+ INDEPENDENT RESEARCH & ENGINEERING
Online learning is my starting point.
Embodied intelligence is where I’m taking it.
N-TEK is where I bring my research and engineering together. My commitment is online learning: intelligence that keeps learning through interaction with a changing world.
I built NEMESIS to learn through continuous control. NEXMIND grew from that engineering into my main intelligence architecture: persistent belief, multiple horizons, and online adaptation. NOMAD carries this work into physical-platform design.
A note from the ArchitectMY CONTRIBUTIONS TO ACTIVE INFERENCE
Models that can change structure.
Evidence that keeps its context.
DEMONSTRATED WORK
I build systems that learn, run experiments that answer questions, and carry what works into the next architecture.
I demonstrated autonomous search and target finding in simulation, bringing perception, persistent belief, multirate inference, and safety-gated guidance together.
I sustained autonomous operation for more than 34 minutes in a persistent, populated simulation, with all five temporal levels active.
I trained NEMESIS across tens of billions of simulated timesteps, developing learned control with custom GPU-accelerated training infrastructure.
ONLINE LEARNING IS MY ORGANIZING PRINCIPLE.
I’m developing NEXMIND around a changing world, multiple horizons, and experience that keeps shaping what comes next.
Horizon bands shown on a logarithmic scale.
INTERMEDIATE HORIZON
The 16-second horizon bridges immediate choices and longer-range intent. I use it to organize action over a meaningful sequence of events.
I separate how far a level predicts, how often it replans, and the stride of its predictive rollout. Longer-horizon reasoning does not impose one slow clock on local guidance.
The prediction horizons are 1, 4, 16, 64, and 256 seconds. I connect longer-term context to immediate decisions, with evidence moving back through the hierarchy.
PERSISTENT WORLD BELIEF
I organize belief around objects that persist through time: their identity, motion, uncertainty, relevance, and the evidence of where they are—or are no longer.
Search involves more than detection. Occlusion, reappearance, and missing observations all matter. My perception architecture keeps learned semantic evidence and independent geometry distinct as they inform belief and planning.
02 / ADAPTIVE MODEL STRUCTURE
I’m developing NEXMIND to do more than adjust an existing prediction. My research also asks when one explanation should become several, and when unnecessary distinctions should disappear.
A surface might behave differently when wet. A model could begin with one explanation, then consider whether separating those conditions helps predict what happens next.
I compare proposed changes against subsequent experience. Replaying the past can refine a candidate; fresh observations must provide the evidence for its structure.
Let both explanations predict subsequent experience before learning from it.
03 / EVIDENCE THROUGH THE HIERARCHY
I want the hierarchy to share information without mistaking repetition for confirmation. An observation can shape an immediate decision, inform a longer horizon, and return as context. Its origin still matters.
My architecture tracks where information came from and whether different reports share the same source. Related evidence needs different treatment from independent observations.
Confidence should follow what the world adds, not how many times information travels through the system.
04 / EXPERIENCE & LEARNING
I’m preparing the scientific description of how I organize experience for online learning.
PUBLIC TERMINOLOGY / IN PREPARATION05 / LEARNING FROM PREDICTION ERROR
I’m developing mechanisms that compare predicted consequences with observed outcomes and direct greater learning attention to meaningful model errors.
My aim is to revisit the relevant ordered sequence at higher learning exposure, while keeping its context intact. Replay should deepen learning without pretending that the same experience is new evidence.
Replay the experience.
Keep the evidence intact.
I use expected free energy to reason about possible actions, uncertainty, and preferred outcomes. Prediction-error learning asks a different question: did the world behave as the model predicted? An unwanted outcome may still have been accurately predicted.
I’m continuing to implement, test, and refine these mechanisms, with online learning at the center of the architecture.
See how I’m bringing the research into the workMy intelligence system. My physical platform.
Research that keeps both moving forward.
01 — THE PHYSICAL PLATFORM
Loading 3D model…
Move to deform · Drag to rotate · Scroll or pinch to zoomKeyboard: arrow keys · + / − · Home to reset
I’ve established a whole-aircraft digital design. I’m refining NOMAD Mk. II’s geometry and running fixed-airframe CFD, feeding aerodynamic analysis back into the next design iteration.
Whole-aircraft CAD established. Geometry refinement and aerodynamic analysis advancing.02 — THE INTELLIGENCE PROGRAM
My main intelligence system. Built around online learning.
NEXMIND is my main intelligence system. I’m advancing its native software, online model adaptation, predictive perception, and integration with physical systems.
I use controlled simulation to investigate active inference and online learning. This research helps me test ideas about adaptation and decision-making as I develop NEXMIND.
HELICOPTER-SCALE SIMULATION
I’ve flown 700 kg-class helicopters in simulation with NEXMIND, online learning, and safety gating operating during flight.
I’m carrying this experience into the next generation of NEXMIND and NOMAD’s physical-platform engineering.
03 / IMPLEMENTED LEARNING SYSTEMS
Where learning became
an engineering discipline.
I built NEMESIS as a GPU-accelerated C++/ROCm reinforcement-learning system for multi-agent continuous control. Its development spans PPO, self-play, opponent curricula, predictive learning, and detailed runtime diagnostics.
My implemented research lineage also includes recurrent world models, DreamerV3, PonderNet adaptive computation, and high-performance LibTorch training and inference.
NEMESIS achieved learning. Once it answered the questions I was pursuing, I carried that foundation into persistent world state, explicit uncertainty, temporal organization, and online model revision. NEXMIND grew from that experience.
IMPLEMENTED RESEARCH LINEAGE · ORIGINS IN 2024HOW I MOVE THE WORK FORWARD
I move between formal reasoning, executable software, controlled simulation, and digital design. Each makes the next question more concrete.
I work on how an adaptive system should learn, organize its model, and use evidence to guide decisions.
I turn the mechanism into code, then test whether its behavior matches the reasoning behind it.
I investigate what happens when predictions meet changing conditions, and use those failures to sharpen the next experiment.
With NOMAD, I work through geometry, assembly, and analysis. Physical constraints make the broader ambition concrete.
+ Simulated flight is already part of my research history. I’m advancing the current architecture and aircraft design toward physical deployment.
What I’ve established. What I’m advancing.
What I’m building toward.
FOUNDATIONS ESTABLISHED
I’ve built high-performance learning systems, developed original active inference contributions, demonstrated integrated simulation, and established a whole-aircraft CAD baseline.
ADVANCING NOW
I’m advancing NEXMIND, my active inference research, and NOMAD’s design and analysis.
THE LONG-TERM DIRECTION
I’m working toward continually learning intelligence in physical systems, with capability earned through evidence.
İBRAHİM ALİCAN / FOUNDER & ARCHITECT
I’m İbrahim Alican.
This is N-TEK.
I’m an Aerospace Engineering undergraduate at Istanbul Aydın University. Through N-TEK, I bring inference and learning research, native software, controlled simulation, and rotorcraft design together.
I came to active inference through building learning systems. NEMESIS taught me what it takes to make a policy learn. NEXMIND asks how an autonomous system should maintain beliefs and revise its understanding while it operates.
I set the architecture and connect those disciplines under one direction: continually learning intelligence with a path into physical systems.
Explore what I’m buildingMy interest in systems that learn through interaction begins.
I build the first systems in a sustained reinforcement-learning project.
My technical program begins as learning infrastructure scales and active inference becomes central.
Recorded embodied simulation, ongoing cognitive architecture, and a whole-aircraft digital design.
