental AI research and the systems that turn those ideas into something people can actually use.
AI Research
LLMs
Latent Reasoning
Software
01 / Research
Exploring what happens inside the model.
Experimental work on latent representations, structured deliberation and the relationship between internal reasoning and generation.
RESEARCH / 001
J-Space Deliberation Engine
Structured Latent Deliberation for Gemma 4 E4B using Jacobian Lens.
J-Space is an experimental latent reasoning architecture that introduces a structured five-slot workspace into the residual stream. The system organizes internal deliberation into Objective, Hypothesis, Evidence, Critic and Commit before final text generation.
Objective
Defines the goal
Hypothesis
Candidate solutions
Evidence
Relevant signals
Critic
Consistency check
Commit
Final consolidation
System Architecture
User Prompt
↓
Gemma 4 E4B-it
↓
J-Space Workspace
↓
Jacobian Lens
↓
Final Generation
02 / Models
Models built to do something.
Specialized models designed around coding, technical reasoning, multimodal understanding and local inference.
Nexora-Gemma-4-E4B-Coder
A coding-focused multimodal language model fine-tuned from the Gemma 4 family and converted to GGUF for efficient local inference. Designed for programming, debugging, refactoring, technical reasoning and multimodal code understanding.
03 / Selected Work
The rest of the experiment.
Research is only one part of the work. These projects explore software, AI, interfaces and systems from different angles.
Nexora
A coding-focused multimodal model built for technical reasoning, programming and efficient local use.
View model →Projeto Lunar
An independent software project exploring a different direction of engineering and experimentation.
Open repository →Lumi Motion
An experimental project exploring motion, interaction and digital experiences through technology.
Open repository →Agner Browser Project
An independent browser project exploring software construction at the application and platform level.
Open repository →MergeCraft
An experimental software engineering project focused on building and exploring new development ideas.
Open repository →IA RAG Nexa
An AI project exploring Retrieval-Augmented Generation and the connection between language models and external knowledge.
Open repository →
04 / Areas
What I work on.
01
Latent Reasoning
Structured internal representations and experimental reasoning architectures.
02
Large Language Models
Fine-tuning, model behavior, coding models and technical reasoning.
03
RAG Systems
Connecting language models with external knowledge and retrieval systems.
04
Multimodal AI
Models that combine language with visual and technical information.
05
Local AI
Quantization and inference systems for practical local deployment.
06
AI Infrastructure
The tools and systems required to build and run modern AI.
07
Software Engineering
Building complete applications, tools and experimental platforms.
08
Experimental Systems
Projects that explore ideas outside conventional categories.
05 / Stack
Tools of the trade.
Python
PyTorch
Transformers
Hugging Face
Unsloth
llama.cpp
GGUF
BitsAndBytes
Accelerate
CUDA
RAG
LLMs
Multimodal AI
Software Engineering
06 / About
Miguel Penha Reis
I'm an AI Engineer and researcher working across large language models, latent reasoning, model fine-tuning, multimodal AI and software systems.
I build both the models and the systems around them, moving between research experiments and practical engineering.
My work explores what happens inside intelligent systems, how their internal representations can be structured, and how those ideas can become useful tools in the real world.
Build the model.
Build the system.
Understand both.