octagono

Building agentic systems with DSPy & Dapr. Declarative LLM programming meets distributed runtime architecture.

Available for consulting & collaboration
Get in touch GitHub Labs Blog
Expertise
01

DSPy — Declarative LLM Programming

  • Declarative framework — Signature-driven module composition (ReAct, ChainOfThought, ProgramOfThought, MultiChainComparison).
  • Optimizer-driven compilation — BootstrapFewShot → MIPROv2 → GEPA, with multi-objective Pareto optimization and UCB tree-guided evolution.
  • Adapter-level configuration — Chat, JSON, and XML protocol adapters with adapter-level config.
  • LLM serving — vLLM, SGLang, and TurboQuant for efficient batch inference and quantized deployment via Ollama and OpenRouter.
02

Dapr — Distributed Agent Runtime

  • DurableAgent execution — Stateful workflow execution with automatic checkpointing and recovery. Scale-to-zero with persistent agent state.
  • LLMOrchestrator — Dynamic multi-agent coordination with agent-as-tool composition for hierarchical delegation.
  • MCP integration — Standardized tool and resource access across agent boundaries via Model Context Protocol.
  • Knowledge & memory — GraphRAG, Weaviate, and FAISS for semantic search. Pydantic-structured schemas, vector-based conversation memory, and agent registry for discovery.
03

Systems Architecture

  • Self-evolving pipelines — LSE (Learning to Self-Evolve) with UCB tree-guided evolution, improvement-based rewards, and Meta-LSE hyperparameter self-evolution.
  • Distributed orchestration — Ray for parallel execution and scaling. LiteLLM unified gateway across 100+ LLM providers with fallback and rate-limiting.
  • Observability — MLflow-tracked experiments, OpenTelemetry distributed tracing, and model evaluation pipelines.
  • RAG & microservices — Weaviate, FAISS, FalkorDB for semantic retrieval. FastAPI with Redis caching, PostgreSQL persistence, Kubernetes deployment. Formal verification with Lean 4 and Z3 on NixOS.
04

Infrastructure & Security

  • Declarative environments — NixOS for reproducible system states. Kubernetes and Docker container orchestration.
  • GitOps CI/CD — GitHub Actions, GCP managed services, and Redis-backed caching.
  • Zero-trust security — WireGuard networking, AST-based code safety scanning, sandboxed execution with OS-level resource limits.
  • Continuous validation — Kali Linux tooling for security auditing. Post-quantum cryptography (Kyber, Dilithium) for transport security.
Stack
Core
DSPyDaprdapr-agentsPythonFastAPILLMs
Ecosystem
LLM Serving & Inference
vLLMSGLangOllamaOpenRouterLiteLLMTurboQuant
Databases & Search
WeaviateFalkorDBFAISSPostgreSQLRedisGraphRAG
Infrastructure
KubernetesDockerNixOSGCPWireGuardGitHub ActionsOpenTelemetry
Languages & Tools
MCPPydanticReactAstroRayMLflowLean 4Z3Kali LinuxHugging FaceLoRA
Approach
The best agentic systems aren’t built—they’re programmed and evolved. Turn runtime failures into growth. Let UCB tree search find the path. Measure improvement, not perfection. Ship skills that self-evolve—then let Meta-LSE evolve the evolution itself.
Connect

Looking for DSPy consulting, Dapr agent architecture, or building production agentic systems? Let’s talk.

GitHub LinkedIn Instagram Facebook ai@octagono.org