The starting point
Project 001 is an independent effort to develop an AI agent for fibre-to-the-home (FTTH) design. The intended ingredients are local AI models, public Polish geospatial data, computational tools and explicit tests.
This first entry establishes the public scope. It does not report a completed experiment, a working end-to-end design workflow or a verified network design.
The engineering question
A model can produce a convincing explanation of a route. That explanation still needs to be checked against geometry, input quality, numerical constraints and the assumptions of the task.
The question for this project is whether an agent can coordinate those checks in a way that remains inspectable. A useful record needs to show how an output was obtained and why it should, or should not, be trusted.
What a useful test entry must contain
- A bounded question. State the behaviour being tested and the expected outcome before the run.
- Traceable inputs. Identify the public source or explicitly synthetic fixture, date, reuse conditions, coordinate reference system and transformations.
- An environment. Record model and tool versions, relevant settings, hardware and the code revision.
- Observed output. Preserve the actual result and relevant logs. Distinguish a tool result from the model’s interpretation of it.
- An independent check. Compare against a reference or clearly stated invariants, including units and tolerances where relevant.
- A consequence. Explain what failed, what changed and what the rerun showed. Leave unresolved questions visible.
This is the reporting format for future work, not a claim that these tests have already been run.
Planned first check
A small synthetic geometry fixture is one possible starting point: known coordinates, a declared coordinate reference system and an independently calculated expected distance. The test would check whether the workflow uses the intended units and exposes enough intermediate information to verify the calculation.
The run has not been documented here. There is no measured error, timing result or pass/fail outcome to publish yet. Even a successful distance check would establish only that narrow behaviour; routing feasibility and network design would remain separate questions.
Publication boundaries
This notebook excludes customer data and employer materials. Public data is not assumed to be complete, current or unrestricted simply because it is accessible. Each dataset used in a published test needs its own source and reuse record.
No illustrative map will be presented as an engineering result. No performance figure will be added without the conditions under which it was measured.
Next step
Document one actual run with its inputs, expected behaviour, observed output and check. If it fails, record the failure before describing the fix. The next useful entry is evidence from that work.