AI chess coaching platform for kids — live at pawnprint.com
| Built: 2025–2026 | Status: Live — Phase 0 tester program |
Pawnprint is an AI chess coaching platform targeting children ages 7–17. It turns a player’s own games into a personalized training plan — blunder puzzles scheduled via spaced repetition, coached in language matched to the learner’s age and vocabulary level.
The core problem it solves: LLM-powered chess tools hallucinate. Language models cannot reliably evaluate chess positions, which makes them dangerous in a learning product for children. Pawnprint’s architecture makes this impossible by design, not by instruction.
Chess improvement tools fall into two categories: generic platforms (Lichess, Chess.com) that serve adult-level content to all ages, and human coaching at $50–150/hour that most families can’t sustain. Neither option adapts to a specific player’s mistakes, age, or vocabulary level.
LLM-based alternatives introduce a new problem: models confidently describe tactics that don’t exist on the board. For a child learning the game, that’s actively harmful.
A three-layer pipeline that enforces strict separation between chess analysis and language generation:
The LLM receives a structured object of pre-computed facts. There is no path in the system for it to perform chess analysis, regardless of what it might otherwise generate.
Personalized blunder puzzles Games imported from Chess.com are analyzed by Stockfish. Each blunder becomes an SM-2 scheduled puzzle tied to the specific position where the player went wrong — not generic tactics, their actual mistakes.
5.4 million puzzle library Training pulls from a 5.4M Lichess puzzle corpus difficulty-matched via Glicko-1 rating, benchmarked against Chess.com global percentiles sampled from 9 community clubs. Players get puzzles at the right difficulty, not arbitrary buckets.
Age-adaptive coaching language Four age brackets (under 10, 10–14, teen, adult). The system tracks each user’s chess vocabulary acquisition — introduced, reinforcing, acquired — across 7 exposures per term. Every coaching response is constructed from what this specific user already knows and what they don’t yet.
RAG-grounded explanations Coaching responses are grounded in 8,543 enriched chunks from classical chess literature via FTS5 BM25 retrieval. The same tactical theme retrieves different passages for different age brackets.
Archetype-aware session planning Players are profiled by learning archetype — attacker, puzzle-solver, competitor — which influences retrieval ranking and session composition.
Pawnprint operates as a commercial online service that knowingly collects personal information from children under 13, putting it in scope for COPPA (16 CFR Part 312, as amended by the 2025 amendments effective June 23, 2025, with compliance deadline April 22, 2026 — passed). Pawnprint launched on the live amended Rule, not the legacy 2013 baseline that most operators in the market still target. The platform operates under a comprehensive end-to-end compliance build with both halves shipped: application infrastructure verified in production plus procedural artifacts audited for cross-document consistency.
Application infrastructure:
Procedural infrastructure:
Verification at closure: end-to-end consent and re-consent flow tested in production against real closed-beta cohort with verified email delivery, append-only audit trail intact across multiple policy versions, and cross-document consistency validated (15-of-15 cross-artifact + 10-of-10 operational probes passing).
For the full case study covering execution discipline, key architectural decisions, lessons captured, and how the engineering patterns transfer to HIPAA / FERPA / GDPR / GLBA / PCI-DSS / regulated-industry compliance more broadly, see: Pawnprint COPPA Compliance Case Study →
| Metric | Value |
|---|---|
| Live URL | pawnprint.com |
| Uptime | 99% |
| Infrastructure cost | $60/month |
| RAG knowledge chunks | 8,543 |
| Puzzles indexed | 5.4 million |
| Age brackets | 4 |
| Phase 0 testers | 7 |
| Tester retention | 100% |
| Billing | Stripe (pre-revenue phase) |
| COPPA compliance posture | Comprehensive end-to-end build under amended 2025 Rule — both halves shipped, audited at closure |
| Layer | Technology |
|---|---|
| Frontend + API | Next.js 14 (App Router), TypeScript, Tailwind |
| Database | Turso (libSQL + FTS5) |
| Chess engine | Stockfish 18 (WASM + native binary) |
| Chess logic | chess.js |
| Coaching LLM | Gemini 2.5 Flash |
| Index-time generation | Claude Haiku 4.5 |
| Spaced repetition | SM-2 |
| Puzzle rating | Glicko-1 |
| Auth | JWT + TOTP 2FA |
| Infrastructure | DigitalOcean, PM2, Cloudflare Tunnel |
| CI/CD | GitHub SSH deploys |
| Compliance | Termly Pro+, Resend (DKIM), Cloudflare Email Routing, append-only audit (3 tables), formal § 312.X self-assessment, eight processor DPAs |
Preventing hallucination architecturally. The standard approach — instruct the model not to make things up — doesn’t hold reliably in a product children depend on for accurate instruction. The solution was to remove the possibility at the system design level, not the prompt level.
Age-bracketed retrieval at query time. The same tactical theme requires genuinely different explanations for a 7-year-old versus an adult. Built a multi-factor retrieval ranking system that produces different top results for the same query depending on the learner’s profile.
Vocabulary scaffolding at scale. Tracking term acquisition state per user across an open-ended chess vocabulary, then dynamically constructing each coaching prompt from that state, required careful schema design and prompt architecture to remain consistent across millions of potential query combinations.
Pawnprint is in active development. Phase 1 (paid launch) planned following Phase 0 tester validation.