Welcome to my Agentic AI laboratory. This repository is a monorepo of autonomous, multi-agent systems. Each project is self-contained and showcases a different architectural pattern of modern AI engineering — multi-agent collaboration, agentic tool-use, LLM-as-judge grounding, and local-model orchestration.
Built with AI-assisted development (Claude, Copilot) — consistent with modern engineering practice.
- Languages: C# (.NET 9/10), Python
- Frameworks/Orchestration: Semantic Kernel, OpenTelemetry + Langfuse, FastAPI (ML sidecar)
- LLM Providers: Config-switchable — Ollama (local), OpenAI, Azure OpenAI, Gemini, Claude, OpenAI-compatible endpoints
- Core Concepts: Multi-agent collaboration, function calling, agent self-critique, groundedness judging, deterministic fallbacks, cost guardrails
A multi-agent backend that watches a live manufacturing database for production anomalies and coordinates five Semantic Kernel agents — Auditor, Reviewer, Fetcher, Reporter, and a Groundedness Judge — to investigate, draft a root-cause report, and grade the report against the actual query evidence.
- Architecture: Supervisor cycle with an in-line critic (SQL review before execution) and an LLM-as-judge grading pass.
- Core Tech: C# (.NET 10), Semantic Kernel, Postgres, scikit-learn Isolation Forest (FastAPI sidecar), OpenTelemetry → Langfuse, per-cycle token/cost tracking.
- Explore Project 📂
An autonomous agent that triages server-node hardware failures from a raw diagnostic
bundle spanning five telemetry sources (Redfish BMC events, PCIe AER registers, kernel
dmesg, pre-OS boot-stage POST/UEFI logs, and DPU-internal console telemetry), isolates
the failing subsystem via LLM tool-calling against native C# parsers, and emits a
structured RCA report (JSON + Markdown) with deterministic cross-checking.
- Architecture: Two-phase agent loop (investigate with tools → validated structured submission) with a deterministic rule-based fallback so the pipeline always ships a report.
- Core Tech: C# (.NET 9), Semantic Kernel, local quantized LLM via Ollama (llama3.2), xUnit, Docker Compose.
- Explore Project 📂 | Interactive Demo 🖥️
A hands-on lab for the server management plane: a lightweight Python Redfish client library (shared across the monorepo) driven against the DMTF Redfish Interface Emulator, covering power actions, BIOS attributes, thermal/power telemetry, event logs, and BMC account management — no real hardware needed.
- Architecture: Reusable client library in
Shared/, spec-faithful mock BMC as the demo target. - Core Tech: Python, DMTF Redfish schema, requests-based session auth.
- Explore Project 📂 | Library 📚
A Redfish-compliant mock BMC (Baseboard Management Controller) that exposes DMTF Redfish REST APIs to report component inventory (CPU, GPU accelerators, PCIe devices), stream platform telemetry, and run automated diagnostic passes — including under injected, edge-case hardware failure modes. No real hardware needed.
- Architecture: Layered ASP.NET Core service (Api → Core domain → seed Data) validated by three test suites — unit, Redfish/OData contract-conformance, and concurrency + fault-injection stress.
- Core Tech: C# (.NET 9), ASP.NET Core, DMTF Redfish data model (DSP0266) with OData annotations, Swagger/OpenAPI, xUnit.
- Explore Project 📂 | Architecture 📐
A reference project where a single product-catalog domain is exposed over five transports
— REST, GraphQL, gRPC/protobuf, raw WebSocket, and SignalR — each a thin adapter over the
same application service, so the protocols can be compared apples-to-apples. A second tier
layers an agentic AI surface on top: an MCP server that exposes the domain as tools,
retrieval-augmented semantic_search over a vector store, and a Semantic Kernel agent that
orchestrates both the in-house MCP and an external community MCP.
- Architecture: Shared Domain/Application layer as the single source of truth; every protocol (and the MCP tool layer) is an interchangeable transport adapter — adding one is additive, never a rewrite.
- Core Tech: C# (.NET 9), ASP.NET Core, HotChocolate (GraphQL), gRPC, SignalR, Model Context Protocol (MCP), Qdrant (RAG), Semantic Kernel with a config-switchable LLM (Ollama / OpenAI / Gemini).
- Explore Project 📂
Shared/redfish_toolkit— Python Redfish/BMC client (systems, chassis, managers, accounts) used by the Redfish API Lab and available to any project needing out-of-band hardware telemetry.
Note: this repo uses a git submodule (the DMTF Redfish emulator in the Redfish API Lab). Clone with
git clone --recurse-submodules, or rungit submodule update --initafter a plain clone — otherwise that folder will be empty.
Each project directory has its own README with setup instructions. Several use Docker Compose for their infrastructure (Postgres/Langfuse for AeroMind IQ; Ollama for RCA Engine), and degrade gracefully when optional services are unavailable.