Active Builds
Four systems I'm building and operating on my own time — governed AI infrastructure, compliance-grade pipelines, and encrypted knowledge architecture. The problems they solve are the same ones I see inside enterprise environments.
A 550+ document personal vault built to enterprise content management standards. ISO 27001-aligned data lifecycle policy — nine review classes enforced at every git commit via pre-commit hooks that block any active file missing classification. Multi-lane routing: legal and sensitive content stays strictly local; cloud-eligible content routes separately. Version-controlled with a formal dev → test → prod branch promotion protocol. The vault is the authoritative source; everything downstream — AI-generated enrichment, vector indexes, exports — is derived and read-only.
Same governance architecture I'd build for any organization that needs to trust what it knows.
NVIDIA Quadro P5000 (CUDA) and AMD Radeon Pro WX 9100 (Vulkan) — two independent inference lanes on a single workstation running a structured five-seat model council. Each seat is validated through 4-hour soak testing before production admission: 2,800+ inference cycles, zero failures, VRAM telemetry logged throughout. Seat types span dense models (7B–30B parameters), MoE architectures, and a vision seat for image annotation and document OCR. An 80% VRAM ceiling is the only admission rule — no exceptions, no override.
I know the difference between what documentation claims and what a model delivers under sustained load — because I ran the test.
vault_enrich.py converts binary vault files — PDF, DOCX, XLSX, images — into structured knowledge records: governed YAML frontmatter with semantic fields (who / where / when / why / key_facts / semantic_tags) plus an immutable source snapshot. First production batch: 48 enrichment outputs and 4 vision-annotated context records, all human-reviewed. Sensitive documents are quarantined before entering the pipeline and routed to the compliance handler. Idempotent by design — re-enrichment runs against the frozen source extract, not the original file. A degraded model pass cannot corrupt a clean record. Downstream: Qdrant vector indexing for semantic search across the full knowledge base.
Same pattern as any enterprise enrichment pipeline — the difference is I own the failure modes.
GPRF is the master orchestrator for all document ingestion requiring sensitivity controls. Every file is classified before any AI council call — legal, restricted, and privileged content never touches a cloud LLM. Pre-determination policy gate: classification happens at entry, not after. Built to exceed LKIF (Legal Knowledge Interchange Format) and SALI LMSS standards: immutable chain-of-custody audit trail, privilege flags, jurisdiction tracking, and OQ-3 confidence calibration gates on every council output. Tokenization replaces PII and identifying information before inference; rehydration restores it after. Encrypted LUKS storage at rest. The full pipeline — tokenize, dispatch, rehydrate, audit — is sequenced and auditable end to end.
This is the compliance architecture I'd propose for any organization handling regulated or legally sensitive content at scale.
These aren't side projects. They're a working proof of how I approach complex, sensitive, multi-system problems — with the same rigor I'd apply inside any enterprise: governance before feature, compliance before speed, test it under real conditions before you trust it.
I'm building infrastructure that most organizations would staff a team to design. I'm building it alone — which means I understand every layer, every failure mode, and every tradeoff.