03 · Backend · Growth · Data & AI
LangGraph + FastAPI + Postgres
Pipelines and agent workflows as explicit graphs, each step typed and retried, served through the same FastAPI you started with.
Agent fit4/5 Growth · Charge for it Reviewed
The stack
- Framework
- FastAPI + LangGraph
- Runtime
- Python 3.14 with uv
- Data
- Postgres with pgvector, object storage for raw files
- Hosting
- Fly or Modal for the API, Modal or a worker pool for jobs
- Also
-
- Pydantic models as the graph state
- A queue (Redis or Postgres-backed) with retries and dead letters
- Langfuse or OpenTelemetry for traces
- Evals in CI for the LLM steps
Why agents do well here
4/5- A graph with typed state makes every step's input and output explicit, which is what lets an agent modify one node without breaking the rest.
- Idempotent steps with retries turn flaky external calls into a visible queue instead of silent data loss.
- Traces and evals give the agent something to check against, which matters more here than in any other profile.
Avoid at this tier
- Hiding LLM calls inside helper functions; each is a node with a schema.
- A separate vector database while pgvector still fits.
- Shipping prompt changes without an eval run.
How it fits together
Growth-tier data work is pipelines that fail partway and workflows that call models. Make both explicit graphs with typed state so the agent can see each step, and put retries and traces around every external call.
Folder layout
app/
api/ FastAPI routes
graphs/ <pipeline>/ {graph.py, nodes.py, state.py, test_graph.py}
models/ Pydantic state and schemas
jobs/ queue workers running graphs
evals/ datasets and scoring for LLM nodesFirst five decisions
Lock these in before the agent writes the second feature. Each one removes a choice it would otherwise make differently every time.
- 01 Typed graph state
- 02 One queue with retries
- 03 Pgvector in Postgres
- 04 Traces on every node
- 05 Evals in CI for prompt changes
Also fits this cell
Same tier, same profile, a different stack. Pick one of these only for the reason given.
- FastAPI + Jev Decision Service (Python) (opens in a new tab)
The pipeline makes typed decisions (route, classify, approve) that must be logged, thresholded and replayable, not generated text.
When to move on
Stay here until the product forces the next tier. You are ready for Scale when:
- More than one team deploys backend code independently
- Data lives in more than one region or under residency rules
- An auditor will ask who changed what
- Latency and throughput are contractual, not aspirational