Skip to main content

PortfolioCase study

UR-01Live · simulated fleet

Urja

The trip is done. Where did the money go?

An AI munshi for small fleet owners: it reconciles each trip's diesel, tolls and km, and every morning tells the owner, in Hindi or English, which rupees don't add up and how sure it is.

Urja's Today screen for a simulated fleet: “Your trucks earned ₹1,86,400 yesterday. ₹11,430 of it doesn't add up, across 3 trips.”, a night-highway truck reconstruction and three flagged trips.
  • 5 rules

    raise every flag, each with a confidence level

    Structural

  • 2 of 23

    flags wrong in the simulated month, shown to the owner

    Structural

    The false-accusation guardrail is under 10%, and the product shows it as “When Urja was wrong”.

Evidence keyStructural

Why it exists

Diesel is the biggest cost, and the owner only sees it after the trip.

Owners of 10–100 trucks run the business from a phone and WhatsApp. Fuel is the largest operating cost, and vendor sources claim a large share of it leaks — directional figures, not proven ones. The owner's question is simpler: why did a finished trip make almost nothing?

The truck finished the trip — why did I make almost nothing?

— The owner's question, Urja Discovery PRD

The morning

Read the brief, open the evidence, ask the driver.

A morning brief in Hindi or English names the trips that don't add up. Each flag opens its evidence — the route, the fuel curve and what the truck recorded — with a confidence level and the driver's side, and nothing is deducted until the owner decides.

A simulated trip's evidence page: Jaipur to Delhi on NH48, a map with the stop near Behror marked, “38 L diesel unaccounted, ₹3,420”, high confidence, what the truck recorded, and a button to ask the driver on WhatsApp.
One flag's evidence (simulated fleet)
The morning brief in Hindi on a phone-width page: yesterday's earnings of ₹1,86,400, a 14-day bar chart, and three trips to look at.
The morning brief, in Hindi

Decisions

Rules find the money; the model only answers questions.

  • Could have
    A broad telemetry dashboard
    Chose
    One problem, found and explained
    Because
    Building is the cheap part; the work is finding the right problem for one owner.
  • Could have
    Let a model find the leaks
    Chose
    Deterministic rules, with a live model only in Ask Urja
    Because
    A fully live design risked invented rupee figures and a fragile demo; rules make every figure explainable.
  • Could have
    Measure engagement
    Chose
    ₹ recovered per truck, guarded by the false-accusation rate
    Because
    A flag that wrongly accuses a driver costs trust, so every flag shows its confidence and the driver's side.

How it works

From the truck's data to the owner's morning.

  1. Trip datafuel sensor, GPS, tolls, bills
  2. Five rulesfuel drop, refuel, km/L, route, toll
  3. Confidencehigh, likely or check
  4. Morning brieftemplated, Hindi or English
  5. Ask UrjaGemini, with a labelled fallback
One trip through Urja

Evidence

Live on a simulated fleet — no users yet.

Urja runs live on a fictional 24-truck fleet. 1,235 unit and 515 end-to-end tests passed and the security review found nothing. The Ask Urja eval did not pass: the final run hit the model's free-tier quota, so the live Ask box usually shows its fallback answer.

What isn’t measured

  • No owners, trips or recoveries from a real fleet are recorded.
  • The Ask Urja eval gate is recorded as failed.
  • The Hindi copy has not had a native-speaker review.

Sources

Everything here is backed by a real artifact.

  • Research
  • PRD
  • Design
  • Evaluation
  • Live data

Evidence behind the Urja case study

Private working documents are listed with what they support; only public sources link out.

  1. Discovery PRD

    Research28 Sep 2026

    Supports: The owner, the problem and the pain signals

  2. Solution PRD

    PRD28 Sep 2026

    Supports: The screens, the five rules and the AI design

  3. Decision log

    Design28 Sep 2026

    Supports: Problem-first scope, rules over a live model, the guardrail

  4. QA report and Ask eval

    Evaluation5 Oct 2026

    Supports: Test results, the security review and the failed eval gate

  5. Live captures

    Live data18 Oct 2026

    Supports: The screens shown here

Links and next steps

Solo build · PM, design and AI-agent buildSep–Oct 2026Live · simulated fleet