LoadMaster checks a loaded trailer against the federal cargo securement rules (49 CFR Part 393, Subpart I) and tells the hauler, in plain words, what to fix before the trailer rolls.
It is a rules system, not a picture describer. SAM 3.1 (Meta Model API) segments the load, the trailer bed, the wheels and axles, and the tie-downs, and turns those masks into measurements: tie-down count and coverage, rear overhang, load balance against the axle position, tarp coverage. A deterministic rules engine applies the actual securement rules to those measurements and returns PASS, CHECK THIS, or COULD NOT TELL for every check, with the rule citation and one specific fix for every problem. Muse Spark (Meta Model API) turns those verdicts into a plain-words report and independently cross-checks photos; when the two models disagree, the verdict drops to CHECK THIS. A fix-and-recheck loop lets the hauler correct what the camera missed and see exactly which verdicts changed.
- The working app: https://littlestjames82-sys--loadmaster-fastapi-app.us-east.modal.run Open it on a phone, take or choose a photo of a loaded trailer, and get the full check back. The first run after idle can take a minute or two while a container starts; later runs are much faster.
- Sample runs: https://littlestjames82-sys--loadmaster-fastapi-app.us-east.modal.run/gallery Three real pipeline runs (firewood side view, excavator lowboy rear view, ratchet-strap anchor close-up), captured from the app itself.
- Field results: https://littlestjames82-sys--loadmaster-fastapi-app.us-east.modal.run/results Every captured run of the live pipeline, unedited: 11 loads, 68 rule verdicts (2 PASS, 6 CHECK THIS, 60 COULD NOT TELL — where a photo cannot prove a check, the app says so instead of guessing).
- No photo handy? The app itself has a one-tap sample run: it puts a real firewood load through the full live pipeline, same as any photo you would take.
Real output, straight from those runs (detections outlined on the photo, verdicts one tap away in the app):
- Cargo securement was the official vehicle focus area of CVSA Roadcheck 2026, and it still ranked #4 among vehicle out-of-service violations: 1,724 violations, 12.4 percent of the vehicle total, in three days of inspections (CVSA, 2026 results).
- FMCSA's 2025 full-year numbers (via CVSA) count 18,108 violations for cargo not secured to prevent leaking, spilling, blowing, or falling, plus 16,054 for unsecured vehicle components and dunnage.
- The AAA Foundation's 2026 update counts road debris, most commonly unsecured cargo and fallen vehicle parts, in more than 319,000 crashes with more than 32,000 injuries and 433 deaths across 2018 to 2023.
The rules exist and are enforced. What is missing is a way for the person tying the load down to check it before an inspector, or the road, does it for them.
- Impact (30%): the numbers above; a tool aimed at the independents who have no safety department (91.1 percent of US carriers run 10 trucks or fewer, ATA Trends 2026); a Spanish report option and a listen-aloud report for the workforce that actually hauls (Hispanic drivers are 25.7 percent of US truck drivers, EPI/CPS 2025; 28 percent of Hispanic adults are smartphone-dependent for internet, Pew 2025).
- Use of Meta AI Tooling (25%): SAM 3.1 does the geometry (segmentation, calibrated measurement), Muse Spark does two distinct jobs: an independent second read of the photo that can demote verdicts, and the plain-words report written strictly from the engine's verdicts. Both models, two roles each, one deterministic referee between them.
- Demo (25%): the live app link above works right now; the gallery shows real runs with no upload; the 3-minute video shows a failing load, the fix, and the recheck flipping it to PASS.
- Technical Depth and Execution (20%): docs/ARCHITECTURE.md (design decisions, including the measured reason straps are counted from close-ups plus tap marks, never from full-frame guesses), 150 automated tests, examples/output/ (real pipeline artifacts), BUILD_LOG.md (the whole build, in the open).
The closest product, CargoPro, is a per-driver checklist app: it asks the driver to count the tie-downs, does working load limit math on the numbers typed in, and attaches photos as unanalyzed evidence. Telematics platforms (Samsara, Motive) digitize inspection forms at fleet prices with hardware. The incumbent for most independents is a plastic sliding calculator. LoadMaster is the one that looks at the photo: it sees the load, measures it, applies the actual rules, cites them, and has a second model check the first one's work.
- SAM 3.1 pipeline: prompt packs keyed to trailer type, load type, and photo role; parallel calls with retry; mask decoding; an editable detection model designed for tap-to-correct; annotated overlay renders.
- Calibrated measurements: bed-length calibration turns pixels into inches (rear overhang, load centroid vs axle position with an uncertainty range, tarp coverage, tie-down coverage gaps). The layer refuses to guess a scale it cannot support.
- Deterministic 49 CFR 393 rules engine: the 8 checks below, each verdict linked to its evidence (mask region, measurement, rule citation) and a minimum fix, risk-ranked.
- Dual-model cross-check: Muse Spark reads the photo independently; disagreement demotes a verdict to CHECK THIS, never upgrades one.
- Muse Spark report: a short, plain-words, fix-first summary written strictly from the verdicts (a template fallback covers any model failure). Available in English or Spanish, with a listen-aloud readout in the app.
- Tap-to-correct + instant recheck with a visible before/after delta; guided capture with photo quality gates; a printable one-page report card; a judges' sample gallery of real runs.
- Multi-photo runs: side + rear + close-up photos of the same load are analyzed together and judged as one set, the way the rules engine was designed to see a load. The side view carries the measurements, the rear view judges lights and plate, close-ups carry the tie-down count. Installable as a phone app (PWA).
- Scoring harness (tools/score_runs.py): runs photo sets through the same path the app uses and diffs every verdict against expected statuses, built for the staged test set where the ground truth is known.
- Measured on itself: across the 11 full runs captured during development (68 rule verdicts), the second opinion's independent answers are recorded in the evidence of 19 verdicts, and in the live close-up run its strap count disagreed with SAM's (2 vs 1) and moved the edge-protection verdict on its read alone.
- 150 automated tests (pytest), all passing. Architecture and design decisions: docs/ARCHITECTURE.md.
- Hosted as a FastAPI app on Modal, deployed by GitHub Actions on every push to this repository.
Built by Ryan Cotten (Ghost Developer Studio, Oakdale, Tennessee) for the Meta Global AI Developer Hackathon, Indie Developer track, theme: Build for Your World. Ryan hauls firewood, equipment, and trailers in East Tennessee. This is a real problem from his own work.
A photo check can count visible tie-downs and flag placement. It cannot verify strap tension, actual working load limit ratings, or load weight. LoadMaster says so in every report, and any check that depends on something a photo cannot show comes back COULD NOT TELL, with the reason named, never a guess.
- Tie-down count by load length (393.110)
- Tie-down strength: aggregate working load limit at least 50 percent of load weight (393.106, 393.108); photo-limited, see honesty line
- Heavy equipment, 10,000 lb or more: at least 4 tie-downs, accessories lowered and secured (393.130)
- Forward blocking against sliding under braking
- Edge protection where straps or chains cross sharp edges
- Overhang, and lights or plate obscured
- Tarp coverage of loose material
- Weight placement: centered, heavier end forward of the axles (reported as a visual estimate only)
Everything in this repository was created inside the hackathon window. Pre-window work was research only: a written spec grounded in 49 CFR 393, with no code and no repository.
- Oct 26: repository created; SAM 3.1 and Muse Spark access proven on the Meta Model API (first SAM call: one trailer mask, 0.863 confidence, on a development photo).
- Oct 26 (Day 1): the SAM pipeline is built: prompt packs keyed to
trailer type, load type, and photo role; mask decoding; an editable
detection model built for tap-to-correct; and annotated overlay
renders. Proven on development photos, outputs in examples/output/.
Run it with
python -m loadmaster.pipeline --photo PHOTO --trailer-type utility --load-type firewood --role side --out-dir examples/output(dependencies in requirements.txt). - Oct 26 (Day 2): the measurement layer: calibration, overhang, centroid vs axle as a labeled visual estimate, tarp coverage, tie-down gaps, each with honest refusal paths.
- Oct 26 (Day 3): the deterministic rules engine (all 8 checks) and the Muse Spark cross-check, then the working app itself: FastAPI service + mobile web UI, deployed live on Modal the same day.
- Oct 26 (Day 4 build session): Muse Spark report layer, recheck delta, guided capture + quality gates, printable report card, and the judges' sample gallery. First real end-to-end use on a phone: a 19-bale hay load, every bale found, verdicts matching the photo.
Build progress is logged in BUILD_LOG.md in the order it happened.
- Meta Model API: SAM 3.1 (segmentation) and Muse Spark 1.3 (report reasoning), accessed with an API key that is never stored in this repository.
- meta-sam-parser (Meta's SAM output parser) for mask decoding; the build log records the Day 1 route decision.
- Pillow for image handling and overlay rendering.
MIT. See LICENSE.