Work

Systems with evidence.

Three flagship builds across local AI, generative audio, and real-time game systems. The result comes first; implementation depth stays available without overwhelming the initial scan.

Featured work

Start with the system closest to your interests, then open its implementation detail or full case study.

P-01 / Local AI Ramone Private intelligence infrastructure with local inference, retrieval, and bounded public edges. P-03 / Generative audio SONIN An autonomous audio-visual instrument built around dual engines and an analysis feedback loop. P-02 / Game audio SlamPunk A real-time dynamic mix engine that responds to competitive play without masking gameplay cues.
All projects

The full build record.

Media, role, result, and primary links stay visible. Specs, tools, and supporting outcomes open on demand.

P-01
AI / Local LLMs RAG DevOps

Ramone — Local AI System

Architect & Systems Engineer

A fully self-hosted private AI infrastructure — zero cloud dependency, zero data egress. Five LLMs served via Ollama on local hardware, wrapped in Docker, accessed through Open WebUI with ten specialised workbots backed by RAG knowledge bases. Everything lives on a dedicated NVMe drive: portable, rebuildable in under 30 minutes.

Result

Five local models and ten specialised workbots run without cloud inference or data egress, and the stack remains portable enough to rebuild in under 30 minutes.

Technical detail
GPU
RTX 5070 — 12GB VRAM
RAM
64GB DDR5
Models
5 LLMs / 3b–32b params
Workbots
10 specialised agents
Ollama Docker WSL2 Ubuntu Open WebUI RAG Pipelines Windows Batch
Full data sovereignty — conversations, documents, and RAG knowledge bases never leave local hardware
Custom ATLAS_BOOTSTRAP.bat auto-starts full stack on boot — Docker, Ollama, Open WebUI with live health checks
RAG knowledge bases built from chunked university lecture PDFs — Academic Vector Index grounded in real course material
Hardware portable — entire system on L:\ NVMe, rebuildable on new hardware in under 30 minutes
Model selection matched to task profile: 3b for speed, 32b for deep reasoning and RAG retrieval
Case Study GitHub
Intensity:
P-02
Audio Systems Game Dev Unreal Engine 5

SlamPunk — Dynamic Mix Engine

Lead Audio Systems Designer & Music Producer

A logic-driven dynamic audio engine for SlamPunk — a competitive future-sport game blending tag and basketball. Engineered a 15-stem interactive music system that mirrors player performance in real time, escalating from a minimal atmospheric floor to a dense triumphant climax while maintaining surgical clarity for competitive gameplay cues.

Result

A 15-stem music system moves across three intensity layers on a locked 140 BPM grid while preserving gameplay cues through sidechain ducking and spectral carving.

Technical detail
Timeline
14 weeks / ~600hrs
Grade
100% — Grade A
Architecture
15-stem / 3 intensity layers
BPM Grid
140 BPM locked
Unreal Engine 5 MetaSounds Blueprints FL Studio 24 REAPER v7 Vital (Wavetable)
15-stem dynamic music system split across 3 match intensities and 5 discrete instrument busses
Real-time sidechain ducking — music stems duck dynamically when parkour SFX or commentator lines trigger
Hierarchical submix tree with spectral carving — Data Bus hard high-passed at 200Hz to preserve low-end headroom
Global Music Manager Blueprint locks all transitions to 140 BPM grid — zero rhythmic stutter on intensity escalation
Fully modular pipeline — new arenas and themes hot-swap into established mix framework without breaking thresholds
Case Study
Walkthrough: Watch on YouTube
P-03
Generative Audio Max/MSP DSP

SONIN — Generative Audio-Visual System

Designer & DSP Engineer

An autonomous digital instrument built in Max/MSP that composes its own evolving music and visuals in real time by continuously listening and reacting to the sounds it creates. A dual-engine architecture (melodic and granular) connected by an internal audio feedback loop; the system analyses its own output and routes that data back into its own generative parameters.

Result

Two generative engines remain musically stable through an internal analysis feedback loop while GPU-rendered visuals leave the audio thread unaffected.

Technical detail
Timeline
Sept – Dec 2025
Grade
100%
Architecture
Dual-engine + feedback loop
Visual
GPU pipeline via Jitter
Max/MSP Jitter poly~ pfft~ pattrstorage jit.gl.slab
Scale-constrained melodic engine using coll lookup tables — infinite variation that never leaves the key
Granular engine with dynamic micro-fade voice management — runs at maximum density without DSP artifacts
Internal RMS and spectral brightness feedback loop creates genuine cause-and-effect between engines
Non-linear logarithmic scaling prevents feedback runaway — system settles into stable evolving states
Full GPU visual pipeline via jit.gl.* — pixel math offloaded to graphics card, audio thread unaffected
Case Study Walkthrough