Portfolio
What I have built
50+ agents in production across 14 clients, a deliverable factory exposed as an MCP server, and the governance to keep them honest. One project per slide: what it is, who uses it, what it does, and one number that proves it.
- 50+
- agents in production
- 14
- active clients
- 20+
- sectors
Microsoft MVP, M365 Copilot and Copilot Studio · Lausanne
10 projects, one slide each. Click a diagram to jump to it.
International scientific research programmeClient anonymised
11 Copilot Studio agents over a multi-terabyte document estate
Engineers and business users ask one Teams entry point. An orchestrator routes each question to the specialist agent and the retrieval mode that fits it, so the user never has to know which one.
- Semantic search over technical documentation, Graph RAG over SharePoint, text-to-SQL over the data warehouse
- A 9-agent procurement chain from request to payment, with multi-step approvals and cost forecasting
- 2 to 4 hours saved per user per day, as reported by the client
Delivery tooling, built at WitivioOwn build
A deliverable factory exposed as an MCP server
Consultants and Copilot agents call the same 178 tools to compose client decks, contract documents and architecture diagrams. The delivery method is encoded as guardrails in the tool, so a junior ships senior work.
- 36,848 lines of TypeScript, 364 test assertions, all green in continuous integration
- Finalisation is blocked on a broken brand rule or an internal slide left in a client deck
- The local RAG refuses to run off loopback: a check in code, not a promise in a charter
Pre-sales, in production at WitivioOwn build
A chain of agents that qualifies AI use cases before anyone builds them
An employee describes an idea in Teams. 7 specialist agents dig into it, a separate qualifier scores it out of 25 and drafts the report in the requester's own mailbox. It has the authority to say no.
- 5 rejection filters no score can override, each with 2 concrete reframes
- The report is drafted from the requester's mailbox, never a service identity, and a human sends it
- Featured by Microsoft as System Prompt of the Week
Governance, public repositoryPublic
AgentLens: an agent that audits a tenant's agents
It reads Azure Resource Graph, Microsoft Graph, Dataverse and Cost Management, read-only, and answers inside Microsoft 365 Copilot: who owns what, which environments are open to everyone, and what the estate costs.
- 5 tools: inventory, DLP posture, value and cost, consolidation, estate map
- One read-only service principal, Bicep, Azure Container Apps, GitHub Actions and Vitest
- Zero and unknown are different answers: it never invents a number
Global energy group, 120+ countriesClient anonymised
An agent that runs tenders under EUR 50k end to end, inside Teams
A buyer describes the need, budget and scope in chat. The agent drafts the statement of work and pricing grid as real Word and Excel files, opens a Teams channel per tender as the audit trail, and routes the approval.
- One orchestrator and 5 agents specialised by purchasing segment: a single generalist answered badly
- 4 months from framing to V1 in production, France first, designed for 15,000+ employees
- Audited by Microsoft engineers and published by Microsoft as a customer story
International insurerClient anonymised
IT support rebuilt around one Teams app and one agent
Employees open tickets, browse the service catalogue and talk to the agent in a single Teams app. When direct access to the ticketing system was refused on security grounds, version two read through read-only endpoints over the data lake instead.
- 26 support providers consolidated behind one front door
- 9,740 SharePoint pages indexed into the unified knowledge base
- Retested on every Microsoft release, with a KPI dashboard and history in version two
International food and agriculture groupClient anonymised
R&D knowledge in 6 languages, 6 single-language agents
Research experts ask in their own language. A router sends the question to the agent built for that language, so instructions stay short, the knowledge base stays clean, and no answer drifts into English halfway through.
- Migrated from Azure AI Foundry to Copilot Studio, then industrialised with retries, caching and history
- Silent single sign-on with a per-user token, no shared secret in the web client
- Off the premium model tier, cost per interaction down
Demo, replayable at any clientOwn build
A procurement agent that checks the request against policy first
4 Copilot Studio agents qualify a purchase request against a written policy, reject what fails it with 2 reframes so a no is still a lead, and score the supplier responses. It ships with its own synthetic data.
- Instructions do not restrict an agent; the capability has to be cut, so the tool surface was
- Qualifies, rejects and scores against the policy document, not a hidden prompt
- Replayable anywhere: synthetic suppliers, requests and responses come with it
Full application, own buildOwn build
Growth Engine: one backend, two surfaces
Goals, activities, news feeds and LinkedIn publishing in one React app on a Python backend. The same backend is exposed as an MCP server, so an agent and a person work from one source of truth with nothing to keep in sync.
- 49 endpoints across 18 routers, React front end, Python backend
- Local model inference by default, cloud fallback case by case
- LinkedIn's terms are quoted in the config next to the code they constrain, and publishing needs a human action
Picked up by MicrosoftPublic
Prompt Architect: an agent that writes prompts in 4 steps
Understand, draft, refine, ship. It asks at most one clarifying question, then produces a persona, instructions and an output specification. Microsoft PnP featured it as Prompt of the Week.
- Bilingual, packaged, reusable at a client in half a day
- One of 8 declarative agent packs ready to load into Microsoft 365 Copilot
- Open on GitHub, with the prompt library alongside the agent
Next step
Everything here is verifiable.
If you dispute a figure, I will show you where it comes from. A half-day workshop on a real use case of your choosing is the fastest way to see how I work.
Elliot Margot · Microsoft MVP · e-margot.ch