Principal AI Architect
I'd own the seams, set the operating model, and keep the model plane honest — 6-spine review with Demo vs Strict labeled, not a catalog dump.
Panel receipt
Demo vs Strict labeled. Free cold starts are not warm SLOs — probe these artifacts before you book a loop.
Decision
Hit live /health before the call so cold-start latency is labeled, not surprise.
Signal
Portfolio probes the D1 spine APIs server-side and surfaces warm vs cold honestly.
Limitation
Render Free idle spin-down is real — first hit after idle can take 15–40s.
Decision
Panel default is Strict-local packs — not the public Free demo as an always-on SLO.
Signal
AegisAI and Enterprise RAG ship STRICT_PANEL_PACK docs for local / JWT panel paths.
Limitation
Public Free demos stay honesty-labeled; Strict may require local or GCP setup.
Decision
Replay ask→RAG→govern→meter without a guided tour when you want a stranger-checkable receipt.
Signal
GOLDEN_PATH.md + latest JSON artifact in the architecture portfolio.
Limitation
Full ask→answer needs API keys; ACF live publish still requires Clerk.
Decision
Say posture statuses aloud — both vLLM and PEFT receipts are now real Mistral-7B/L4 runs, not fixtures. No fiction either way.
Signal
ModelForge /api/v1/posture ready + vllm_cuda.json real (Mistral-7B on L4, cuda=true) + peft_gpu.json real (QLoRA SFT + DPO on Mistral-7B/L4, cuda=true).
Limitation
vLLM receipt is real but single-request, not a throughput-under-load claim; PEFT receipt reports real training config/timing, not a quality/win-rate score (see the receipt's known_gaps) — ADR-035/037.
Decision
Every portfolio-wide claim should resolve to Verified, Demonstrated, or Contextual.
Signal
/proof marks soft career metrics as Contextual with explicit limitations.
Limitation
Contextual does not mean false — it means primary records are confidential or external.
Start here
Recruiters & Hiring Leaders
Forward the brief, skim the spine, then decide if a loop is worth it. Embed is how architecture lands — not a second career brand.
For hiring teamsExecs & org leaders
Decision rights, ARB invariants, and what I refuse to automate. Leadership is the other half of Principal+ — not a sidebar.
Open leadershipTechnical Reviewers
Diagrams, 18 catalog products, repos, and case studies on one page — start with the spine, not a scavenger hunt.
View portfolio6-spine review path
Start with AegisAI (governance), then orchestration, access-aware RAG, ModelForge (weights / serve / proof), and a governed publish path. Open the Model Plane taxonomy glassbox for LoRA · QLoRA · Multi-LoRA · task types · classical ML stack.
Monitor → Govern → Remediate — a runtime control plane for tool authorization, policy violations, HITL approvals, signed audit, and agent registry lifecycle.
Reviewer evidence
Control plane over agent builder — enterprises need governance across fleets, not another chat UI.
Website deploy tools forced through approval_required policy on live platform.
Registry defaults to in-memory on free tier; Postgres path documented for production.
Supporting · Multi-agent OS
LangGraph stateful graphs over linear chains — enterprise workflows need checkpoints and HITL.
Case studySupporting · Knowledge layer
Authorization before ranking — vector DB is implementation; access control is architecture.
Case studySupporting · Model Plane
One ModelForge flagship beats burying PEFT/vLLM in a teaching drawer (ADR-034).
Case studySupporting · Content automation
AegisAI gateway blocks publish until policy allows — side effects never bypass governance.
Case studySpine and AgentOps consumers wire critical suites as merge gates: RAG answer + adversarial (Enterprise RAG), harness QA + repo fix (LoopForge), mission gate + collaboration scorecard (AegisLoop), graph HITL (Content Factory), brief gate (Sentinel), triage preference (DomainForge), and router_invariant across VAP, AegisAI, and FinOps. Builds fail on regression — not just fixture validation.
Get in touch
Hiring teams start on the brief. Panels start on the 15-minute review. Subscribe for weekly architecture essays.