7
GFIS stages
Systems I build · AI infra
I don’t orchestrate — I build.
Credibility rests on owned infrastructure: a research system, a corpus, hardware and agent platforms — not third-party black boxes.
3,885
Automated tests
6.78M
Corpus chunks
2× GB10
GX10 nodes
235B
Local model
Qdrant
Self-hosted
The stack
Five layers, in my own hands.
Research system
GFIS
A proprietary 7-stage research and synthesis system (L1 fact / L2 dynamics / L3 hidden mechanism). ~3,885 automated tests, per-client dimension configs.
Corpus & RAG
Corpus V3
13.8K books, 6.78M leaf chunks, 31 shards. Own embedder (Qwen3-Emb) + reranker (BGE-v2-m3), self-hosted Qdrant. Public figure: a 1.48M-fragment book corpus.
Hardware
GX10 lab
Two ASUS Ascent GX10 (GB10) nodes over Tailscale; vLLM inference (122B/235B models), Ollama routing; a Hetzner RAG node.
Agents
Agent platforms
Hermes (primary), AGENTICWEB, plus the AGÓRA (multi-model) and DIALEKTIKA (adversarial) reasoning engines; IDEATEST business-idea validator.
PKM
LLM-wiki
A layered personal knowledge OS on top of Obsidian: markdown → Qdrant → graph → agent query.