QualiteknikQA/ML
Work

Selected engagements

Client names and locations are withheld. These are real delivered engagements, and the technical detail is accurate — we're happy to walk through any of them in depth, and to arrange references.

  • Public Infrastructure

    Crowdsensed road surface monitoring at state scale

    A road network of roughly 589,000 km was monitored by manual photo-and-GPS reports: weeks to action, inconsistent data, and no way to separate genuine defects from ordinary vibration. We built a crowdsensing pipeline that reads smartphone accelerometer, GPS and camera data as vehicles drive, classifies the road surface it is travelling on, and distinguishes potholes from speed bumps and driver handling — the false-positive problem that makes accelerometer-based detection unreliable in the field. Smartphone imagery serves as the source of truth, so the models are retrained against evidence rather than assumption.

    Modelling approach follows published research on deep learning for road surface monitoring — Varona, Monteserin & Teyseyre, Personal and Ubiquitous Computing (2020).

    Focus
    Deep learning (CNN, LSTM, reservoir computing), crowdsensing, computer vision
    Engagement
    Proof of concept and MVP — a PWA, live at machinelearnspothole.site
    Outcome
    93% accuracy on pothole and surface classification, with an imagery-backed retraining loop
    Signal Capture

    A phone in a moving vehicle senses the road through the suspension. Each trace is plotted against road position, so every spike sits directly beneath the defect that produced it.

    The tyre bridges what is narrower than its contact patch, so a street gutter registers differently from a pothole of the same depth — and a speed bump produces a clean signature that looks a great deal like a defect. Separating those three is the classification problem.

    Model Evaluation

    Detecting a pothole is the easy half. Not calling a speed bump a pothole is the half that decides whether a road authority can act on the data — so the evaluation is reported per class, against five alternative models, rather than as one headline number.

    01

    Accuracy by model

    Stability event classification. The CNN is compared against two other deep models and three non-neural baselines on identical data.

    CNN
    0.00
    LSTM
    0.00
    Reservoir
    0.00
    TClass
    0.00
    DTW
    0.00
    EasyGR-NB
    0.00
    02

    Where the errors go

    Overall accuracy hides which mistakes a model makes. The diagonal is correct; everything off it is a misclassification, and the shade is that error's share of the true class.

    Model
    Call
    Door
    Message
    Pothole
    Speed bump
    Street gutter
    Call
    256
    6
    9
    1
    3
    0
    Door
    0
    302
    0
    0
    0
    0
    Message
    3
    13
    246
    1
    0
    0
    Pothole
    0
    0
    1
    246
    5
    0
    Speed bump
    9
    0
    7
    28
    200
    12
    Street gutter
    0
    0
    3
    1
    15
    241
    Rows: true · Columns: predictedCorrectMisclassified
    03

    The classes that actually matter

    Per-class F-measure on the stability-event task. Every model finds potholes. Only one holds up on speed bumps and street gutters — the two events that generate false pothole reports.

    Call
    0.94
    Door
    0.97
    Message
    0.93
    Pothole
    0.93
    Speed bump
    0.84
    Street gutter
    0.94
    F-measure, 0 → 1CNNFive other models

    Varona, Monteserin & Teyseyre — “A deep learning approach to automatic road surface monitoring and pothole detection”, Personal and Ubiquitous Computing 24:519–534 (2020). Figures reproduced from the published results.

  • Government & Civic Services

    Civic infrastructure complaint and repair platform

    A development authority tracked road repairs, footpaths, parking and waste hotspots through disconnected manual processes, with no shared view of what had been reported or resolved. We ran the feasibility study and wrote the SRS, then built a web and mobile platform with separate agency and administration dashboards: the full complaint lifecycle from report to resolution, categorisation by type and by responsible agency, configurable zones and sectors, and an audit log on every transaction so accountability survives a dispute.

    Focus
    Systems analysis, full-stack platform, real-time reporting
    Engagement
    Feasibility study, SRS, and build
    Outcome
    Real-time complaint dashboards with agency-level accountability and a complete audit trail
  • Legal Technology

    Resilient extraction pipeline for public court records

    A records pipeline against a public high-court portal kept breaking. It was built on absolute XPaths that failed the moment the page shifted, and — the more serious fault — it reported success it had never verified, so silent failures were indistinguishable from clean runs. We rebuilt it on ID- and content-based selectors, and moved success detection next to the operation performing it, so no step can report done without positive confirmation. One year's cohort produced around 150 structured case tables and 157 judgment documents, each retrieval checked rather than assumed.

    Focus
    Data engineering, failure detection, structured extraction
    Engagement
    Rebuild and hardening of an existing pipeline
    Outcome
    Verified retrieval in place of assumed success; silent failures made visible
  • Judicial Administration

    Court case notification and reporting system

    A court administration tracked hearings, orders and case status across spreadsheets and paper, which made it slow to answer two basic questions: what is listed today, and where does this case actually stand. We built an operator system covering the full case lifecycle — registration, order details, current status and final disposal — with hearing-wise listings, filtered search across the case register, and direct access to judgments and orders, so notifications of hearings, filings and deadlines are issued from the same record everyone works from.

    Focus
    Workflow design, full-stack application, reporting
    Engagement
    Design through delivery
    Outcome
    A single case register with hearing-wise listings and status visible to operators and administration
Have Something Similar?

Let's talk about what you're building.

We'll ask direct questions about your data and constraints before we tell you what's possible.