The KYD engine sits between a raw phone sensor stream and a driver-risk score an insurer can put a premium on. Six load-bearing systems, three fidelity tiers, one continuous score.
GPS, accelerometer and gyroscope sampled with battery- and thermal-adaptive rates, converged to a stable ~5 Hz across every handset OEM, queued durably on device and uploaded in segments throughout the drive. Doze, a network drop or an app kill never loses a trip.
Orientation recovery fuses gravity, GPS-derived acceleration and gyroscope to rebuild the vehicle’s motion frame. Mid-trip re-orientation, low or absent GPS and OEM background-killers are handled by named safety contracts rather than silent guesses.
Per-channel anti-gaming runs before scoring: falsified location, passenger-vs-driver, signal-denied and capture-completeness checks, each versioned and explicit. A trip that cannot be trusted is not quietly scored anyway.
Six dimensions, published weights, vehicle-class-specific matrices, exposure-correct denominators over moving time, and population percentile banding with anchor hygiene.
Per-trip impact, confidence interval, percentile band, event map and timeline — plus an explicit ‘not measurable’ where a dimension abstained and the weights renormalised.
Fleet surfaces read the operational view; the insurer partner API reads a consent-scoped risk view. One capture, two audiences, zero extra collection — and no raw location leaves the platform.
An explainable score is not a nice-to-have in insurance — it is the difference between a signal an actuary can defend and a number a regulator will reject.
| Dimension | Weight | What it measures |
|---|---|---|
| Speed compliance | 25% | Speed against mapped limits, cached offline per metro |
| Braking | 20% | Harsh-deceleration events, rate-corrected for exposure |
| Acceleration | 15% | Aggressive launches and throttle behaviour |
| Cornering | 15% | Lateral force through turns, in the vehicle’s recovered frame |
| Smoothness | 15% | Jerk and variance against the driver’s own baseline |
| Context | 10% | Time of day and road class — night on a highway is not noon in a colony |
Two-wheelers and four-wheelers use separate weight matrices, and a third exists for the case where on-board diagnostics data is present. Two-wheeler risk is not car risk, and scoring a bike on a car matrix is how incumbents get emerging markets wrong.
±3.1 pts
p78–p84
41 min moving
A dimension the data cannot measure returns nothing and the remaining weights renormalise — the engine says “not measurable” rather than inventing a 100. Illustrative interface; figures are representative, not a live customer trip.
Reconstructs the vehicle’s real motion frame from gravity, GPS-acceleration and gyroscope. A phone loose in a cup-holder or re-oriented mid-trip still yields valid physics. This is the make-or-break of phone telematics, and it is solved.
Versioned, explicit, per-channel checks — falsified location, passenger-versus-driver, signal-denied, capture-completeness. A GPS blackout can no longer certify as a clean drive.
An unmeasurable dimension is withheld from score, badges, journeys and coaching; weights renormalise. The engine says ‘not measurable’ rather than inventing a 100. Rare — and the credibility of the whole product.
Moving time, not wall-clock — a lunch stop no longer costs a driver points. Population baseline and percentile banding with anchor hygiene, so tiny-exposure drivers cannot distort p5 and p95.
Partner JWT scoped to INSURER_DATA_SHARING — serves score, history, feature breakdown, trips, portfolio distribution and quote-handoff verification. A legacy unscoped endpoint was found in audit and removed.
OpenStreetMap speed limits cached offline (~36,900 cells, ~12 MB per metro). One-command deploy with auto-rollback, monitored to a single paging path, a WAF auth-flood penalty box, and CI gates.
Fidelity layers compose into the same engine — they never rip and replace it. A fleet that starts smartphone-only and later adds dashcams or GPS tags does not restart its score history.
Sakshya does not ship hardware. Dashcam fidelity arrives through an OEM partner API; the Tier 3 tag is an AIS-140-style device sourced, not manufactured. No bill of materials, no supply chain, no field-operations organisation — which is exactly what makes geographic expansion a software problem rather than a logistics one.
Phone IMU and GPS. Works on any device, no hardware, no install. Powers today’s scoring.
Forward video for event corroboration and dispute resolution, integrated via an OEM partner API.
A dedicated AIS-140-style tag for underwriting- and claims-grade evidence. Same engine, higher bar.
Managers ask in natural language; the assistant reasons over live fleet data, surfaces what matters, takes safe actions and escalates the rest — and it watches for risk without being asked.
Safety-critical actions never fire on a language model alone. The assistant can suggest; the rule engine decides. Every step is logged — for the fleet manager, the driver, the insurer and the regulator.
“Show me all drivers who braked harshly in the last hour.” “Which route had the most incidents this week?” No SQL, no filters, no training.
The bot does not wait to be asked — it surfaces fatigue, overspeed streaks, low-score drivers and SOS events as they happen, with a recommended action attached.
Pull a trip’s timeline, event map, weather and driver history on request. The bot assembles an incident report with a one-paragraph executive summary.
“Move these three drivers off the night shift.” “Send a coaching card to anyone below 65 this week.” The bot drafts the action; the manager approves.
“Which zone is getting riskier?” “Compare this month’s score distribution to last.” The bot reads the analytics so the manager does not have to.
Every read, recommendation and action is recorded with the model’s proposal and the deterministic verification behind it. Insurer-grade by design.
Straight answers, including on the things we have not proved yet.
Ask us something elseKYD - Know Your Driving - is Sakshya's scoring backend. It converts raw smartphone sensor data into an objective, anti-gaming, explainable driver-risk score through a durable workflow pipeline, and exposes that score to insurers through a consent-scoped partner API. It is the asset insurers and fleets pay for.
Because the phone is not bolted to the vehicle. A handset in a pocket, a cup-holder or a jacket - or one that is re-oriented halfway through a drive - produces accelerometer readings in the phone's frame, not the vehicle's. Without recovering the vehicle's true motion frame from gravity, GPS-derived acceleration and the gyroscope, braking and cornering measurements are meaningless. It is the make-or-break problem of phone-based telematics.
If the data cannot support a dimension - poor GPS, insufficient exposure, a sensor that failed to report - Sakshya withholds that dimension from the score, badges, journeys and coaching, and renormalises the remaining weights. The engine states that the dimension is not measurable rather than inventing a value. It is rare, and it is the credibility of the whole product.
Rates are computed over moving time rather than wall-clock time, so a lunch stop or a loading delay no longer costs a driver points. Population baselines and percentile banding apply anchor hygiene so drivers with very little exposure cannot distort the p5 and p95 anchors for everyone else.
Not yet, and we say so plainly. Score-to-claims validation is the single proof that converts pilots into insurer contracts, and it is the first milestone the current round funds - correlating the score against real claims and incidents using Tier 3 data, with a co-signed study design alongside an insurer partner. Until that study exists, we describe the score as explainable and anti-gaming, not as claims-validated.
Engineering, actuarial and data-science teams are welcome to interrogate the methodology, the abstention rules and the anti-gaming channels in detail. We would rather answer hard questions early.
No cookies, no trackers, no third-party scripts on this site. Your enquiry reaches the founders directly.
