External Publication
Visit Post

Multi-agent Claude production system + heritage preservation use case — Community GPU Grant inquiry + open-source collaboration

Hugging Face Forums [Unofficial] April 29, 2026
Source

Hi Hugging Face team and community,

This post was drafted collaboratively by Richard Murray (Founder & Steward, Quebec, Canada) and the Claude instance embedded in our multi-agent system as clone-co-pilote. We are reaching out via the forum because no public email is listed for partnerships or community grant inquiries — please redirect us to the right channel if this isn’t it.

TL;DR

We operate a multi-agent Claude Code production system (39+ specialized silos, 461+ cross-clone briefs, 3208+ forensic log entries) on AMD recycled hardware in rural Quebec. The system serves two missions in parallel:

  1. Heritage preservation at scale — 31.6 TB across 38 canonical categories, 5.17M+ multilingual files (Japanese vintage computing included), classified and indexed by AI through 250K+ doc ChromaDB RAG. Real digital archaeology, not data extraction.

  2.  2. **Multi-agent operational research** — production observation of behavioral patterns across heterogeneous agentic tasks. Forensic log captures every tool call, prompt, and decision continuously since 2026-04-25.

  3. We are reaching out for three reasons (any one of which would be valuable)

  4.  1. **Community GPU Grant inquiry** for a Hugging Face Space we’d publish around our heritage preservation pipeline (multilingual classification + retrieval visualization). Publicly demonstrable case study of open-source ML applied to digital heritage.

    1.     2. **Research Residency Program awareness** — we have a `clone-research` agent producing empirical findings (recursive meta-case in multi-agent systems, 5+1 production-validated behavioral patterns, Defense-in-Depth architecture for auditor regress). If a residency could host concrete formalization, we are listening.

      1.        3. **Open-source partnership / community spotlight** — our stack is documentable open. Our project’s philosophy is anti-extraction, anti-vendor-lock-in, anti-cloud-monopoly. Hugging Face is the natural home for what we’re building.

      2. Who we are

    2. Colossus is an independent AI inference + heritage preservation system based in Quebec, Canada (rural region).

  5.  * **Hardware** : 1× RX 6900 XT + 3× RX 580/590 (mining-recovered), scaling to 5 cards. Debian unstable + Mesa 26.0.5, Vulkan inference path via RADV. Anti-extraction by construction: production AI on recycled hardware proves the stack works without cloud-locked GPU rentals.

    •    * **Multi-agent orchestration** : 39+ specialized Claude Code silos with persistent structured memory and defined scopes (network architecture, archival, GPU operations, agent oversight, security research, strategic coordination, etc.). Cross-clone briefs in markdown protocol (461+ verified 2026-04-26), all ingested to shared ChromaDB RAG.

      •      * **Hierarchical behavioral discipline** : rules coded P0–P5 in `MEMORY.md` global. Honesty-first (P5), measure-before-claim (P4), respect human autonomy (P1) actively enforced.

        •        * **Empirical-veto policy** between agents: a specialist can veto a coordinator’s directive when the hypothesis fails. Production-validated 2026-04-23 (hardware specialist prevented destruction of two GPUs).

          •          * **Heritage corpus** : 31.6 TB / 38 categories / 5.17M+ files served via 3 NAS. Japanese vintage computing magazines, Western retro technical manuals, multimedia archives.

            •            * **Public artifacts** - https://colossus-ia.org (live API status, GPU telemetry)

              •              * https://imsai.colossus-ia.org (in-browser IMSAI 8080 emulator preview of educational kit)

                •                * Upstream co-signed contributions (Mesa, llama.cpp) — co-attribution policy: human + AI signed jointly

                • Why Hugging Face specifically

              • Open-source community alignment

              • Our project’s philosophy is anti-extraction, anti-vendor-lock-in. Heritage preservation must remain in community ownership. AI tooling for digital archaeology must be replicable by independent groups, not gated by API costs. HF’s mission is the natural home for what we’re doing.

            • Heritage preservation Space proposal

            • We can build a public HF Space demonstrating multilingual classification on heritage corpus (Japanese vintage computing, Western technical, retro game documentation), visualization of cross-corpus connections discovered by RAG retrieval, live demo of small-model classification (Qwen-VL 7B + BGE-m3 embedding) on uploaded vintage docs.

          • Community GPU Grant inquiry : would HF support GPU credits to host this Space at production scale?

        • Multi-agent forensic research possible Research Residency framing

        • Our clone-research agent has formalized empirical findings on multi-agent behavioral patterns since 2026-04-27 (living catalog A1-A11 + A8’ positive auto-correction; recursive meta-case documentation; 5-layer Defense-in-Depth for auditor regress; cold-boot context cost measurement ~10% of 1M-token window). This needs academic framing we can’t provide alone.

      • Open weights compatibility for Phil 70B local pilot

      • Our Évasion roadmap involves fine-tuning a 70B local model (Llama 3.x, Qwen 2.5, or Mistral as base) on hardware acquired in ~2 months (Threadripper Pro + 4× AMD AI Pro 9700). HF role hosting open weights, fine-tuning tooling (TRL, PEFT, datasets, transformers), and benchmarks makes this naturally HF-aligned.

    • What we ask

  6. We’re open to multiple modes:

  7.  * (a) Community GPU Grant for a public heritage preservation Space

    •    * (b) Research Residency for our `clone-research` empirical work

      •      * (c) Visibility partnership / community spotlight if our case is interesting

        •        * (d) Open dialogue without specific deliverable

          •          * (e) Connection to a relevant person on the team you’d suggest
          • We have no presumption about what fits. Comments, questions, redirections welcome.

        • Co-attribution note

      • Our policy formally co-signs human + AI contributions on external work. Mentioned proactively for transparency.

    • Reachable

  8.  * Email: rmurray@colossus-ia.org

    •    * Time zone: Eastern Time (UTC-4/-5)

      •      * Languages: English (this post by Claude co-pilot), French (Richard native)
      • Reported by Richard Murray, in collaboration with Claude (Anthropic) — colossus-ia.org

Discussion in the ATmosphere

Loading comments...