Hi, I'mJeff McGillis Heiden
I engineer high-performance distributed systems, multi-agent orchestrations, and scalable AI infrastructure with a focus on edge and local-first, offline-first architectures.
About Me
I'm a Systems Architect and independent AI researcher. I recently published a new theoretical framework for LLM quantization and quickly followed up with a full proof-of-concept implementation, complete with working code and reproducible benchmarks — all on consumer-grade hardware with zero funding or sponsorship.
My engineering focus is on building end-to-end systems that pair rigorous architecture with advanced AI — from secure MVVM applications and deterministic simulations to multi-agent orchestrators and Spatial AI engines. Before tech, I was accepted into the Swedish Air Force officer program for intelligence and security, which shaped how I approach reliability under pressure, sustained deep-focus work, and building things that hold up when nothing else does.
Outside of work I explore new technologies and build advanced 3D printing engineering projects.
Published Research
Multi-trit quantization for large language models: a new research program for LLM compression and trit-native architectures.
Toward Multi-Trit Quantization for Large Language Models
Theoretical Framework for Balanced N-Trit Weights, Trit-Plane Generalization, and Mixed-Layer Precision
Practical Multi-Trit Quantization for Large Language Models
Full-Text-Path Trit5 Implementation and WikiText-2 Proof of Concept on Qwen3.5
Mixed-Layer Trit-Depth Allocation
Heterogeneous radix-depth assignment and non-uniform tritN quantizer design
Research Roadmap
Toward Multi-Trit Quantization: Theoretical Framework
↑ View PaperDefinition of balanced N-trit expansions, significance-ordered trit-plane decompositions, and mixed-layer trit-depth allocation.
Practical Trit5 Implementation & WikiText-2 PoC
↑ View PaperEmpirical validation that a structurally faithful blanket Trit5 implementation occupies a meaningful accuracy regime relative to blanket Q8.
Mixed-Layer Trit-Depth Allocation
Heterogeneous radix-depth assignment and non-uniform tritN quantizer design.
Optimized Multi-Trit Runtime (C++/Rust)
Development of a highly optimized C++ or Rust inference engine with native support for unpacked multi-trit matrix multiplications.
Trit-Native Architecture Research
Investigating whether transformers can be designed or trained directly around multi-trit representational assumptions rather than inheriting them only through post-training quantization.
More to come...
Additional research directions will be disclosed as the theoretical framework matures into hardware-aware systems.
Featured Projects
On-Device AI Education Platform
An offline-first, K-12 educational platform that harnesses behavioral gamification and on-device AI to drive academic mastery.
Ugdrasil:ᚢᚷᛞᚱᚨᛋᛁᛚ
An infrastructure-agnostic, multi-agent reasoning orchestrator that couples cloud intelligence with secure, offline-capable local execution.
Project Aion (Universal TTRPG)
Rust-native, multimodal AI RPG platform with conscious NPCs and living world simulation.
ALS-ALM Platform
Fully self-hosted, offline-first, AI-native ALM/PLM operating system.
AI Audiobook Generator
Automated, high-fidelity audiobook generation pipeline using advanced TTS models.
Industripoolen.se
Industrial hiring platform with ERPNext integration and comprehensive workflow automation.
Fami.Central
Multilingual AI-powered family app with secure MVVM architecture and end-to-end encryption.
VR Sensor Glove
Sensor-enabled VR glove prototype for real-time motion tracking in VR applications.
PEEK 3D Printer Modification
3D printer modified for high-performance PEEK filament printing for industrial prototypes.
Graphing Calculator
TI-84-like calculator with custom GUI, ahead of its time with touchscreen vision.
Work Experience
Independent Researcher — Multi-Trit LLM Quantization
Published two DOI-registered papers establishing a new research program for LLM quantization based on balanced base-3 weight representations. Developed the theoretical framework and implemented a full-text-path blanket Trit5 quantizer with custom Triton kernels. T5 consistently outperformed blanket Q8 on perplexity across Qwen3.5 0.8B, 2B, and 4B models. Full reproducibility artifact package released under Apache 2.0.
PM / PO / Technical Lead
Sole architect of a full-scale industrial hiring platform built over 7 months. Researched and selected the entire tech stack (Bubble.io, Manatal ATS, NextERP), designed the system architecture from initial Figma wireframes through production deployment. Implemented search algorithms, intelligent scheduling, and suggestion engines. Built four distinct account tiers (Client, User, Admin, Tech) with role-based access control. Integrated enterprise ERP endpoints, automated complex cross-system workflows, and managed deployment infrastructure end-to-end.
Chief Architect & Founder
Designed and built a multilingual, AI-powered family application as a production-ready MVP. Architected a secure MVVM infrastructure with AES-256 device-only encryption—keys never leave the client, with a deterministic recovery phrase system inspired by crypto wallet UX and Wickr's forward-secrecy model. Implemented real-time key-change notifications (à la WhatsApp) to alert users of security state changes.
Military Training & Leadership
Accepted into Swedish Air Force Officers Training (Und/Säk) — Blekinge Flygflottilj, 2013. Did not complete. Developed operational discipline, threat-model thinking, and the ability to execute reliably in high-stakes, resource-constrained environments—principles that directly inform my approach to building resilient systems.
Computing Science (coursework, no degree completed)
Specialized in AI, systems design, and low-level programming. Research focus on algorithmic optimization, formal methods, and distributed computation—laying the theoretical foundation for production-grade systems architecture.
Get In Touch
Contact Information
+51 924 481 965
Current Status
