🍔 TripleDB

Every restaurant from Diners, Drive-Ins and Dives — structured, searchable, and mapped.

🌐 tripledb.net · 📺 805 Episodes · 🍔 1,102 Restaurants · 🍽️ 2,286 Dishes · 📍 1,006 Mapped · 💰 $0 Cost

TripleDB processes 805 YouTube videos from Guy Fieri's "Diners, Drive-Ins and Dives" (DDD) into a structured Firestore database of restaurants, dishes, ingredients, and iconic Guy Fieri moments. The name is a triple play: Triple D (the show's nickname) + DB (database). It's live at tripledb.net — a Flutter Web app where you can search by anything, find a diner near you, and watch the exact moment Guy walks in.

YouTube Playlist (805 videos)
    | yt-dlp (local, tmux)          --remote-components ejs:github
    v                                --cookies-from-browser chrome
MP3 Audio
    | faster-whisper large-v3       LD_LIBRARY_PATH=/usr/local/lib/
    v (local CUDA, tmux session)    ollama/cuda_v12:$LD_LIBRARY_PATH
Timestamped Transcripts
    | Gemini 2.5 Flash API          Free tier. 1M context. No chunking.
    v
Extracted Restaurant JSON
    | Gemini 2.5 Flash API          Dedup by name+city. Merge dishes.
    v
Normalized JSONL (1,102 restaurants)
    | Nominatim (OpenStreetMap)      1 req/sec. geocode_cache.json.
    v
Geocoded Data (1,006 with coords)
    | Google Places API (New)        Text Search -> Place Details. >=0.70 match.
    v
Enriched Data (582 verified)
    | Firebase Admin SDK             Merge updates. Never overwrite originals.
    v
Cloud Firestore
    | Flutter Web
    v
tripledb.net

What you can do on tripledb.net:

• Search by anything — dish name, cuisine type, city, chef, ingredients ("ground beef" → hamburgers, meatloaf, etc.)
• Find a diner nearby — location-aware, consent-gated, proximity-sorted
• Watch the exact moment — deep link to the YouTube timestamp where Guy walks in
• See Google-verified ratings — star ratings and open/closed status from Google Places
• Discover 150+ trivia facts — rotating facts about the show with no-repeat system
• Browse name changes — restaurants that changed names get "Now known as" badges

Phase 10 complete. Three tracks across 3 iterations: Retrospective (v10.44), Technology Radar (v10.45), UAT Handoff + IAO Template (v10.46). UAT artifacts ready for Gemini CLI. IAO Project Template ready for any new TachTech project.

TripleDB is built using Iterative Agentic Orchestration (IAO) — a development methodology where LLM agents execute project phases autonomously while humans review versioned artifacts between iterations. Every iteration produces five artifacts: design, plan, build, report, and changelog. The report informs the next plan. The methodology itself evolves alongside the project.

IAO crystallized through 46 iterations into the Nine Pillars: Artifact Loop, Agentic Orchestration, Zero-Intervention Target, Pre-Flight Verification, Self-Healing Execution, Progressive Batching, Post-Flight Functional Testing, Mobile-First Flutter + Firebase (Zero-Cost by Design), and Continuous Improvement.

Bulk pipeline processing and UAT auto-chain execution run in tmux sessions for crash resilience and unattended operation. The Phase 5 production run (805 videos, 14 hours) and all Phase 10 UAT execution are tmux-driven. 46 iterations produced 19 cataloged failure modes, a 14-item plan quality checklist, and 10 lessons learned. See docs/archive/ddd-retrospective-v10.44.md.

See docs/ddd-design-v10.46.md for the full Nine Pillars framework. See docs/iao-template-design-v0.1.md for the project-agnostic IAO methodology template.

Evaluated 13 tools across 5 axes (architecture fit, cost, token efficiency, integration, TachTech breadth). Full analysis in docs/ddd-radar-v10.45.md.

~/dev/projects/tripledb/
├── CLAUDE.md                    ← Version lock (Dev)
├── README.md                    ← Public, changelog (NEVER truncate)
├── .gitignore
├── docs/                        ← Current iteration only
│   ├── ddd-design-v{P}.{I}.md
│   ├── ddd-plan-v{P}.{I}.md
│   ├── ddd-build-v{P}.{I}.md
│   ├── ddd-report-v{P}.{I}.md
│   ├── screenshots/
│   └── archive/                 ← ALL previous iterations
├── pipeline/
│   ├── GEMINI.md                ← Version lock (UAT/legacy)
│   ├── scripts/
│   ├── config/
│   └── data/                    ← gitignored
├── app/
│   ├── pubspec.yaml
│   ├── firestore.rules
│   ├── firebase.json
│   ├── lib/
│   ├── web/
│   └── build/web/               ← gitignored

v0.7 (Phase 0 — Setup)

• Monorepo scaffolded with pipeline/, app/, and docs/ directories.
• 805 YouTube playlist URLs collected. IAO methodology established.
• Learned: fish shell has no heredocs — use printf or nano.

v1.8 (Phase 1 — Discovery, Attempt 1)

• Failure: Nemotron 3 Super (120B, 42GB) could not run on 8GB VRAM RTX 2080 SUPER. Model spilled to CPU RAM, causing indefinite timeout loops during context pre-filling.

v1.9 (Phase 1 — Discovery, Attempt 2)

• Pivot: Swapped Nemotron for qwen3.5:9b. Added transcript chunking for 8K context window.
• Extraction still too slow — 5-10 minutes per video with frequent timeouts.

v1.10 (Phase 1 — Discovery, Attempt 3)

• Pivot: Shifted extraction to Gemini 2.5 Flash API (free tier, 1M token context).
• 93% extraction success rate. 186 restaurants, 290 dishes from 30 videos.

v2.11 (Phase 2 — Calibration)

• 60-video dataset: 422 unique restaurants, 624 dishes. Gemini Flash handles 200K-char transcripts.
• CUDA LD_LIBRARY_PATH must be set at shell level, not Python level.
• Marathon extraction timeout increased to 300 seconds.

v3.12 (Phase 3 — Stress Test)

• Zero interventions achieved for the first time. Autonomous batch healing.
• 511 restaurants, 896 dishes, 98 dedup merges. 4-hour marathon logged as edge case.

v4.13 (Phase 4 — Validation)

• 608 restaurants, 1,015 dishes, 162 dedup merges. Zero interventions.
• Extraction and normalization prompts locked for production. Group B green-lit.

v5.14 (Phase 5 — Production Setup)

• Fixed null-name restaurant merging bug (14 records collapsed into one entity).
• Built Group B runner: group_b_runner.sh + checkpoint_report.py.
• Hang detection: 600-second signal.alarm around model.transcribe.
• IAO Eight Pillars documented in README.

v5.15 (Phase 5 — Production Run)

• 14-hour unattended run via tmux. 778 downloaded, 774 transcribed, 773 extracted.
• 4 videos exceeded 600s timeout — skipped. Resume support confirmed.

v6.26 (Phase 6 — Firestore Load)

• 1,102 unique restaurants loaded to Firestore. State inference: UNKNOWN reduced from 126 to 33.
• App wired to Firestore. Search, trivia, and list views functional.

v6.27 (Phase 6 — Geolocation Fix)

• Fixed geolocation prompt but broke Firestore with temporary bypass. Reverted in v6.28.
• Downgraded geolocator to 10.x for Flutter compatibility.

v6.28 (Phase 6 — Geocoding)

• 916/1,102 restaurants geocoded via Nominatim at 1 req/sec. Map showing pins across the US.

v8.17–v8.25 (Phase 8 — Flutter App)

• Two-pass app build. Pass 1: scaffold + core features. Pass 2: design tokens, component patterns.
• DDD Red (#DD3333), Orange (#DA7E12), Outfit + Inter fonts. Lighthouse A11y 92, SEO 100.
• 3-tab bottom nav (Map/List/Explore), restaurant detail with YouTube deep links, dark mode.

v6.29 (Phase 6 — Polish)

• Trivia state count fixed (excludes UNKNOWN → shows 62 states).
• Map pin clustering via flutter_map_marker_cluster. Orange clusters with counts.
• Explore page: multi-cuisine string splitting for accurate counts.

v7.30 (Phase 7 — Enrichment Discovery)

• Google Places API (New) enrichment pipeline built. 50-restaurant discovery batch.
• 66.7% match rate. 90% of matches rated 4.0+. 4 coordinate backfills. API cost: $0.

v7.31 (Phase 7 — Enrichment Production)

• Full run on 1,102 restaurants. 625 enriched at 55.9% match. 32 permanently closed identified.
• App UI: rating badges, open/closed status, website and Google Maps links.
• 1 intervention (API key not set in environment).

v7.32 (Phase 7 — Enrichment Refinement)

• Refined search on 462 no-match restaurants (4 query passes). 83 recovered, 18% recovery rate.
• Gemini Flash LLM verification: 112 confirmed, 126 false positives removed, 26 uncertain.
• Net enrichment: 582 verified. Geocoding coverage: 91.3%.

v7.33 (Phase 7 — AKA Names + Closed UX)

• google_current_name backfilled for all enriched restaurants. 283 genuine name changes.
• Grey map pins for closed restaurants. "Show closed" filter toggle. Closed banners on detail pages.
• "Now known as" display for renamed restaurants. Step-level checkpointing introduced.

v7.34 (Phase 7 — Cookies + Analytics + Polish)

• Cookie consent: accept/decline/customize with 3 categories. 365-day cookie, SameSite=Lax.
• Firebase Analytics with consent mode v2. Events gated by user consent.
• Name-change threshold tightened 0.95 → 0.90 (86 reclassified, 279 genuine remain).
• 26 UNCERTAIN records resolved (15 kept, 11 removed).

v9.35 (Phase 9 — App Optimization)

• Executor change: Gemini CLI → Claude Code.
• Riverpod 2.x → 3.x migration. Geolocator 10.x → 14.x upgrade.
• Trivia expanded from ~9 facts to 70-80+ with no-repeat shuffle system.
• "Nearby Restaurants": 15 results with distance in miles, "Show all nearby" → 50.

v9.36 (Phase 9 — Production Fix)

• CRITICAL: Fixed white screen crash on tripledb.net caused by eager provider initialization in main() before runApp(). CookieConsentService constructor accessed dart:html document.cookie during Riverpod 3 provider graph construction, killing the app before Flutter could render.
• Fix: Removed ProviderContainer from main(), switched to standard ProviderScope with lazy providers. Made CookieConsentService lazy-initialized with try-catch fallback. Moved analytics initialization to widget tree post-frame callback.
• Restored complete README changelog (v0.7–v9.36) — all 21 entries preserved.

v9.37 (Phase 9 — Post-Flight Protocol + Location Consent)

• Post-Flight Protocol (Pillar 9): Automated runtime verification using Puppeteer headless browser. Serves release build locally, navigates app, checks rendering, reads console for errors. 6 gates: bootstrap, navigation, features, cookies, console, changelog integrity. Prevents white-screen deploys like v9.35. Permanent part of IAO methodology.
• Location on consent: Accepting cookie preferences now triggers browser geolocation permission request. Grants location → populates "Nearby Restaurants" immediately. Decline → no location prompt. Reduces permission fatigue from 2 prompts to 1 flow.
• Changelog gate: Post-flight verifies README changelog entry count ≥ 22. Agent cannot declare iteration complete if changelog has been truncated.

v9.38 (Phase 9 — Cookie Banner Fix + Functional Post-Flight)

• Root cause: Cookie Secure flag silently prevented document.cookie writes on HTTP, and _writeCookie() used ISO 8601 for expires instead of RFC 1123. Additionally, _readCookie() used fragile split('=') with parts.length == 2 check that could fail if cookie values contained = characters.
• Fix: Conditional Secure flag (only on HTTPS), RFC 1123 date format for expires, indexOf-based cookie parsing instead of split, and essential key validation to reject malformed/stale cookies.
• Post-Flight v2: Two-tier system. Tier 1: health gates. Tier 2: iteration-specific Playwright/Puppeteer functional playbook with click-verify-confirm actions. Canvas screenshots replaced with accessibility tree verification and interactive button clicking.
• Cookie banner verified via 6-test playbook: renders, dismisses, persists, validates cookie structure, fresh context, decline path. All 6 tests PASS.

v9.39 (Phase 9 — Nearby Filtering + Location Consent Fix)

• Bug 1 fix: Filtered "Unknown" city/state restaurants from nearby results. Only restaurants with valid, real city and state values appear in "Nearby Restaurants" and proximity-sorted search.
• Bug 2 fix: Deduplicated nearby results by restaurant_id and normalized name. No restaurant appears twice. Belly and Snout (3 entries in JSONL with different IDs) now shows once.
• Bug 3 fix: Accept All cookie consent now correctly triggers browser location permission request before dismissing the banner. Location grant populates "Nearby Restaurants" immediately.
• Design doc: Comprehensive living ADR with Eight Pillars, full environment setup guide, complete project state, and work remaining to MVP.
• Post-flight: 7/7 tests PASS — no Unknown in nearby, no duplicates, Accept All → location, cookie persistence, decline path correct.

v9.40 (Phase 9 — dart:html Migration + Firestore Security Rules — FINAL DEV ITERATION)

• dart:html → package:web: Migrated cookie_consent_service.dart from deprecated dart:html to package:web + dart:js_interop. Eliminates the persistent analyzer deprecation warning (info • 'dart:html' is deprecated) and enables future WASM compilation. All cookie read/write/HTTPS-detection functions verified via 7/7 post-flight playbook.
• Firestore security rules: Created firestore.rules with read-only public access. restaurants and videos collections allow public reads, deny all client writes. Admin SDK writes (pipeline scripts) bypass rules. Default deny on all other collections. firebase.json updated to reference rules file.
• Final dev iteration: Phase 9 complete. All P0/P1 items resolved. flutter analyze shows 0 errors, 0 warnings, 0 infos. App is production-ready for Phase 10 UAT handoff to Gemini CLI.

v9.41 (Phase 9 — Nine Pillars + README Overhaul)

• Nine Pillars of IAO: Methodology evolved from Eight to Nine Pillars. New Pillar 8 (Mobile-First Flutter + Firebase, Zero-Cost by Design) elevated from tech stack choice to architectural principle. Continuous Improvement renumbered to Pillar 9.
• Agent permissions updated: Agents CAN now run flutter build web and firebase deploy. Agents CANNOT git add/commit/push. Kyle commits at phase boundaries.
• README overhaul: Full rewrite with feature badges, ASCII pipeline diagram, layered architecture table, and updated project stats reflecting 41 iterations of development.
• CLAUDE.md template v2: Updated with CAN/CANNOT permissions table.

v9.42 (Phase 9 — Hardening Audit)

• Lighthouse baseline: Accessibility 93, SEO 100, Best Practices 77, FCP 1.5s (local) / 0.8s (prod), CLS 0. Performance score N/A (Flutter canvas rendering prevents Lighthouse FCP detection in headless). Bundle size 2.8MB (tree-shaken icons 99.4% reduction). No source maps in build.
• Error boundary testing: 5/5 PASS — Firestore offline (graceful degradation), location denied (app loads without distance), cookies disabled (app functional), invalid URL (GoRouter handles), empty search (no crash). No white screens. No inline fixes needed.
• Security hardening: 6 security headers deployed to Firebase Hosting — X-Frame-Options: DENY, X-Content-Type-Options: nosniff, Referrer-Policy: strict-origin-when-cross-origin, Permissions-Policy (camera/mic blocked, geolocation self), HSTS 1 year, and CSP with allowlists for Flutter CanvasKit, Firestore, Analytics, map tiles, and fonts. Cache headers: JS/CSS immutable 1-year, index.html no-cache. Firestore rules pen-tested (writes blocked). Dependency audit: no CVEs, 5 major upgrades available (deferred).
• Browser compatibility: Chromium PASS across 4 viewports (375x667, 768x1024, 1024x768, 1920x1080). No horizontal scroll. Keyboard navigation 10/10 focus changes. Firefox: app loads but Flutter runtime Invalid language tag: "undefined" error (upstream Flutter bug, P2). WebKit: skipped (missing system deps on CachyOS).
• Versioned changelog: Introduced ddd-changelog-v{P}.{I}.md as 5th artifact for resilience.

v9.43 (Phase 9 — Package Upgrades + Trivia + Preferences Fix)

• Package upgrades: flutter_map 7.0.2→8.2.2, flutter_map_marker_cluster 1.4.0→8.2.2, go_router 14.8.1→17.1.0, google_fonts 6.3.3→8.0.2. Zero breaking changes. flutter analyze: 0 issues. flutter build web: success.
• Trivia expansion: Increased from ~55 to 151 unique facts. 65 curated + 86 dynamic data-driven facts. Fixed duplicate display with Set-based dedup. Implemented shuffle-based rotation with no-repeat guarantee at cycle boundary.
• Save Preferences → location: Clicking Save Preferences with geolocation enabled now correctly triggers the browser location prompt, matching Accept All behavior. Fixed by reordering: location request fires before modal dismissal.
• Firefox ESR testing: Added Firefox ESR to browser test matrix. Chrome Stable + Firefox ESR are the two supported test targets. Firefox: app loads, renders, no blocking errors. A11y tree not accessible via Puppeteer BiDi (CDP required) — documented as test gap.
• Browser test results: Chrome 9/9 PASS. Firefox 6/6 PASS + 1 SKIP (a11y CDP limitation).

v10.44 (Phase 10 - Retrospective)

• Pillar 9 retrospective: Comprehensive archive review across 43 iterations and 10 phases. Produced intervention timeline (36 total interventions, 89% in Phases 1-2), tool efficacy matrix (14 tools scored), pillar evolution history (6 -> 8 -> 9), plan quality analysis with IAO checklist (14 items), failure mode catalog (19 categorized failures), and top 10 lessons learned distilled into actionable rules.
• Em-dash formatting rule: Established standard - no em-dashes in any artifact. Baked into CLAUDE.md, GEMINI.md templates, and design doc Section 13.
• Phase 10 architecture: Three sequential tracks defined - Retrospective (v10.44), Technology Radar (v10.45), UAT Handoff (v10.46+).

v10.45 (Phase 10 - Technology Radar + README Overhaul)

• Technology Radar: 13 tools scored across 5 axes (architecture fit, cost, token efficiency, integration, TachTech breadth). 5 tools rated Adopt (Claude Skills, Gemini Flash, Puppeteer, Lighthouse, Sonnet 4.6). 5 rated Trial (Ruflo, NemoClaw, Local LLMs, Context7, Gemini Skills). 2 rated Assess. 1 rated Hold (Playwright - dropped).
• README pipeline overhaul: Replaced ASCII-only pipeline diagram with Ruflo-style layered pipeline table showing stage, tool, input, output, and runtime. Added execution model table (Group A/B/UAT with tmux callouts).
• tmux visibility: Added tmux to README tech stack, architecture layer, IAO methodology section, and pipeline table. tmux is a primary component for bulk processing and UAT execution.
• Retrospective integrated: Plan quality checklist (14 items) referenced in CLAUDE.md template. Retrospective highlights added to README. Track C deliverable spec finalized - 4 artifacts: UAT design + plan (Gemini), IAO Project Template design + plan (Claude, any new project).
• Phase 10 Track B complete. Ready for Track C (UAT Handoff + IAO Template).

v10.46 (Phase 10 - Track C Capstone)

• UAT handoff artifacts: Produced ddd-design-uat.md and ddd-plan-uat-v0.1.md for Gemini CLI to replay the full TripleDB pipeline from scratch. Same Firebase project, hosting preview channel, no Firestore writes. Pipeline validated by JSONL diff against dev output.
• IAO Project Template: Produced iao-template-design-v0.1.md and iao-template-plan-v0.1.md - generic Nine Pillars framework for any new TachTech project. Includes plan quality checklist (14 items), failure mode catalog (19 modes), top 10 lessons. Quick start: 5 commands to scaffold.
• Phase 10 complete. Three tracks delivered across 3 iterations: Retrospective (v10.44), Technology Radar (v10.45), UAT Handoff + IAO Template (v10.46).

Kyle Thompson — Solutions Architect @ TachTech Engineering

Built as a passion project for finding the best diners after long motorcycle rides.

Built with IAO - Iterative Agentic Orchestration Phase 10.46 - UAT Handoff + IAO Template across 46 iterations