The work, not the summary of it.
Architecture is easy to claim and hard to show, so most of what is here is the thing itself rather than a description of it — diagrams you can click into, node by node, with the controls and the cost model attached; one design carried all the way through the TOGAF method; and writing about what actually broke. Two of the architectures come from production systems. The rest are reference designs for problems I think are worth solving, and they say so.
Seven you can click through.
Not slideware. Press play, or open any node for its responsibilities, the tech and data behind it, the controls that constrain it, and what it costs to run.
Schema-to-Screen
A GraphQL contract and the React screen on top of it, generated over thousands of tables. The model sees a retrieved subgraph, never the catalog; entitlements compile into the SQL; writes are proposals until a human approves them.
Open the walkthrough →Agentic AI in Enterprise ERP
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 user’s, and a SOX-auditable trail on every write-back.
Open the walkthrough →HealthQCopilot
The clinical copilot behind the venture, drawn end to end. Consent checked before any model call, PHI redacted at the boundary, and every clinical claim verified against FHIR before a clinician sees it.
Open the walkthrough →Trustworthy Education AI
A tutor a district could actually approve: retrieval grounded in the school’s own curriculum, student PII redacted before it reaches a model, and a teacher in the loop on anything that reaches a learner.
Open the walkthrough →Teacher Workflow
Most education-AI demos target the student. This one targets the teacher’s day — planning, grading, differentiation, family communication — each a narrow agent over one grounded retrieval layer.
Open the walkthrough →Local Agentic AI for Healthcare
A quantized open model on a mini-PC, cache-augmented generation over the guideline corpus, on-device retrieval, and cite-or-refuse when the evidence is thin. Zero PHI egress — as a topology, not a policy.
Open the walkthrough →Local Agentic AI for Education
The same offline pattern under classroom constraints: local curriculum, a model small enough to run on hardware a school can afford, and no per-token cost to ration. Student data never leaves the room.
Open the walkthrough →One design, carried through the whole ADM.
Schema-to-Screen is the one architecture here with its complete TOGAF set published beside it — eight views from vision to roadmap, each with its own catalogs, matrices and diagrams. It is what the method looks like when it is load-bearing rather than decorative.
Navigator
The ADM wheel, clickable, with the diagram inventory and reading conventions.
Open →Phase A · Architecture Vision
Vision statement, scope, solution concept, stakeholder map, driver → outcome chain, business scenario.
Open →Phase B · Business Architecture
Heat-mapped capability map, value stream, swimlane process flow, RACI matrix, business services catalog.
Open →Phase C · Data Architecture
Conceptual model of 20 entities, data lifecycle and dissemination, data security matrix, CRUD matrix.
Open →Phase C · Application Architecture
Component and communication diagram, interaction matrix, generation-run and runtime request sequences.
Open →Phase D · Technology Architecture
Platform decomposition, environments and locations, Azure deployment with a local-first equivalent, standards catalog.
Open →Phases E & F · Roadmap
Gap analysis, 17 work packages, transition architectures TA-0 to TA-4, milestone plan M0 → M6 with exit criteria.
Open →Preliminary · Requirements · G · H
Ten principles (six absolute), requirements catalog with traceability, risk register, governance checkpoints.
Open →Field notes from the platform.
What broke, what it cost, and what I would do differently. All writing →
TOGAF, applied for real
When the method is load-bearing rather than decoration — and what it costs when it is treated as paperwork.
Read →The arithmetic decides
Where the model may not go in a money-moving agent system, and why the boundary is arithmetic rather than judgement.
Read →Making enterprise RAG a platform, not a project
What broke when every team built their own retrieval stack, and the contracts that replaced them.
Read →Building agentic AI for healthcare
Lessons from HealthQCopilot: narrow agents, verification gates, and why “the model said so” is not a defensible answer.
Read →Shipping a micro-frontend shell on Module Federation
The contracts that kept a multi-team frontend sane, and the ones that were not worth the cost.
Read →Collapsing a fragmented REST surface into one GraphQL gateway
What broke first, and the migration order that would have hurt less.
Read →Things I am building.
HealthQCopilot
A clinical copilot built on narrow agents over FHIR, with consent, PHI redaction and a verification gate in front of every clinical claim. The product tour runs in the browser; the architecture is published in full and the code is public.
Take the product tour →Bangla LLM
Language models for a language with far less training data than its speaker count would suggest — and the evaluation problem that comes with building for a language the benchmarks ignore.
More on the home page →Peer-reviewed, and a while ago.
Computer vision and neural networks for industrial defect detection — the problem that got me into this work in the first place.
- Automated System to Detect Textile Defects ICCIT 2006
- A Suitable Neural Network to Detect Textile Defects ICONIP 2006, Hong Kong
- Automated Textile Defect Recognition System using Computer Vision & ANN Enformatika 2006, Budapest
Send me the one workflow you wish already worked.
If any of the above looks like the shape of your problem, the fastest way in is a concrete one.