Six architectures you can click through.
Not slideware. Each one is an interactive walkthrough — press play or click any node to see its responsibilities, the tech and data behind it, the controls that constrain it, the non-functional targets it has to hit, and what it costs to run. Two are drawn from production work; four are reference designs for problems I think are worth solving. Every one of them ships with a carousel deck and a static diagram if you would rather read than click.
Agentic AI in Enterprise ERP. Read freely, act carefully.
The asymmetry that makes ERP agents tractable: reading the system of record is cheap and reversible, writing to it is neither. Policy RAG over the procurement rulebook, agent authority capped at the requesting user’s, idempotency keys on every write-back, and a human approval gate above a value threshold — with a SOX-auditable trail recording who acted, why, and against which model and prompt version.
Open the walkthrough →HealthQCopilot. Narrow agents and a verification gate you can audit.
The clinical copilot behind the venture, drawn end to end. Consent checked before any model call, PHI redacted before a prompt leaves the boundary, per-agent tool allow-lists, and a hallucination guard that verifies every clinical claim against FHIR before a clinician ever sees it. Every decision streams to an audit log — because in this domain “the model said so” is not a defensible answer.
Open the walkthrough →Trustworthy Education AI. Grounded, safe, and teacher-led.
How I would architect a tutor that a district could actually approve: retrieval grounded in the school’s own curriculum rather than the open web, student PII redacted before it reaches a model, age-appropriate content safety on every response, and a teacher in the loop on anything that reaches a learner. Pedagogy is a design constraint here, not a prompt.
Open the walkthrough →Teacher Workflow. The same platform, pointed at the adult in the room.
Most education-AI demos target the student. This one targets the teacher’s day: planning, authoring, grading, class insights, differentiation, and family communication, each as a narrow agent over the same grounded retrieval layer. Outputs are fairness-checked and nothing reaches a student or a family without the teacher approving it first.
Open the walkthrough →Local Agentic AI for Healthcare. Offline, on a ~$300 box.
Rural clinics and field medicine do not have a reliable link to a hyperscaler, and the patients who need decision support most are the ones least likely to be sitting near one. A quantized open model on a mini-PC, cache-augmented generation over the guideline corpus, on-device FHIR and vector retrieval, hot-swappable LoRA adapters per specialty, and cite-or-refuse when the evidence is thin. Zero PHI egress — not as a policy, as a topology.
Open the walkthrough →Local Agentic AI for Education. A tutor that works when the internet doesn’t.
The same offline pattern under classroom constraints. Cache-augmented generation over the local curriculum, RAG for the long tail, a quantized model small enough to run on hardware a school can actually afford, and no per-token cost to ration. Student data never leaves the room, which makes the compliance conversation a short one.
Open the walkthrough →Each walkthrough has a “Desktop 16:9” and a “Vertical 9:16” layout — the vertical one exists so the animation can be screen-recorded for social without reflowing. The four reference designs are exactly that: designs, not shipped products. Where a diagram describes work done under an employer, it is the generalised pattern rather than their deployment.