Skip to content
ayan://bhardwaj (home)
Built at
AITX & MLOps World
Where
Austin Tech Week
When
Oct 2026

Runway

Know the day your startup's cash runs out, and find the spend pulling it closer.

Runway reads a bank feed read-only, works out burn, runway and the zero-cash date with plain arithmetic, and finds wasted spend. Agents draft the fixes. A person approves every one.

Open Runway (opens in a new tab)Read the code on GitHub (opens in a new tab)
Flight
Taco Robotics, the synthetic demo company
Runway
12.2 months
Cash out
November 3, 2027
Waste
$2,570 a month
Status
On time: no fixes applied

Taco Robotics, Inc. is Runway's synthetic demo company: 967 transactions, $952,350 in cash and $77,791 a month of net burn. Built at the AITX & MLOps World Community Hackathon, Austin Tech Week, October 2026, for the Rho Fintech and Antler Agents tracks. Now an independent early-access product, not affiliated with Rho.

Cancel the waste and watch the date move

These are the six findings Runway pulled from the demo company's bank feed alone, with no receipts and no integrations: $2,570 a month. Flip them and the board re-flaps.

Cash runs out

November 3, 2027

Runway
12.2 months
Saved
$0/mo
Gained
0 days
Refunds
$0

Runway is cash divided by net burn minus what you stop paying. A refund adds to cash. Same numbers as the demo dashboard, worked out here in your browser.

Every month, one pair of lights goes out

The product's own landing page counts a board down from 12 to 0. Here it is as a runway: scroll and the year plays forward. The far end is November 3, 2027.

12.2 months of cash leftas of Oct 2026

  1. 12Oct 2026lit
  2. 11Nov 2026lit
  3. 10Dec 2026lit
  4. 9Jan 2027lit
  5. 8Feb 2027lit
  6. 7Mar 2027lit
  7. 6Apr 2027lit
  8. 5May 2027lit
  9. 4Jun 2027lit
  10. 3Jul 2027lit
  11. 2Aug 2027lit
  12. 1Sep 2027lit
  13. 0Oct 2027lit
  14. 0.0Nov 3, 2027zero cash

A range to land in, not a single guess

One number for runway is a guess. Runway runs 1,000 Monte Carlo paths on the last six months of burn and shows where 80% of them touch down, next to what happens if you act and if revenue falls 30%.

Cash forecast for the demo company, drawn as a glide slopeCash starts at $952,350 and falls to zero. At today's burn it touches down at 12.2 months, November 3, 2027. The P10 to P90 band touches down between 11.6 and 13.0 months. If every fix is made it lands at 12.7 months. If revenue falls 30% it lands at 11.0 months.Revenue −30%11.0 moToday's burn12.2 mo · Nov 3, 2027If you act12.7 moP10–P9011.6 to 13.0 moOct 2026Dec 2026Feb 2027Apr 2027Jun 2027Aug 2027Oct 2027Dec 2027$0$500k$1MCash forecast for the demo company, drawn as a glide slopeCash starts at $952,350 and falls to zero. At today's burn it touches down at 12.2 months, November 3, 2027. The P10 to P90 band touches down between 11.6 and 13.0 months. If every fix is made it lands at 12.7 months. If revenue falls 30% it lands at 11.0 months.Oct 2026Feb 2027Jun 2027Oct 2027$0$500k$1M
Drawn from the demo's published range: 1,000 simulated paths of the last six months' burn trend plus noise, summarised as P10 to P90. The thin lines are 36 illustrative paths, not the model's output.

Agents draft. A person clears every one.

Autopilot runs a fixed plan, chosen before it reads any bank text, so a vendor name cannot steer it. Its memo is rejected unless every dollar figure matches the analysis, and the feed is read-only: it can never move money.

Autopilot · monthly reviewcomplete

SYNC       ok 967 transactions, read-only feed
ANALYZE    ok burn, runway, 4 detectors ~8 ms
FORECAST   ok 1,000 Monte Carlo paths ~10 ms
PLAN       ok steps fixed before any bank text is read
MEMO       ok every $/month checked against the math: pass
FILE_TASKS ok 6 drafts sent to the inbox, 0 sent out
HOLD       awaiting a person for every draft

Inbox, by dollar impact0 of 6 cleared

  • Downgrade request to Datadog$1,200/mo · Hold short: awaiting approval
  • Cancellation email to HubSpot$800/mo · Hold short: awaiting approval
  • Jira ticket: cancel three seats$423/mo · Hold short: awaiting approval
  • Refund request to Canva$119.99 back · Hold short: awaiting approval
  • Cancellation email to Webex$90/mo · Hold short: awaiting approval
  • Cancellation email to Confluence$57/mo · Hold short: awaiting approval

A replay for this page. Clearing here sends nothing. In Runway a drafted email or ticket goes nowhere until a person approves it, and savings only count once the next bill confirms the charge stopped.

The real thing, on the demo company

Screens from the live site. The recording shows findings being ticked on the dashboard and the runway updating as they go.

Ticking findings on the dashboard. No sound.

Runway dashboard for the demo company: 12.2 months of runway, a range of 11.6 to 13.0, revenue down 30% at 11.0 and fix-all-waste at plus 0.4, with cash and burn cards and this week's alerts.
Waste found, $2,570 a month: six findings with checkboxes and links to the transactions behind each, next to the Copilot panel's suggested questions.
Cash forecast chart with a mint P10 to P90 band, a dashed if-you-act line and an orange revenue-falls-30% line, above the default-alive planner.
Runway's landing page: the headline Know the day your cash runs out beside a runway check card showing 12 months and Nov 3, 2027.
Landing page section What it found: a grid of six findings such as Datadog bill up 50%, $1,200 a month.
Landing page countdown: thirteen navy tiles from 12 down to 0, captioned every month the board counts down one.
Runway's landing page on a phone.

Rules first, models last

Everything that decides a number is ordinary code with tests. Models only name the rare vendor the rules miss, and draft text a person reads.

  1. Bank feed, read-only

    Paged, retried and checked with zod. The bank API it was built for cannot move money.

  2. Vendors by rules

    A vendor catalog and alias rules name about 97% of vendors with zero model calls. Unknowns go once to a small model, cached, forced into a closed list.

  3. Plain arithmetic

    Burn, runway and the zero-cash date in whole cents, from tested TypeScript. The model never does the math.

  4. Four detectors

    Overlapping tools, price increases, subscriptions on idle cards and double charges, from bank data alone.

  5. Agents that draft

    Copilot plans with the same pure tools. Autopilot runs a fixed plan. A database trigger rejects any agent action that is not a draft.

Detector eval on synthetic data
DetectorPrecisionRecall
Overlapping tools100%100%
Price increases100%100%
Idle-card subscriptions100%100%
Double charges100%100%

Synthetic data only, not real ledgers: 5 seeds at 2 noise levels, with planted waste as ground truth. CI fails the build below 95%. A separately written held-out set found real defects (usage growth read as a price increase, at 43% precision), which were fixed.

Speed
~950 transactions in 7–10 ms, Monte Carlo ~10 ms more
Tests
396 unit tests, axe-core clean on the agent tabs
Data
Supabase Postgres, row-level security on every table, append-only audit log
Free tool
A CSV runway grader (opens in a new tab) that runs entirely in the browser; nothing is uploaded
Stack
Next.js, TypeScript, Tailwind, Vercel AI SDK, Supabase Postgres, Recharts, zod, vitest, Playwright

Cleared for takeoff

Try the live demo (opens in a new tab)Code on GitHub (opens in a new tab)