Bartosz Lenart
Iturncomplexbusinesslogicintoreliablesoftware.
Safe. Optimized. Maintainable.
Curious about how minds and machines reason. Drawn to life, nature, physics.
Currently focused on AI reasoning systems and multi-agent architectures.
ABOUT
My interests are hardwired to:
- AI & Machine Learning
- Brain-computer interfaces
- Robotics & Automation
- Neuroscience
- Quantum computing
- Physics & Medicine
Writing: essays on AI reasoning, consciousness, and the neuroscience of thought.
What I build
Focus: AI systems, multi-agent workflows, and regulation-aware stacks for domains where failure is costly.
End-to-end: systems, product architecture, and UI/UX.
Interface, deployment, and maintainability get the same attention as the internals.
The standard: measurable behavior, traceable decisions, and code that survives handoff.
Technology Stack:
Python · FastAPI · LangGraph · Pydantic · GenAI · ML · React · Docker · SQL · Cloud
- Evaluation & verification: benchmarks, regression catches, and explainable reasoning.
- Medical AI & FHIR-native tooling for clinical workflows with auditable state.
- Financial crime prevention: AML and anti-fraud pipelines with traceable decisions.
- Multi-agent orchestration: typed contracts, observable handoffs, bounded tool use.
- Legacy modernization: data pipelines and migration paths for AI adoption without downtime.
BLOG.CKPT
LATEST
The Map Without a Vision: One Sicilian Lemon, Many Cognitive Perspectives on Knowing Without Seeing
Much of brain computation stays below conscious access. Only a small subset of competing states ever wins broadcast. A family of conditions shares one fault line where the map and the report come apart. That split is also the architecture we are building agents on.
TOP.READS
The Mirror Has No Face: Why AI Only Sounds Conscious When You Ask It To
LLMs produce self-awareness language only when prompted. Empirical evidence shows zero unprompted psychological language in 100,000 neutral prompts, context-dependent register switching, not emergent consciousness.
AI as Cognitive Prosthetic: You are the art, AI is an interface, the result is value.
Moving from answer generation to insight extraction. Agentic systems should not replace human thought - they're high-performance 'cyberpunk chrome' designed to extract and amplify the wisdom already present in your mind.
Beyond Tree-of-Thought. Yggdrasil: Parallel AI Reasoning Architecture
A single tree still grows in one direction. Yggdrasil doesn't just explore parallel paths. It detects when insights from one cognitive realm unlock breakthroughs in another, then builds the bridge.
The Event Horizon of Thought: Where Attention Meets Architecture
Exploring the hard limits of attention - whether human working memory or AI context windows - as fundamental structures to be navigated. A framework for designing workflows that respect the physics of thought.
The Geometry of Compression: Why Angles Beat Naive Storage
A visual essay on Moroccan zellige, transformer KV-cache compression, and TurboQuant: good compression keeps the pattern that still lets a system work.
The Neural Bifrost: Bridging Deep Brain Signals and the Surface of Consciousness
A 2026 paper put electrodes inside ten human brains while patients tried to see a near-invisible flicker. The result is a high-resolution view of how conscious access is gated, and the architecture is a bridge, not a destination.
PROJ.DIR
Syntrace
One markdown file for project history, decisions, and reusable lessons. Portable, local-first, and built to help LLM-assisted workflows preserve context across sessions without databases or API keys.
FHIR Flame
FhirFlame is a modular, Dockerized medical AI toolkit for processing healthcare data with multi-agent workflows. It features FHIR R4/R5 compliance, NLP entity extraction, DICOM support, local/cloud deployment (GPU-optimized), and secure API access.
Gaia A1
General purpose multi-agent AI automation framework (Python + FastAPI + LangGraph) with REST APIs, scalable Docker deployment, dynamic tool integration, and strong process automation features. Achieves 55% GAIA L1 benchmark accuracy.
Can Run
CanRun is a game compatibility checker that works as a NVIDIA G-Assist Plugin. It instantly analyzes your PC's hardware against game requirements, predicts performance, and provides upgrade recommendations-all through simple voice commands or API calls.
NEWSLETTER
New essays on AI reasoning, cognition, and system design.
No noise.
Available for selective consulting on AI reasoning systems, multi-agent product architecture, and high-stakes AI where explainability and compliance matter.