Aleph Lab.

Aleph Lab is a San Francisco AI startup in Y Combinator's Fall 2025 batch. Its product, Aleph Kids, teaches English through Minecraft: “Annie,” a voice-enabled AI character, is a co-playable character — she plays alongside children inside the game rather than at them, so they practice speaking a new language while they play.

• Packaged the in-game AI character as a versioned SDK that studios outside the company can build on — a game mode becomes data: 47 in-world event types, a 16-verb action surface, capability negotiation per world, and server-side allowlist enforcement so a mode can never make her do something its world didn't declare — behind 1,301 tests over a hosted multi-tenant service
• Proved the SDK end to end: a game mode authored outside the repository, from the handover docs alone, came back valid, loaded its own content at runtime, and merged as a first-party mode
• Fixed the character picking fights she shouldn't: her threat check keyed on a hostility flag that is behaviorally wrong for provoke-only mobs (endermen, piglins — hit one and the group swarms) and lethal bosses, corrected additively so the change could only ever remove a bad target, never drop her defense against real threats
• Taught her to see modded content — a registry bridge regenerated at every server start, so on a partner studio's mode she learned 44 custom block states and 152 items and names what a kid is holding instead of reading back a raw id
• Took hosted voice end to end, so a builder can hold a spoken conversation with the character from a browser with no game client running
• Shipped a Claude skill that authors and validates a new mode's configuration, so integrating the agent into a mode stopped being bespoke work each time
• Delivered an app-wide React Native redesign from Figma to production — shipped the reorganized screens first so users kept their bearings, then rolled the 33-section design system onto every screen
• Built the product's lifecycle notification system end to end: re-engagement, streak reminders and class-completion alerts, with per-stage copy, back-off so message types cannot pile onto the same family, and a holdout group so the effect could be measured rather than assumed
• Then made it autonomous — it proposes sends, clears its own safety guardrails and delivers unattended, where every send had previously needed a person to approve it
• Widened notification eligibility after validating the targeting against production data, substantially increasing the number of families the system could reach
• Hardened public endpoints that trusted whatever user ID they were handed, adding bearer authentication with no changes required from existing server-to-server callers
• Built an agent-based analysis tool over production data that showed a reported retention gain was a measurement artifact, and surfaced the drivers that actually mattered
• Gave the character a consistent presence in the app — its message-bubble system, a branded crash recovery in English and Korean, and dashboards that tell a failed request apart from an empty one

Specifics about the product's internals and metrics stay with the company — happy to talk through the engineering in more depth over a call.