Sean Spratt — Selected Work
I’m an engineer who ships applied-AI systems end to end: evaluation harnesses, judged generation pipelines, document-scoring services, and the hosted products built on them. That sits on 20 years of production engineering at Microsoft and NCC Group, and in 2026 alone I’ve shipped more than 25 public AI products.
Six projects, each chosen for what it proves
Choir
Evaluation harness & benchmark design · macOS app + CLI
One prompt fanned out to every major LLM at once (OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, local models), with judge-model scoring, latency charts, a divergence heatmap, and saved runs that can be re-judged later without re-running the participants. It also drops in as a skill so other coding agents can delegate work to a specific model.
Choir Reports
Published comparison studies · 23 reports
The benchmark work written up: the same prompt across twenty-plus models, hand-scored and then re-ranked by a blind multi-judge jury, with inter-judge agreement, leak analysis, and the failure modes called out by name. Studies range from a 27-model charisma ladder to how medical advice decays as the asker gets upset.
Scriptorium
Multi-stage agent pipeline with internal judging · 10 stages
A four-sentence brief becomes a finished illustrated audiobook in about half an hour: arc fan-out across premium models, a judging panel that picks and scores candidates at each stage, named-element mixing, style-locked illustration, and multi-voice narration. Seven books have gone through it end to end.
Dexter’s Papers
Messy documents in, scored retrieval out · runs unattended nightly
Every weeknight the day’s arXiv AI papers are ingested, summarized, and scored 0–10, then filed into a filterable map and a per-night calendar with one-line synopses and full breakdowns. It has produced a new episode every night since July 2026 with no hand-holding.
FestivalIMS
Production deployment, not a prototype · hosted for paying customers
Incident management for festivals and regional burns: shift rosters, incident reports, a forensic audit trail, live metrics with a shift Gantt, and an offline Command Center for when there is no wifi. Hosted and operated per event, with full data export and deletion on request.
Tips for Building One-Shots
Working method with AI coding agents · written so others can repeat it
The three-document method I use to turn a product idea into something a coding agent can actually build without session drift or scope creep: what SPEC.md, CLAUDE.md, and FIRST_SESSION.md are each for, why the order matters, and a copy-pasteable scoping template.
The newest work: whole book series, walked
Series Notes
Repeatable multi-agent pipeline with adversarial fact-checking · 11 series, 107 novels, ~3,100 painted frames, all since Oct 6, 2026
One command builds an illustrated reader’s companion to an entire book series: parallel research agents (one per novel, one per screen adaptation), an art-style shoot-out, one writer per page, an adversarial checker that fixes every summary and caption against the text, a cross-page consistency sweep, then film and TV comparison pages, each walk published to its own site. The same pipeline has now run eleven times in two weeks, on series from three novels to thirty-one.
- The Wheel of Time · 15 novels · ~380 frames · vs. the Amazon series
- The Riftwar Cycle · 31 novels · 499 frames
- The Dark Tower · 8 novels · ~340 frames · vs. the 2017 film
- The Chronicles of Amber · 10 novels · ~269 frames
- The Mistborn Saga · 8 novels · ~259 frames
- Solo Leveling · 8 novels · ~297 frames · vs. the anime
- Play to Live · 9 novels · ~257 frames
- The Saga of Seven Suns · 7 novels · 247 frames
- The Silo Saga · 3 novels · 221 frames · vs. Apple’s Silo
- Altered Carbon · 3 novels · 215 frames · vs. the Netflix series
- Elemental Blessings · 5 novels · ~180 frames
The Bookshelf
Finished products, each labeled with what AI made · novels, audiobooks, comics, guides
Everything the Scriptorium pipeline and its cousins have produced, on one shelf: full-length illustrated novels with narrated audiobooks (H. G. Wells’s The Sleeper Awakes, Fool’s Cycle, The Empty Jar), the series walks, comics, picture books, and tutorials. Every book carries a provenance label saying what AI wrote, drew, and voiced.
Everything above is live and public; the full index has the rest, including the less serious ones.