/aiman.tech
01Selected work · 2021–2026

Selected work

Research, products, shipped titles and tooling. The lines I am most invested in have their own deep-dive page - the rest are listed below in case order. Tap a card to go deeper.

[R-001]
Research
Year
2024 - present
Role
Research · author · trainer
Stack
PyTorchCUDACustom kernelsCWSDEQ-flavoured

Parallax

A modality-agnostic cognitive core that settles into a metastable regime in a strange-attractor landscape. Cognition decoupled from token-clocked autoregressive generation. Reasoning depth at O(1) memory.

Parallax is the line I have been chasing for three iterations. By v3 it is a Central Workspace (CWS): a continuous-state dynamical system whose only job is to settle into a metastable basin in an attractor landscape shaped by associative perturbations from a heterogeneous expert pool. The CWS is not a language model. Tokens are sensory observations, summarised by experts; the workspace integrates them through K reasoning steps that are decoupled from token steps; a thin motor decoder reads the settled state. Equilibrium is detected via momentum saturation, not by chasing a fixed point. In the limit, this is a candidate replacement for the residual stack itself - reasoning depth as a separate axis from parameter capacity, paid for at constant memory.

Cognition detached from autoregressive generation - K reasoning steps ≠ token steps
DEQ-flavoured contraction toward a metastable regime, not a fixed point
O(1) memory in K verified - K=48 reasoning depth at 12-layer-TX memory cost
Modality-agnostic core - language is one expert / motor pair, not the substrate
Public lab notes; v3 code private until the 'replaces ResNet' claim survives scale
[R-002]
Research
Year
2026
Role
Author · researcher
Stack
PyTorchCUDATransformers

Dendritic Unit

A side track. What if you took the activation function off the MLP and let the unit itself be nonlinear, through learned competition between dendrite-style branches.

An exploration of structural plasticity at the unit level. Each block holds K small affine branches that compete via a learnable selection rule - sharp enough that one branch can dominate, soft enough that several can coactivate. There is no ReLU/SiLU/GeLU anywhere in the block; the nonlinearity is the structure of the selection. Lives on its own track, not part of Parallax.

Selection between learned affine branches replaces a bolted-on activation
Branch specialization emerges during training (uneven, but informative)
Drop-in compatible with a standard transformer MLP slot
Public research log - notes, code, ongoing experiments
[T-003]
Tooling
Year
2026
Role
Creator · maintainer
Stack
RustC#WebSocketJSON-RPC 2.0

Unity Control Protocol

A Rust CLI + C# editor bridge that exposes Unity to agents over JSON-RPC. For the AI-native Unity workflow.

UCP turns the Unity Editor into a structured tool surface. Scenes, GameObjects, components, materials, prefabs, build pipelines, tests, packages, profiling - all addressable from the terminal or from an agent over a WebSocket bridge. Built because heavy agentic dev workflows in Unity were impossible without it.

Full editor & lifecycle access from outside Unity
Selective .unitypackage import & build-pipeline control
Native debugging and profiling hooks
Distributed as a Claude Code skill - agents can drive the Editor end-to-end
[P-004]
Product
Year
2024 - present
Org
Brainseed Factory · Bonn
Role
Backend & AI engineer
Stack
TypeScriptPythonNLPTTS · STTAgentic runtime

Monamy

An AI-native CRM running fully autonomous mass campaigns and channel-agnostic support - phone, chat, anywhere. monamy.de.

Evolution of 'Mona', a Unity-side AI assistant demo, into a production CRM platform - see monamy.de. I work on the backend that drives autonomous outbound campaigns over the phone and platform-agnostic chat support - orchestration, prompt + eval loops, channel adapters, and the production glue that keeps autonomous agents from doing dumb things in front of paying customers.

Fully automated mass-campaigns over voice & chat
Platform-agnostic conversational layer (CRM, web, phone)
NLP / TTS / STT pipelines wired into the runtime
Production AI for B2B - undisclosed customer base
[S-005]
Shipped
Year
2024
Org
XR Hack Cologne · Meta-sponsored
Role
Team of 5 · gameplay & MR
Stack
UnityMixed RealityMeta Quest

Orbit Brawl

Mixed-reality space brawler that turns your living room into a battleground. Built in 48 hours. 1st place gaming, XR Hack Cologne.

Two days, five people, one Meta Quest. Orbit Brawl uses mixed reality to convert real-world geometry into a zero-gravity arena - your couch becomes cover, your ceiling a hazard. We took first place in the gaming category, sponsored by Meta.

1st place - gaming category, XR Hack Cologne
Real-room geometry as gameplay surface
Built end-to-end in 48 hours
[R-006]
Research
Year
2025
Role
Solo project
Stack
UnityC#Custom NNEvolutionary algorithms

Neural Evolution

Self-driving cars in Unity, evolved without supervision. Custom neural net + population-based training, configurable from the inspector. Generalises to unseen tracks under pure evolutionary pressure.

A playground for unsupervised driving behavior. Cars get a custom neural network and a fitness function - evolution does the rest. Architecture, mutation rate, population size, track topology are all configurable; you watch in real time as a generation gets less terrible at not crashing.

Configurable network architecture (no fixed topology)
Population-based training with real-time mutation
Diverse track environments, live evaluation
Generalises to unseen tracks at full speed under pure evolution
[S-007]
Shipped
Year
2021 - 2023
Org
MobX GmbH · Bonn
Role
Game developer
Stack
UnityC#VR

Everslaught Invasion

Fast-paced VR action title shipped on Meta Quest 2 & 3. Years on the engine side reworking and optimizing core systems.

I worked on Everslaught Invasion as part of MobX's team - most of my time went into reworking and optimizing already-shipped systems and gameplay mechanics. The kind of work that doesn't make trailers but is the difference between a VR game that ships and one that gets returned.

Shipped title on Meta Quest 2 & 3
System rework, optimization, QA
Real-frame-budget discipline that I still apply to ML
[S-008]
Prototype
Year
2022 - 2023
Org
MobX GmbH · Bonn
Role
Lead developer (prototype)
Stack
UnityC#VR

Sci-Fi Stealth Shooter (VR)

Solo-led VR stealth shooter prototype - concept to playable MVP, including enemy AI, stealth systems and player interactions.

A full prototype I drove from concept to MVP - designing the gameplay loop, building the enemy AI, the stealth state machine and the player-interaction layer. Internal, not public. The work that taught me how to own a system end-to-end.

Concept → playable MVP, solo-led
Enemy AI, stealth systems, player interactions
End-to-end system ownership
[P-009]
Product
Year
2024
Role
Freelance · AI flow design
Stack
LLMsConversational AIEval loops

LIYA

Redesigned an AI-driven onboarding & interview flow - unbiased candidate profiling, auto-generated company-side dossiers. myliya.com.

Freelance redesign of the AI-driven onboarding and interview process on the LIYA platform - see myliya.com. The new flow gives candidates a fully customizable, conversational way to showcase skills and mindset to an unbiased AI, and on the company side, generates tailored profiles that act as the first point of contact.

Customizable, personalized interview flow
Bias-aware AI profiling on the candidate side
Auto-generated dossiers for hiring teams
[T-010]
Tooling
Year
2026
Role
Author
Stack
PythonAgent frameworks

conv-proxy

A conversational proxy with minimal intelligence, designed to sit on top of agent frameworks like Openclaw.

A thin proxy layer for agentic systems - minimal intelligence on purpose. Sits between the user and the heavier framework, mediating the conversation without becoming the bottleneck. Built to be cheap, predictable and inspectable.

Designed for agent-framework composition
Minimal-intelligence-by-design (no sneaky prompt magic)
Inspectable conversation surface