Technical Review
Fifteen minutes. Six systems. Decision first.
Fifteen minutes. Six systems. Decision first. Public Free demos are honesty-labeled — not warm SLOs. Prefer AegisAI and Enterprise RAG Strict packs when you care about fail-closed policy. Say both PEFT and vLLM CUDA receipts are real (Mistral-7B on a rented L4) — PEFT reports training config/timing, not an adapter quality score; do not overclaim always-on serve. Spine-health probes honesty, not an SLO gate. Cloud proof is ephemeral AWS/GCP receipts, not always-on EKS.
- 13 min
Governance — AegisAI
I'd put an independent gateway in front of irreversible tools — policy, HITL on the scary ones, signed audit. The agent graph doesn't get a back door. UI labels Demo vs Strict from PRODUCTION_STRICT; panel default is the Strict-local pack.
- 23 min
Orchestration — Venkat AI Platform
Multi-agent OS for what agents should do — specialists, RAG strategies, gateway-wrapped notify. Demo strip admits fail-open notify by default; Strict is the panel path when you care.
- 33 min
Knowledge — Enterprise RAG
Filter by who the caller is before you rank. Optimizing recall with unauthorized neighbors is how demos look smart and prod leaks. Dual path: Demo (body Principal) + Strict JWT (ADR-0006/0009).
- 43 min
Models — ModelForge
Agents decide what to do; the model plane decides which weights, where they run, and how we prove it. Live posture: PEFT + CUDA vLLM + SLM + gateway ready — both receipts real (Mistral-7B on a rented L4: QLoRA SFT+DPO training, 13.74 tok/s serving). Training config/timing is not an adapter win-rate. Taxonomy glassbox is for the grill (LoRA · QLoRA · Multi-LoRA · task types).
- 52 min
Application — AI Content Factory
Side effects stay behind gates. Research → drafts → HITL → publish. Banner shows Demo vs Strict; publish fails closed when Strict. Golden path uses /health only — Clerk for live publish.
- 62 min
Decisions — ADR portfolio + eval proof
Why these seams exist: 37 ADRs (orchestration vs governance, RAG auth, ModelForge ADR-034, PRODUCTION_STRICT), golden-eval CI gates, and a replayable ask→RAG→govern→meter artifact.
Convinced the spine holds?
Full catalog is 13 platforms, 5 teaching patterns, and 17 repos — secondary to the five above.
View full portfolio