Problem
Teams building agentic workflows need orchestration beyond a single LLM call — but full MLOps platforms are overkill for webhook-triggered pipelines with a handful of steps. Existing tools either lack state management or require cloud-only inference.
Approach
- — LangGraph state machines model multi-step agent workflows with explicit state transitions.
- — FastAPI webhook ingress accepts external triggers and dispatches async jobs.
- — Ollama integration for local inference with automatic fallback to cloud providers.
- — CLI tooling for pipeline definition, testing, and local development.
Outcome
- Multi-step agent pipelines deployable with a single FastAPI process — no Kubernetes required for development.
- Local Ollama inference eliminated cloud API costs during pipeline development and testing.
- State machine model made debugging agent failures traceable step-by-step.
Metrics
- Pipeline steps
- Unlimited
- LangGraph state graph
- Inference
- Local + cloud
- Ollama with provider fallback
- Ingress
- FastAPI
- Async webhook processing
- Dev setup
- 1 command
- CLI init + local run
Architecture
Webhook IngressFastAPI async
CLI ToolkitInit · test · deploy
↓
LangGraphState machine
↓
OllamaLocal LLM
↓
Cloud ProviderFallback inference
Trade-offs
Orchestration
Chosen
LangGraph state machines
Alternative
Sequential function chains
Explicit state graphs support branching, retries, and human-in-the-loop steps that function chains can't model cleanly.
Inference default
Chosen
Ollama local with cloud fallback
Alternative
Cloud-only (OpenAI/Anthropic)
Local inference removes API cost and latency during development. Fallback ensures production reliability.
Next project
0%