Request for feedback: offline OSM fixture and map-matching contracts

vrajpatell · September 4, 2026, 7:34pm

Hello OSGeo community,

I’m building OpenTrace ML, a pre-alpha Apache-2.0 Python library for experiments that combine computer-vision road observations, incremental traffic forecasts, consent-gated GPX preparation, map-matching contracts, and transparent route-reliability scores.

The current work is deliberately fixture-based and offline. This example runs from small original fixtures, without a routing service or bundled third-party data:

git clone https://github.com/vrajpatell/opentrace-ml.git
cd opentrace-ml
python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
OPENTRACE_PSEUDONYM_KEY='replace-with-a-secret' python examples/map_match_fixture.py

OpenTrace treats OpenStreetMap as externally licensed network data. It does not bundle OSM extracts, upload automated edits, or treat road-damage detections as a source for OSM edits. Any future contribution workflow would require an independently authorized source, licence compatibility, and human review.

I’d value feedback from geospatial Python and routing developers on two bounded tasks:

an offline OSM integration fixture: Add a tiny offline OpenStreetMap integration fixture · Issue #3 · vrajpatell/opentrace-ml · GitHub
a timestamp-validation defect in private-trace cleaning: Reject conflicting source points with the same timestamp before trace filtering · Issue #16 · vrajpatell/opentrace-ml · GitHub

Repository: GitHub - vrajpatell/opentrace-ml · GitHub

In particular, what would make an offline fixture useful enough for practitioners who use PostGIS, Valhalla, OSRM, or FMM?

vrajpatell · September 12, 2026, 9:20pm

Progress update — 12 September 2026

OpenTrace ML’s latest main branch adds a native Go core and portable linear-model inference: a model trained in Python can be exported as data-only JSON and used for single-step or recursive forecasts in Go, with conformance tests.

Python also now has an experimental CPU neural forecaster, persistence/seasonal baselines, and reproducible rolling benchmarks with strict timestamp alignment and per-lead errors. Neural export to Go is still an open design task. The initial public-data benchmark is only one UCI traffic window, and its neural fits reached the iteration limit; broader evaluation is needed.

The geospatial workflow remains at the library-contract and fixture stage: consent-gated GPX preparation, map-matching result validation, observations exported as GeoJSON, and transparent route signals. This is still pre-alpha work.

Current geospatial contribution requests:

#3 — tiny offline OSM fixture: a small local graph test with documented source, attribution, and checksum.
#8 — split traces around long recording gaps: avoid implying a continuous journey across missing observations.
#10 — minimum-contributor gate for unmatched-trace aggregates: review the aggregation contract before any publication workflow.

These last two safeguards remain open work. Pseudonymous location traces are still sensitive, and generated observations remain separate from OSM; no automatic OSM edits are submitted.

If you work with PostGIS, Valhalla, OSRM, or FMM, which small fixture and expected outputs would help you test an integration? A sample contract review or reproducible edge case is a useful contribution.

Forecasting setup and limitations