Evidence

What we built, and how we checked it

Two of these are deployments. Two are engineering write-ups — because how a measurement problem was actually solved tells you more about whether a vendor can solve yours than a logo does.

  • Methodology

    Internal validation programme

    • Traffic counting
    • Validation
    • IRC:64-1990

    Measuring our own error, and publishing it

    Challenge

    Automated counting accuracy can be stated two ways, and quoting only the flattering one is common practice. We wanted a figure that survives an engineer checking it.

    Approach

    Three surveys were counted by hand, independently, on live uncontrolled traffic — 16,560 vehicles. Software output was compared against those manual counts both by total volume and by weighted per-class accuracy. Categories with fewer than 50 manual counts were excluded as statistically unreliable rather than averaged in.

    Outcome

    95.0% volume accuracy and 92.1% weighted class accuracy. We publish the weighted figure as the headline because it is the stricter of the two — it does not allow over- and under-counts to cancel.

  • Client

    NKMPV

    • Crack detection
    • ATCC
    • Deployed

    Road survey software for a civil engineering consultancy

    Challenge

    A civil engineering and materials testing consultancy running road condition and traffic survey work needed software fitted to Indian survey convention rather than adapted from Western tooling.

    Approach

    We built and deployed a crack detection system for pavement distress from survey footage, and an automatic traffic counting and classification platform calibrated for the local vehicle mix. We also built and maintain their public website.

    Outcome

    Both systems run in production. The distress and counting work directly informed the products now offered here.

  • Engineering

    Playcraft ATCC

    • Computer vision
    • Verification
    • Engineering

    Three defects that were costing 289 vehicles

    Challenge

    Counts were stable but conservative. Aggregate accuracy alone could not say whether a change was a genuine improvement or a metric artefact.

    Approach

    Three interacting defects were isolated: a deduplication rule that swallowed genuinely distinct vehicles, a segment authorisation window that expired before the cooldown it depended on, and an isotropic dedup bubble that could not distinguish an identity switch from the next motorcycle. Each was fixed and verified separately.

    Outcome

    A full production re-run recovered 289 vehicles. A 36-window sweep across one clip showed 31 windows counting more, 5 unchanged and zero regressions. Per-vehicle ground truth on a slit-scan window confirmed the gain was real at the level of individual vehicles, not just in aggregate.

  • Engineering

    Playcraft Distress

    • Photogrammetry
    • Measurement
    • Geometry

    Measuring defect area from a single uncalibrated camera

    Challenge

    A defect flagged is not a defect quantified. Maintenance estimation needs square metres, but survey footage arrives from cameras mounted at unknown heights and angles, with no calibration data.

    Approach

    A row-dependent ground-sample-distance model derived from camera height and vertical field of view, with optional homography where calibration exists. A pixel near the top of the frame covers far more ground than one near the bottom; the conversion accounts for where in the frame each detection sits.

    Outcome

    Defect dimensions in real-world units from ordinary monocular video, once camera height and field of view are supplied — without a survey-grade rig. Uncalibrated footage still yields consistent pixel dimensions, which is enough to rank severity along a route.

The honest position

We would rather show you working software than a client list

We are a small team and our commercial history is short. What is not short is the engineering. If you want to see the code, the architecture, or the tool running on footage you choose, that is how we prefer to be evaluated.

See it run on a call