Day 75: Pipeline integration: the one API call design
One call in, one structured answer out
The Stage 2 exit criterion is a single API call: image in → {pose, measurements_cm, size, confidence} out. Design that contract deliberately. It should expose enough for a client to render results and understand uncertainty, without leaking pipeline internals. A clean, versioned response schema is a backend instinct that transfers directly — and it's what makes the engine consumable by the Next.js UI and, later, by the Stage 5 gateway.
from pydantic import BaseModel
class Measurement(BaseModel):
cm: float
confidence: float
sources: list[str] # ["pose", "parsing"] — provenance
class MeasurementResponse(BaseModel):
measurements: dict[str, Measurement] # shoulder_width, chest, arm_length...
size: str # "M"
size_confidence: float # 0.91
pose_detected: bool
warnings: list[str] = [] # e.g. "low visibility: left_wrist"Provenance is not optional
Recording which models produced each measurement (sources) is the Stage 2 seed of the model telemetry that becomes central in Stage 5 — 'when a number looks wrong, which component produced it?'. Design the answer to that question into the contract now, and debugging (and the Day-90 honesty report) gets far easier.
Key terms
- Response contract
- The deliberately-designed schema of a service's output — what it exposes and guarantees to clients.
- Provenance
- Recording which model/source produced each result, enabling later debugging and telemetry.
Why include each measurement's source models (provenance) in the API response?
