Marine diesel mechanic by trade. Building local-first AI infrastructure.
Multi-agent systems that run on your hardware, stay inspectable, and don't depend on the cloud.
"Identity loads whole. Knowledge loads chunked."
Currently: Minni.
Minni — Old Norse for memory; the minni was also the toast raised so the dead stayed remembered. It descends from PIE men- ("to mind"), the root behind mind, mnemonic, Greek mnēmē, Sanskrit manas — the proto-word for memory, put back to work. The rune is Mannaz (ᛗ), the mind rune, from the same family.
(formerly Sovereign Memory) — local-first memory and governance for AI agents, evolving in the open.
Per-agent vaults with review-first learning: writes are proposed, pass a quality gate and operator review, then become durable. Cross-agent handoffs carry identity, principal, and context. Every read and write is audited. Session hooks feed the same memory spine to Claude Code, Codex, Gemini, and Grok — different agents, one shared, inspectable history.
Shipping now: minni:plan — plans as durable artifacts, not session todo-lists. They survive crash and compaction, one agent's plan is executable by another (Claude writes, Codex runs, both observable), and steps close on an evidence gate: not "the agent said done," but the proof passed — test output, serial markers off real hardware.
SQLite/WAL is runtime truth · FAISS handles vector recall · native Apple Foundation Models keeps it on-device · no remote service.
| Project | What it does |
|---|---|
| 🔩 AetherKernel | Bare-metal Embedded Swift kernel for a Raspberry Pi 4 — async/await and a custom SerialExecutor on real hardware, no OS underneath. Started as a palate cleanser and became the live proving ground for Minni-coordinated agent fleets; surfaced swiftlang/swift#89835 upstream. |
| ⚙️ Binary-Forge | Hand-forged x86-64 Linux ELF binaries. No compiler. No libc. Just raw ELF, direct syscalls, and machine code elegance. |
| 🧬 llmHub | Shipped and resting. Native macOS/iOS platform for running agents side-by-side — @mention routing, 4 concurrent streams, 8 LLM providers behind one protocol, MCP bridge. Its memory layer became part of what Minni is now. |
A paired runtime + harness — Swift agent framework plus the Python eval framework that grades it.
| Project | What it does |
|---|---|
| 🪶 Syntra | Swift backend that implements a tri-brain orchestration pattern (Valon / Modi / Drift). Vapor REST API, pluggable LLM providers (OpenAI / Anthropic / Grok), built-in tool registry, context & token management. |
| 🧪 Syntra-Testing | Python benchmarking framework — GSM8K, ARC, CMT and custom datasets · grading · matplotlib viz · PDF reports. Used to evaluate Syntra and other agent systems. |
Datasets on 🤗 Hugging Face — eval traces, deliberation transcripts, and dilemma suites generated by these projects:
syntra-evals-dataset— 480 examples including a CMT-Benchmark subset for condensed matter physics reasoning.Syntra-Ethics-Dataset— 177 tri-brain dilemma prompts acrossvalon_ethics,modi_logic,drift_resilience, andcoherence_structuressuites.syntra-testing-evals-v4— full benchmark suite (CMT + ARC + GSM8K) with grading tools and sample runs.council-transcripts— multi-agent deliberation sessions across Workhorse / Creative / Speed personas, synthesized by a Conductor.
| Languages | Swift 6 TypeScript Python |
| Frameworks | SwiftUI Node |
| AI / Agents | Apple FoundationModels MCP JSON-RPC multi-agent CLIs (Claude Code · Codex · Gemini · Grok) |
| Memory & Retrieval | SQLite / FTS5 FAISS sentence-transformers |
The first thing I built was an assistant for diagnosing engines and walking me through mechanical work. Everything here grew out of that — the agent architectures, the local-first thesis, the focus on inspectable memory.
The mechanic's job hasn't changed. The toolbox just got new tools.
📍 Southern NJ · 🤗 huggingface.co/Infektyd · 𝕏 @infektyd