// service
Agent-Fleet Platform — Memory, Routing, Observability
A governed multi-agent platform with shared memory, model routing, and full token/cost/trace telemetry — so fleets are operable, not just demoed.
shape
8–12 weeks
NVIDIA · Intel · Red Hat · Grafana/OpenTelemetry ecosystem
outcome
Teams deploying more than one agent inherit the three problems that kill them at scale: memory, routing, and knowing what anything costs. You get a shared MCP gateway and GPU-accelerated RAG memory tier, a programmable mixture-of-models routing layer, and complete token/cost/trace observability wired into Prometheus/Grafana-grade tooling with budget alerts. The result is a control plane your ops and finance teams can both live with.
what you get
- Agent-fleet reference architecture: orchestration, memory (RAG), routing, and observability layers
- Shared MCP gateway and 3-tier RAG memory implementation over GGUF embeddings
- Token/cost/trace telemetry stack (LiteLLM + OpenTelemetry to Prometheus, Loki, Tempo, Grafana; tracing via Arize Phoenix)
- Budget alerting and cost-per-workflow reporting
evidence — public, checkable
- qmd-memory-stack — local GPU-accelerated 3-tier RAG memory, ~19× query speedup (Intel Arc iGPU, Vulkan)
- qmd-gateway — shared, warm, writable MCP memory gateway for a multi-agent fleet
- agent-observability-stack — LiteLLM + OTel → Prometheus/Loki/Tempo/Grafana/Phoenix; video walkthrough public
- semantic-router — mixture-of-models routing
- Multi-agent fleet orchestration at scale, delivered (anonymized capability category)
- Guarded, agent-operated control planes with GitOps DR (anonymized capability category)
agent-observability-stack on GitHub qmd-memory-stack on GitHub
The observability stack is self-hostable end to end and GPU-aware (Intel iGPU/NPU plus NVIDIA RTX 3090 eGPU), with dynamic onboarding and Telegram budget alerts. The memory tier fails open: if recall degrades, queries still answer. The routing layer treats models as a portfolio — policy decides who serves what, and telemetry proves the policy right or wrong on dated numbers.
pairs with
Applied Agent Delivery
One named workflow — analysis, triage, content, assurance — delivered as a working agentic application with provenance-bound, unattended operation.
Agentic CloudOps & AIOps — Signal to Action
Your telemetry stops being a wall of alerts and starts closing loops — events correlated, remediated under guardrails, and verified resolved.
Sovereign Inference & GPU Economics
Production-grade LLM serving on your own NVIDIA GPUs, with a dated, benchmarked answer to "local or frontier?" per workload.