A studio-lit car, three-quarter front view.
carvia

Your car already knows.

Carvia tells you in plain language, weeks before it matters.

Carvia’s read

Your battery is failing.

11.9 V · 82% chance within three weeks

Cars never got their Stack Overflow moment.

I grew up around cars. Diagnosing a problem meant digging through old forums, Team-BHP threads, Facebook groups, someone three pages deep in a thread from 2014 who’d had the exact same rattling noise.

That’s the relationship developers used to have with Stack Overflow, paste the error, hope a stranger hit it first, read until something works. Then AI made troubleshooting instant.

Developers caught up. Cars didn’t.

Then: a 2014 forum thread

forums › technical › suspensionarchived

Rattling noise over bumps, front left [SOLVED?]

Page 3 of 7 · posted 11 Jun 2014

  • torqueaddict_9911 Jun 2014 · 21:14
  • desi_drifter12 Jun 2014 · 08:02
  • bhp_oldtimer14 Jun 2014 · 19:47

    ↳ the answer, 40 minutes of reading later

‹ prev · next ›132 replies · 7 pages
2014 → now

Now: the same question, answered in a chat

Ask Carvia

AI

Rattling noise over bumps, front left, what is it?

Reasoned from the symptom, model and service history. Not a forum crawl.

Carvia is what that shift looks like for vehicle ownership, the forum knowledge in your pocket, reasoning in real time, for a car that already knows more about itself than any forum thread ever could.

What exists today

  • OBD dashboard apps

    data without meaning

    Live PIDs, fault codes, gauges: everything the car says, none of what it means.

  • OEM companion apps

    walled gardens

    Polished, but one brand deep, and the manufacturer grades its own homework.

  • Service marketplaces

    booking without trust

    They sell the repair. They can't audit whether you needed it.

Carvia sits where all three stop: meaning, across brands, on the owner’s side.

Four pillars. One car.

Everything a person does with a car falls into four jobs: owning it, understanding it, driving it, and deciding what to do next. Four pillars, not fifteen features. Onboarding is a lookup rather than a form, so the whole garage arrives from one phone number.

The garage home screen: one vehicle, its odometer, a battery reading and a fuel range. Moments later the battery reading falls to 11.9 volts and an alert appears: an 82 percent chance of failure within three weeks.

  • 01

    Own

    “Never lose track of your vehicle.”

    • documents & insurance
    • service records
  • 02the differentiator

    Understand

    “Understand what your vehicle is trying to tell you.”

    • AI mechanic
    • plain-language codes
  • 03

    Drive

    “Drive better every week.”

    • trip intelligence
    • eco & safety score
  • 04

    Decide

    “Help people make smarter vehicle decisions.”

    • buying advisor
    • repair-or-sell

Every pillar feeds one artifact: the vehicle timeline. That’s where the story goes next.

One life. One spine.

The vehicle's version of a health record: chronological, visual, and written to by all four pillars. A service record, a wear prediction, a long trip, a resale decision: all on one line.

  1. FEB ’22 · 0 KMOWN

    Purchased

    Day one: VIN scanned, documents in, warranty clock started.

  2. FEB ’23 · 9,480 KMOWN

    Insurance Renewed

    Reminded 3 weeks early, zero lapse.

  3. AUG ’23 · 15,120 KMOWN

    15,000 km Service

    Invoice OCR’d, next service predicted.

  4. MAR ’24 · 24,300 KMUNDERSTAND

    Brake Pads Changed

    Wear model flagged them 800 km before the squeal.

  5. DEC ’24 · 31,650 KMDRIVE

    Trip to Goa

    1,842 km round trip, eco score 81.

  6. APR ’25 · 38,900 KMDECIDE

    Tyres Replaced

    OEM vs aftermarket compared, community-sourced pick.

  7. FEB ’26 · 47,210 KMUNDERSTAND

    Battery Changed

    The 82% prediction from the top of this page, caught 3 weeks early.

  8. APR ’26 · 49,050 KMUNDERSTAND

    Check Engine Warning

    Explained in plain language, resolved same week.

  9. JUN ’26 · 51,300 KMDECIDE

    Sold

    Repair-or-sell advisor said sell; timeline exported to the buyer.

Every pillar writes to the same spine. That’s the product.

The screen that explains, not just shows.

P0562 and a dial tell you a fault exists. They do not tell you what it costs you, or when. This is what it looks like when the vehicle explains itself: a forecast you can drag forward, a diagnostic you can watch run, and a sentence at the end of both.

Interrogate it, don’t just glance at it.

Drag the horizon on the left and watch risk build on the car itself: the schematic is the interface, not a decoration. On the right, run the same signal as a live diagnostic, right now.

9:41

Risk forecast

KA 05 MJ 4821
82%

Chance the battery fails within 21 days

Carvia’s read

Replace before your next long trip.

hollow, dashed = not observable over OBD-II

Systems · at this horizon

Date
This week
Action
Replace before your upcoming trip
Impact
Cost: ~₹6,500 parts + labour
9:41

Diagnostic

KA 05 MJ 4821
0%

Carvia’s read

Replace the battery before your trip.

82% chance of failure within three weeks.

Cooling loopNominal, last checked 2h ago

Live telemetry

live

BAT

11.94V

IAT

41°C

RPM

812

  • 82%

    Battery fail risk

    at the 21-day horizon, the hero's own claim, now something you can drag

  • −0.48V

    Resting voltage · 11 days

    the trend the forecast is built on

  • 5/ 8

    Signals live over OBD-II

    3 more not yet exposed, see the constraint section below

How do you show 82% without faking precision?

  • A

    A bare number

    reads as fake precision

    “82%” standing alone claims more certainty than any model earns. A naked probability invites either blind trust or instant dismissal.

  • B

    A gauge

    reads as drama

    A needle sweeping into a red zone says panic. It performs urgency instead of communicating one, and trains owners to ignore it.

  • C

    Number + sentence + trend

    the claim arrives with its evidence

    Carvia pairs the figure with the plain-language sentence and the 11-day trend that produced it. You see the reasoning, not just the verdict.

And when a sentence isn’t enough, an enthusiast can open the full correlation model behind it. The same beginner ↔ enthusiast toggle from earlier decides how much of the math you meet first.

Describe the noise. Show it the part.

The moment Carvia stops being a dashboard and becomes a mechanic you talk to: describe what you hear, show it what you see, and get an answer that doesn't need a glossary.

Voice is the right default for this persona. A first-time owner can’t read a fault code, but they can describe a rattle from the front left over bumps, and that sentence carries more diagnostic signal than any symptom picker I could design. So the mechanic listens first, and only then asks for eyes on the part.

The same conversation scales in both directions: from “what’s this orange light?” all the way to walking into a workshop already knowing the fair price range, the part options, and the questions worth asking. One interface for the entire anxiety spectrum of ownership.

  1. 01Listena sound described beats a code read
  2. 02Correlatethe car’s own data joins the conversation
  3. 03Lookthe camera confirms the suspect part
  4. 04Explainplain language, real costs

A live voice demo is its own build, Vapi territory. This one runs from a script.

Simulated, scripted demo. The owner says: There's a rattling noise from the front left when I go over bumps. Carvia cross-references four signals, then asks: Can you point the camera at the front-left wheel well? The camera view identifies the stabilizer-link outer bushing, and Carvia explains: Almost certainly a worn stabilizer-link bushing. Not urgent, but it'll get louder. Fixing it is cheap. Replacing the full link is ₹1,400-2,100 plus about an hour of labour.

An independent second opinion, from the car itself.

Not every service center recommends work only when it's actually needed. Carvia cross-references your vehicle's manufacturer maintenance schedule against your own records, odometer, invoices, service history, so you have an independent basis to question, or confirm, what you're told.

The trust gap doesn’t care whether you’re a beginner or an enthusiast. A first-time owner can’t tell a fair recommendation from an upsell; an enthusiast can, but still can’t prove it in the moment. Carvia gives both the same independent footing, two sources they already own, reconciled before they reply.

Information is where most of this category stops. Carvia continues into judgment: not what the numbers are, but what they mean for the decision in front of you.

Service recommendation, cross-referenced

The service center recommends a brake-fluid flush for ₹1,800. Carvia checks it against the manufacturer schedule, brake fluid due at 40,000 km or 3 years, and your own records: odometer 22,400 km from OBD, last flush 14 months ago from an invoice. Carvia’s read: not yet; worth asking why it was suggested now.

Your service center recommends

Brake-fluid flush₹1,800

Carvia checks this against

cross-referenced
  • Manufacturer schedulechecking…
  • Your recordschecking…

Carvia’s read

Not yet.

Your manufacturer’s schedule says 40,000 km or 3 years. You’re at 22,400 km and 14 months. Worth asking why it was suggested now.

fluid condition itself isn’t OBD-readable, this check reconciles the schedule with your records.

What the port can actually see.

OBD-II is a diagnostic connector, not a full sensor array. Here is what it exposes, what it does not, and where every number on this page comes from.

Reliably available over consumer OBD-II today

  • Battery voltage & health
  • Starter current
  • Temperature correlation
  • Driving-behavior patterns
  • Fault codes & freeze frames

The hero moment is built entirely on real, available signal.

Not reliably exposed today

  • Suspension wear
  • Brake wear
  • Drivetrain condition

Future direction, not hand-waved

External sensors

Aftermarket accelerometers and pad-thickness sensors where the bus is silent.

Service-record inference

Mileage and maintenance history as a proxy for wear the port can’t report.

Manufacturer API partnerships

The long-term path, OEM telematics expose what OBD-II never will.

The gap is flagged as a known constraint and a future direction, not glossed over.

Grayscale, with three earned exceptions.

Hierarchy is weight, scale, and space, not color. Green, amber, and red exist only to mark real component condition, and only at the smallest scale it can be read at, a dot, a node, a bar fill. Dark, consumer-grade, and deliberately nothing like an enterprise B2B system.

Spec sheet · Status system

Palette

  • Page

    #0B0B0C

    The ground. Everything sits on it.

  • Surface

    #131519

    Cards on the page. The phone screen itself is Page — the app is as black as the site.

  • Raised

    #1A1D22

    Panels nested inside another surface.

  • Floating

    #22262C

    Sheets, callouts, the tab bar: things that sit over content.

  • Platinum

    #E5E5E5

    Type and interactive surfaces. Light does the work color used to.

  • White

    #F7F7F7

    Emphasis: the brightest note, for headlines and what matters most.

  • Green

    #3F9E5A

    Nominal: a dot, a node marker, a bar fill. Never anything larger.

  • Amber

    #D99022

    Watch: a tradeoff worth naming, not yet urgent.

  • Red

    #E5484D

    Failing: earned by the moment, never decorative.

Usage discipline

  • elevation

    Surfaces lift by FILL, never by outline. Each step is light enough above the one beneath it to be visible on its own, finished with a 1px inset top highlight so it reads as lit from above. A nested panel is always lighter than its parent, a card darker than its own container is a bug, not a style.

  • emphasis

    Hierarchy is weight, scale, and space. Never hue. Interactive surfaces and emphasis are simply lighter; nothing borrows a color to stand out.

  • geometry

    Component status lives on the car itself, a top-down blueprint and an under-hood cutaway with nodes placed where the part physically sits. Diagnosis is spatial, not a list with icons.

  • status

    Green, amber, and red exist only to mark real component condition, and only at the smallest scale it can be read at: a dot, a blueprint node, a thin meter fill. Never a background, a heading, a button, or a border. A screen should be about 98% grayscale.

  • unmeasured

    What the port can't see gets its own mark: a hollow dashed ring, not a fourth color. Absence of data is a state, not a guess dressed up as one.

  • data

    Data-viz stays grayscale. Charts read as neutral tone and shape, a trend line is never a color.

Color should read as meaningful, not decorative. If in doubt, cut the color.

Type stack

  • Bricolage Grotesque

    Display: characterful, a little mechanical, confident.

    This is what it means.

  • Inter

    Body: invisible on purpose.

    There’s an 82% chance it fails within three weeks.

  • JetBrains Mono

    Data: the voice of the machine.

    P0562, SYSTEM VOLTAGE LOW

Motion principles

  1. Motion supports reading the AI’s reasoning, it never competes with it.
  2. Nothing animates without a comprehension reason.
  3. Reduced-motion is a first-class path, not an afterthought.

MetaThe light-card pattern (plain language on platinum, light = understanding) is this system’s central idea. It’s also the pattern this page uses every time the AI speaks. The page you’re reading is the design-system demo.

Now take the wheel.

The same argument as a thirty-second game. City traffic, the highway to Goa, and what happens when the check-engine light strikes mid-trip.

CARVIACITY · DEMO
← → OR TAP · DODGE TRAFFIC

That’s Carvia’s whole pitch as an arcade loop: the city throws noise at you, the highway throws the check-engine light, and the moment it matters it’s explained in plain language, then you just drive on to Goa. Your tour guide is waiting on the road below.

A car that explains itself is a different product than a car that reports itself.

Three calls I made

  • How does one interface serve a beginner and an enthusiast?

    Always-expandable, beginner-first.

    Modes fork the mental model and make depth a wall. One screen speaks plain language by default, and the full working is one tap away. I prototyped a Beginner/Enthusiast toggle and killed it.

    Still to validate: whether adaptive-over-time ever earns its complexity

  • How do you show an 82% failure probability without faking precision?

    Number, sentence, and trend together.

    A bare number over-claims certainty. A gauge performs urgency. Narrative alone under-informs. Pairing the figure with the sentence and the 11-day trend that produced it lets each part hold the others in check.

    Still to validate: whether owners read the trend, or only the sentence

  • What is the visual language for the Timeline?

    A health record with hard edges.

    Apple Health chronology is the base, but a vehicle's life has harder edges than wellness data: warnings, repairs, a sale at the end. Alert nodes carry those edges without turning every oil change into an event.

    Still to validate: whether a sale or major repair needs a treatment beyond a node

And three I killed

  • A community Q&A feed

    Shipped instead: the AI mechanic answering directly

  • Beginner/Enthusiast mode toggle

    Shipped instead: progressive disclosure, depth as a pull

  • A parts and service marketplace

    Shipped instead: neutral recommendations, no transaction

If this went further

  • OBD data-partner research
  • Owner testing, beginner through enthusiast
  • A working Timeline prototype