Can expertise scale safely?
Turn fragmented track evidence into useful coaching while preserving the coach's judgement, voice and customer relationship.
Portfolio · AI-enabled service venture
HiveMind.
Human-led coaching.
HiveMind is the venture; Hive Coach is its product. It turns track-day evidence into a coach-owned debrief: AI accelerates the analysis, while the coach retains judgement, voice and final sign-off.
Private evidence snapshot · repository main at 7e23cfe · reviewed 11 July 2026
Executive view
A service experiment in preserving expert authority.
HiveMind is deliberately outside MiddleLeap's regulated core. Its relevance is the operating principle: AI may increase the reach and consistency of an expert service without becoming the accountable expert.
Core loop live
The product can assemble evidence, draft a grounded debrief and return it to a coach for review.
The coach decides
AI cannot issue final advice; human edit and sign-off remain the customer-facing boundary.
Human-in-the-loop design
The experiment informs service propositions where AI assists perception and synthesis but cannot own the decision.
Wider pilot evidence
Test credentials, noisy trackside inputs, repeat usage and coach validation before expanding the service.
01
Why it exists
Elite coaching insight should not depend on an elite coaching budget.
Karting produces rich evidence—timing sheets, video, photographs, voice notes and weather—but turning it into useful, personal feedback is slow. Hive Coach makes that synthesis repeatable without replacing the coach.
Evidence is scattered
Track-day observations arrive in different formats and often disappear before they become actionable learning.
One coached debrief
Capture once, assemble the session record, draft a grounded debrief and return it to the coach for judgement.
The coach is the product
AI supports perception and synthesis; it cannot send the final advice or claim facts outside the evidence.
02
Product architecture
A traceable path from track evidence to coach-owned advice.
The runtime is deliberately staged. Vision, transcription and synthesis create structured evidence before final prose is produced. Grounding and confidentiality checks sit before the coach review and PDF delivery boundary.
PeopleDriver + coachCapture · context · judgement
EvidenceMedia + timingPhoto · video · voice · weather
ExperienceNext.js PWATrackside capture and review
Stage 1Vision + WhisperServer-side extraction
Stages 2–3Synthesis + proseClaude · coach voice · grounded facts
Human gateCoach edit + sign-offMandatory before delivery
System of recordSupabase + RLSProfiles · evidence · consent
OutputPDF + WhatsAppApproved debrief delivery
03
Loom-informed delivery
Start with the coach's gold standard, then make every AI step answer to it.
HiveMind applies the Loom's evidence, specification and human-authority principles to venture building. This is disciplined use of the method's ideas—not a claim of full regulated-harness adoption.
Observe
Coach workflow, track-day evidence and a gold-standard debrief
Frame
Human-led product principles, consent boundaries and acceptance criteria
Design
Stitch-assisted screens and a capture-to-debrief architecture
Build
Claude Code loops with server-side AI and repository tripwires
Evaluate
Gold-set scoring, RLS checks and mandatory coach sign-off
Technical build recordAI systems · Technology · Quality controls+
04
AI systems
Delivery AI and product AI are different systems.
Claude Code, skills, hooks, reviewers and Stitch MCP shape how the product is built. Claude and Whisper power the server-side customer workflow. The repository contains no Codex build provenance in the reviewed snapshot.
Claude Code
The repository documents autonomous /loop delivery with explicit acceptance checks and stop conditions.
Claude + Whisper
Claude Sonnet handles vision and synthesis, Opus drafts final prose, and OpenAI Whisper transcribes voice notes.
Two focused skills
eval runs the quality harness; new-migration creates database changes with the project conventions.
Three hooks
Pre-write checks block client-side AI and protected-file changes; post-edit checks run targeted validation.
Two bounded reviewers
Grounding and RLS security are reviewed as separate concerns rather than folded into generation.
Stitch MCP
Google Stitch is documented as build-time design tooling. It is not part of the customer-facing runtime.
Server-only AI→Structured evidence→Gold-set evaluation→Coach edit→Human send
05
Technology
A trackside PWA backed by a secure, observable AI pipeline.
TypeScriptNext.js 15 + React 19PWASupabase PostgresSupabase Auth + RLSSupabase StorageAnthropic ClaudeOpenAI WhisperPuppeteer PDFWhatsApp Cloud APIPostHogSentryVitest + PlaywrightpgTAPVercel
06
Quality system
Quality is measured at the advice boundary.
Build integrity
Types, lint, unit tests and the offline evaluation harness run together.
AI containment
Repository checks prevent browser-side model calls and committed secrets; gitleaks adds a separate scan.
RLS integrity
Supabase migrations and pgTAP policies are exercised in a dedicated database workflow.
Grounded output
A gold set scores whether the generated debrief stays faithful, useful and appropriately confidential.
Product confidence
Playwright covers user journeys, AI-tagged scenarios and row-level security boundaries.
Coach authority
The product requires a coach edit and sign-off before a debrief becomes customer-facing advice.
07
Development record
The core loop is live; expansion remains evidence-led.
One-loop proof
Proven
Core debrief
Phase 0 live
Operations spine
Substantial build
Personalisation
Active
Delivery integrations
Human-gated
Pilot expansion
Roadmap
Known boundary
Phase 0 is a working application, but production confidence still depends on real coach inputs, noisy trackside trials, live delivery credentials and wider pilot evidence. Parent consent and coach authority remain human responsibilities, not model decisions.
08
Evidence register
Claims are grounded in a reviewed private build record.
The repository is not publicly accessible. This register names the material reviewed; detailed source access can be considered during appropriate diligence.
R1Repository overview and product principlesReviewed
R2Binding Claude Code conventionsReviewed
R3Technical architectureReviewed
R4Phase 0 build planReviewed
R5Post-Phase-0 roadmapReviewed
R6Claude Code hooks and pluginsReviewed
R7Continuous-integration workflowReviewed
R8Deployment runbookReviewed
What this venture proves
AI can deepen a human service without taking authority away from the expert.
Bring the product, operating-model and governed-AI learning into your service proposition.