Money Tracker

1. 1Authorize + persist→
2. 2Route intent→
3. 3Run one specialist→
4. 4Render + queue response

Bounded graphs + explicit dispatcher

Separate router and specialist contracts keep trusted profile, time and action references in runtime context.

Adding an intent requires an explicit branch and schema; this buys reviewable authority at the cost of more orchestration code.

Application PostgreSQL + inbox/outbox

Financial writes, command claims and audit records use transactions. Messaging persists across restarts.

A sequential worker can delay later messages during retry. External delivery can repeat a Telegram reply; finance idempotency protects the write.

Local inference + Telegram transport

Inference and financial storage stay on the workstation. Long polling avoids a public inbound endpoint.

The machine must be available for processing. Telegram is an external boundary, so the complete interaction is not entirely local.

Versioned evaluation gates

Datasets, deterministic graders and run manifests bind an accepted configuration to model, prompt, schema and tool versions.

A model or prompt change requires targeted regression runs. A plausible answer or confidence score is insufficient evidence.

• EUR amounts use integer minor units: €36.00 is stored as 3600, avoiding binary floating-point arithmetic.
• Corrections replace a validated aggregate and record an audit event. Undo voids the movement instead of deleting its history.

Dashboard simulation · fictitious data · not screenshots

From the first screen to the retained baseline

1. 1–2 Sep 2026

   Screen the local interfaces

   Pinned quantized weights, llama.cpp, an 8,192-token context and one inference slot. Probes checked structured output, tool calls, tool-result continuation and image transport. A passing screen kept a candidate eligible for further work; it did not establish financial accuracy.

2. 3 Sep 2026

   Turn failures into new probes

   Qwen3 14B filled fields that the message left unspecified. The extended screen tested optional-field omission and union-shaped output separately. Omission failed; union output worked, motivating more explicit model-facing contracts.

3. 8 Sep 2026

   Repair contracts and evaluation evidence

   Versioned schemas, prompts and development labels were revised with recorded reasons. The reporting path was tightened so failed or unmeasured required gates could not become a pass. Later text profiles and tool experiments widened the comparison.

4. 9 Sep 2026

   Promote a complete configuration

   Qwen3 14B passed every applicable blocking gate across the text workflow suites after deterministic read-selection repairs and owner advice review. EVD-004 pinned the accepted model, prompts, schemas and policy. Thresholds were retained; automatic posting stayed disabled.

5. 10 Sep 2026

   Test the larger challenger

   Qwen3.8 27B was measured after its quantized configuration fitted the workstation. It improved categorization but did not clear routing, tools and advice requirements. It remained experimental; the promoted baseline was retained.

Every model, with the scope of its evidence

Outcomes belong to the recorded model/runtime/schema combinations. Earlier Ministral text runs loaded a vision projector before the profile definition was corrected. Suite and instruction versions also changed over time, so these rows are a history of evidence, not a uniform leaderboard.

What made a run interpretable

Image transport is not receipt-extraction evidence; automated receipts remain deferred. Timing observations were specific to this workstation and run configuration. Missing measurements do not count as passes. This case study summarizes the records without reproducing private datasets, outputs or a benchmark score table.

Operations are part of the design: durable retry and dead-letter states, local model-call tracing, and Restic encrypted backups. Trace content remains private. Acceptance evidence exists in the project; no performance score or private artifact is reproduced here.