Skip to content
Sandworld
Book a pilot

Investors · Pre-seed 2026

The practice worlds AI agents need before they touch real apps.

Sandworld builds verifiable digital twins of the apps people use, for AI labs to train and evaluate agents. The product is live. We're raising a $500k pre-seed to land the first paid pilots.

Download the deckTalk to the founderTry the product
Raising
$500k pre-seed
Runway
18 months
Stage
Live product, pre-revenue
Use
First paid pilots

7

worlds live, mobile and web

117

tasks, each with a verifier

117/117

tasks graded A by verifier QA

3

open models benchmarked; best 9%

1

founder, hiring two

01The problem

Agents learn from their graders. The graders are often wrong.

  • 15.3% of FAIL verdicts were wrong in an audit of five public agent benchmarks. arXiv
  • Robustness to reward hacking is the No. 1 buying criterion for RL environments. Epoch AI
  • Real apps can't be used for this: they can't be reset to an exact state, change without notice and have real side effects.

02The product

A twin, a task, and a verifier that can't be talked into a pass.

  • 6 app twins and 1 cross-app world: grocery, payments, airline, invoicing, issue tracking, email and calendar
  • Every verifier attacked before it ships: do-nothing, random, early-stop and collateral-damage agents must fail
  • Tool use and computer use on the same tasks, graded by the same verifier
  • Fits labs' pipelines: REST, MCP, Harbor task packages, ATIF trajectories, a local server for RL at volume
  • Two sides on one platform: a lab workspace and an expert workbench
sandworld.dev/play/worksuite
Worksuite, a cross-app world of email, issue tracker and billing

Real screens from the live twins. Every one is playable at sandworld.dev.

03Why now

Labs spend billions on data. Environments are the new line item.

$2B

Mercor gross run rate, June 2026

Source

$1B

micro1 run rate, October 2026

Source

$3.5B

Snorkel AI valuation after a $350M raise, September 2026

Source

$200–2,000

paid per RL task

Source

~$20k

per cloned website; labs bought hundreds

Source

4–5×

premium for exclusive environments

Source

  • Scale AI launched RL Environments in February 2026.

    Source

  • Mercor bought Deeptune, which recreated hundreds of enterprise apps.

    Source

  • HUD raised a $16M Series A for RL environments.

    Source

  • Halluminate raised a $30M Series A; works with 4 of the 5 top US labs.

    Source

Reported figures from the sources linked; not Sandworld's numbers.

04Where we win

Gaps nobody fills well.

  1. 01

    Verifier assurance as a deliverable

    Few vendors publish how often their graders pass the wrong answer. Sandworld attacks every verifier and publishes the report with every task.

  2. 02

    Results anyone can audit

    Every published run ships every episode, and every failure is categorised. One click turns those failures into expert recovery data.

  3. 03

    Cross-app and computer use on identical tasks

    Facts carried from email to issue tracker to invoice; tools and screens graded by the same verifier, so the cost of acting through a screen is measurable.

  4. 04

    Mobile first

    Research found no vendor selling deterministic mobile app replicas. Three of the seven worlds are mobile apps, with an Android twin next.

Side-by-side vendor comparisons

05Business model

From a free sandbox to licences and expert data.

  1. 01

    Sandbox Free

    The public API and every world, rate-limited

  2. 02

    Pilot from $25k

    One custom twin, started from the app's OpenAPI document, plus 200 verified tasks

  3. 03

    Licence from $20k / twin / yr

    Non-exclusive; exclusive builds at 4–5×

  4. 04

    Expert data per accepted task

    Demonstrations, preferences, rubric grades, red teaming

Indicative pricing, to be validated in the first pilots.

06Traction

Pre-revenue. The product is built, live and public.

No customers yet, and no numbers on this page are projections. Everything below can be checked on the live site.

  • A public sandbox API anyone can run an agent against: curl https://sandworld.dev/api/v1/environments
  • 3 published model runs with every episode public on the Arena; every pass and near-miss checked by hand
  • Verifier QA for every task, published on each world's page

Next 90 days

  1. 0130 conversations with lab and agent-startup eval teams; 3 design partners; the first paid pilot
  2. 02Frontier-model results (Claude, GPT, Gemini) on every task, published with their episodes
  3. 03A sales co-founder and an ML engineer on board
  4. 0410 worlds, including the first packaged Android twin and a second cross-app world

Published runs · pass rate on 117 tasks

Revanth Rajeev, founder of Sandworld

07Team

Revanth Rajeev built all of it.

  • Builds and ships apps, including Android apps on Google Play
  • Built Sandworld end to end: 7 worlds, 117 verified tasks, the any-model eval harness, the expert platform and the live site
  • Ran and audited the first model benchmarks himself

Hiring next

  • A sales co-founder who has sold training data or evaluations to AI labs
  • An ML engineer with RL and evaluation experience

08The ask

$500k pre-seed, 18 months, to the first paid pilots and a seed round.

Milestones

  • 3 paid pilots and the first $250k in annual revenue
  • 10 worlds, including a packaged Android twin
  • 100 vetted experts

Team

55%

A sales co-founder and an ML engineer, based in India

Expert network

15%

Vetting, paid assessments and the first payouts before client payments arrive

Compute and evals

15%

Frontier-model benchmarks, always-on hosting and client evaluations

Security and company

15%

SOC 2 readiness, a penetration test, and a Delaware company with the IP assigned

09Data room

Everything in one place.

Or email revanth@sandworld.dev

Talk to Revanth

Leave a note and a way to reach you; every message gets a reply.

What can we help with?