FIELD_NOTES.md

Field Notes: Tiny Dispatch Coach

What changed after the first prototype

The first version was a normal route optimizer wrapped in Gradio. That was not enough for Build Small. The useful part of the product is not only route math; it is turning messy human dispatch notes into constraints that a planner can actually verify.

The current design uses OpenBMB MiniCPM5-1B-GGUF as the small local model for constraint parsing. The deterministic optimizer then plans routes with capacity, time windows, waiting time, lateness, and a manual baseline comparison.

Why MiniCPM5-1B

• It is an OpenBMB model, matching the hackathon sponsor category.
• It is 1.08B parameters, far below the 32B rule.
• The GGUF release can run locally through llama.cpp.
• Its model card highlights local deployment, tool use, long context, and compact agent workflows, which fit this route-coaching task.

Competition fit

The project is intentionally small in both model size and product scope. It is not a general logistics platform. It handles one common small-business workflow: read the daily order sheet, interpret the dispatcher note, and produce a route plan that a human can audit.

This directly targets the hackathon signals:

• Backyard AI: practical helper for a local delivery operator.
• Off the Grid: no cloud LLM API.
• Llama Champion: MiniCPM5 GGUF is available through llama.cpp, behind an explicit checkbox so the public CPU Basic demo remains responsive.
• Sharing is Caring: the planner trace is included as agent_trace.json.

What the model does

MiniCPM5 receives dispatcher notes such as:

Start at 8:00. School and clinic stops are urgent. Fresh produce should be
delivered before lunch. Van capacity 18.

It returns a compact JSON constraint object:

{
  "prefer_early_priority": true,
  "avoid_late_penalty": 2.0,
  "max_route_load": 18,
  "depot_start": 480,
  "soft_due_before": 720,
  "boost_terms": ["school", "fresh"]
}

The planner treats those constraints as inputs. It does not let the language model invent routes or metrics.

Privacy stance

The demo data is synthetic. The app stores nothing, uses no cloud LLM API, and does not require user secrets. Uploaded CSVs are processed only during the Gradio session.