CASE STUDY / 001

Reliable intake.
Predictable failure paths.

Two separate components, one practical question: how do we verify that the system handles the unhappy path?

AI-assisted · synthetic data · local demonstration

Read the technical brief Download reference source

81tests passed · 2026-09-09

6recorded API scenarios

0live model calls in tests

Evidence captured on 2026-09-09. Results below are historical records, not a live service or a new test run.

RECORDED WALKTHROUGH

What the local example actually does

Silent recording, approximately 57 seconds. Playback was paced for readability; it is not a latency benchmark. The LLM scenes use mock responses.

The PDF and video retain the original 9-scenario browser walkthrough. This viewer shows its 6 API scenarios. The public source ZIP contains regression tests and offline evidence, rather than the complete original demonstration work package.

EXPLORE THE RECORDED RESULTS

Choose a case. Inspect the evidence.

This is a static viewer of recorded test outcomes, not a running API, database, or live model. Nothing you select is sent to an AI provider.

RECORDED RESULT / LOCAL SQLITE EXAMPLE

{
  "mode": "recorded_result / not live",
  "scenario": "valid inquiry",
  "status": 201,
  "result": "created",
  "persisted_inquiries_after_step": 1,
  "audit_events_after_step": 1
}

Database counts follow the original sequential walkthrough.

Two components. Two distinct boundaries.

A / INTAKE API

FastAPI → Pydantic → SQLite

Validates an inquiry, checks a caller-supplied idempotency key, and writes an inquiry and its audit event in one local transaction.

B / MODEL ADAPTER

Provider fixture → parser → typed output

Tests strict output parsing, refusals, incomplete responses, and bounded retry behavior. It is not wired into the intake endpoint.

What is demonstrated

The enclosed local regression suite passes. It exercises request validation, replay/conflict behavior, transaction rollback, selected concurrent submissions, and model-adapter failure handling.

What is not demonstrated

No production deployment, real-model accuracy, customer outcome, or security certification is claimed. No Google Sheets or Telegram integration is included. A different idempotency key represents a new request; this is not distributed exactly-once delivery.

Reproduce the result

The downloadable source includes the tests and setup instructions. Inspect the dependencies and run locally in an isolated environment. The optional --live example can incur provider charges; it is not needed for the tests.

python -m pytest -q
# 81 passed · local verification · 2026-09-09

Read the machine-readable test report ↗

LET’S DEFINE THE FIRST MILESTONE

One clear problem
is a good place to start.

Share the workflow, a sanitized example, and what a successful result should look like. We can agree on scope before work begins.

Start on Upwork

View the brief template

Project goal:
Current stack / workflow:
Sanitized example of the problem:
Expected result / acceptance check:
Deadline and budget range:
Code/data access I am authorized to provide:
AI-assisted development allowed: yes / no / discuss

Upwork-sourced inquiries: please keep pre-contract communication and contracting on Upwork. This site does not accept payment or submit messages.

Reliable intake.Predictable failure paths.