Architected and shipped an AI on-call assistant in C# / .NET 8 that integrates Azure OpenAI (GPT) with 77 read-only tools across 9 backend services via the Model Context Protocol (MCP) — reducing routine incident triage from ~15 minutes of multi-tool context-switching to under 10 seconds.

Designed a dual-runtime architecture (.NET 8 orchestrator + .NET Framework 4.7.2 server) hosting an in-process PowerShell runspace that caches an authenticated session for the process lifetime, eliminating per-call re-authentication and cutting typical query latency to 1–3 s (from a 5–15 s cold start).

Built a streaming ReAct agent loop (Server-Sent Events) with multiplexed tool-calling over two MCP servers, orchestrating up to 16 reason–act–observe iterations per query while holding the production dependency footprint to a single NuGet package.

Engineered "read-only by construction" security with 4 independent enforcement layers and caller-identity passthrough across 4 auth boundaries (Kerberos → dSTS, OAuth Bearer), guaranteeing zero write operations with no standing privileges.

Optimized LLM context efficiency by replacing multi-megabyte raw API payloads with 40+ schema-bounded, filtered tools — shrinking responses from ~18,000 rows to ~50 and eliminating HTTP 429 token-limit failures.

Integrated multi-source RAG grounding (Azure DevOps code + wiki search via MCP), enabling the agent to cite source-of-truth runbooks/TSGs and converting tribal knowledge across ~280 PowerShell cmdlets into shipped, auditable tooling.