Senior Solution Engineer · Apps & Agents
Microsoft Power Platform · Copilot Studio · Dataverse · PCF
I build agents and applications on the Microsoft Power Platform, and the automation and governance that keeps them running once they are real.
This profile is a pattern library. Every repository is a self-contained sample solution that demonstrates a reusable approach to a common problem, published so others can adapt it.
Each repo opens with the problem it solves, and states plainly what has been verified against a live environment versus what has been assumed. Documented negative findings are included, because they save the next person days.
These are samples, not products. Provided as is, not production ready, and not affiliated with or endorsed by Microsoft. Review, test and harden before any real use.
|
Problem: the Copilot Studio VS Code extension is powered by the standard harness. Agents on the GitHub Copilot harness get no clone-edit-sync loop, so the harness built for the most capable agents has the least local tooling. Solves: a VS Code plugin whose skills scaffold, package, review and deploy those agents by
driving |
Problem: most multi-agent demos are one agent wearing several hats. Teams cannot tell when orchestration earns its complexity, so they over-engineer four agents that route badly. Solves: an orchestrator with four domain agents and six deliberately cross-cutting skills, plus the decision framework for when not to split. |
|
Problem: extraction is rebuilt every project because the schema is hardcoded, scanned documents silently return nothing, and AI writes to the system of record unchecked. Solves: user-defined capture fields, automatic scan detection with a vision fallback, and a human validation gate before anything persists. |
Problem: safety tooling is retrospective. It records incidents well and prevents them poorly. The procedure that matters sits in a library nobody opens at the job site. Solves: a JSA sized for the two minutes a worker has, hazard identification from a photo, and controls matched to the hazard rather than the whole library. |
|
Problem: "how do I put my agent inside my app?" has no good published answer. Users get bounced to a separate surface, and the SSO path still prompts them to sign in again. Solves: a PCF component embedding a Copilot Studio agent in a Power App, with the undocumented OAuth card interception that makes SSO silent, plus an offline demo mode. |
Problem: the Planner connector does not work with Planner Premium. Premium plans live in
Dataverse as Solves: the full Planner-to-Dataverse mapping from live metadata, and a working solution with a dedupe guard, testable without a Premium licence. |
|
Problem: organisations deploy agents with no way to answer "what will this cost at scale, and which agent is driving it?" Cost becomes visible when capacity runs out. Solves: collects consumption from the licensing endpoints the admin centre itself uses, verified against a live tenant, with burn rate and days-to-exhaustion forecasting. |
Patterns are extracted and published as they generalise cleanly. Contributions and corrections are welcome, particularly verification results from your own tenant. Several findings are marked unverified precisely so someone can close them out. |
| Principle | In practice |
|---|---|
| Self-contained samples | Each repo stands alone as a demonstrable pattern. No customer IP, data, branding or configuration. |
| Explicit about certainty | Repos separate what was verified live from what was assumed, and publish negative findings. |
| Built for handover | Configuration over customisation. Environment variables, not hardcoded IDs. Solutions ship disabled. |
| Documented for the reader who is stuck | Known gotchas, licensing traps and failure modes, not just the happy path. |
Every repository states what it solves, what is verified, and what is not.
Sample code. Provided as is, without warranty. Not production ready. Not an official Microsoft product, and not affiliated with or endorsed by Microsoft.