🚇 gpu-pilot — project docs

The executable implementation plan for the commutingcost.com redesign pilot: the homepage + the /commute-cost/[metro] template (x53), all four layers end to end — discovery (SSR JSON-LD, sitemap, llms.txt, /methodology, OG cards), motion (the approved Editorial Warm system with per-page transit vehicles), conversion (prefilled CTAs, provenance, email hook), and the data pipeline (GTFS-derived dominant mode per metro). Built from a 13-agent audit of the REAL codebase (115 verified findings, 77 gaps) and hardened by two adversarial critics. 9 stages, 49 files to create + 18 to modify, each stage ending in a verifiable checkpoint. Q1/Q2/Q4 are now closed; Q3/Q5/Q6/Q7 have safe defaults.

The design spec the plan implements. Approach C — balanced pilot slice: take one fully-realized vertical slice (homepage + metro-page template), do all four layers on it end to end, ship it, then fan the pattern out to the remaining pages and to multimodal + losingground. Defines the four success outcomes (get discovered / convert visitors / delight & shares / investor-ready polish), the GEO and AEO artifact set, the motion inventory with its Core Web Vitals guardrails, and the component kit both pilot pages establish.

Newest handoff. The national POI-dest backfill (151,781 dests) is 87% done, ~22h to full, driven by THREE GPU producers write-through to the shared Postgres store: geocoder RTX 5080 (unsharded, owns the heavy giant regions = the bottleneck), gtfs3 RTX 5060 Ti (its shard, slow on heavy dests), and a NEW third producer on tfs's idle RTX 2060 (172.26.1.20) scoped via a new PRECOMPUTE_METROS env to the non-giant metros — it cleared the 25.7K small-metro tail in ~3.5h at ~2/s, then expanded to all 43 non-giant metros (DC's 24.6K-stop region peaks 5.26GB, fits the 6GB card, OOM-tested). Also shipped: mobile UX on multimodal.certihomes.com (tap-to-inspect TouchCard + '?' help tour + filled Train/Bus toggles), all 17 new metros live on BOTH losingground-cities and commutingcost-web rankings, OTP-NE revived as a boot-persistent systemd unit on geocoder (Edison→Penn 70/77min via NJT NEC verified), and gtfs3 made self-healing after a power outage (keepalive fd-inheritance + hung-lock fixes). gpu-pilot main=f2ee10f.

Newest handoff. THE architecture flip: GPU is now an ETL PRODUCER only — production serves disk(L1) → Postgres commute.iso_precompute (L2) → GPU(L3 fallback), rows stamped ENGINE_REV so stale never serves; SERVE_STORE_ONLY=1 removes the GPU from interactive serving entirely. The national backfill went ~18 days → ~8-9: rail-focused nec (78K→5.4K stops, corridor cell 94s→2.8s, Edison→Philly still 75min via NJT→Amtrak Keystone — proven per-leg by trace_trip.py) + priority-2 metros (Baltimore 122×, SD 74×, Philly 18×, DC 11×, LA 2.7×) + a SECOND sharded producer on gtfs3's RTX 5060 Ti (PRECOMPUTE_SHARD=k/m, shared PG store). Also: first-visit warming toast on multimodal.certihomes.com, fare-gap audit+seed (20 metros curated, 8 real $0-placeholder rows fixed, 14 majors seeded), and the repo-custody sweep — every deployed app now on tlcengine GitHub (FE was Bitbucket, adapter/crm-admin had NO git; wildcard TLS key kept out of every pushed tree). gpu-pilot main=5622f9b.

Newest handoff. THE Edison→Philly root cause was a golden-source SHORT-CALENDAR bug: 143/935 agency feeds (NJT NEC mdb-509 +142) define trips' service ONLY in calendar_dates.txt over a short window, absent from calendar.txt → ZERO service on the pinned 20260415. The golden fixer now infers each orphan's weekly pattern and extends it 2020-2050; re-fixed the whole corpus in-place + transit-rebuilt 35 regions. Then, per Krishna, raised the nec corridor region to priority 3 so it serves the WHOLE NYC↔Philly corridor (Manhattan→Philly 99min, Philly→NYC 96min — both were unreachable). Also shipped: static OSRM water masks, Québec City SEO, res-10 dense-core refine, the Zillow B2B pitch + light Certihomes deck, /map auto-paint, and Kenosha/Gary/Vancouver-WA multi-state fixes. gpu-pilot main=637bffe.

The B2B pitch for licensing the commute-cost engine/data to real-estate portals — Zillow first (their 2024 commute filter is time-only and BuyAbility ignores transportation), then Redfin/Rocket, Realtor.com, CoStar (Apartments/Homes), and relocation firms. Product menu: nightly per-listing enrichment feed (the Walk Score pattern), live isochrone API, white-label hex-map embed, two-commuter search. Judge-panel winner: the consumer-feature angle (A: 24, B: 23.5, C: 23). One-pager PDF live at commutingcost.com/partners/commutingcost-data-onepager.pdf.

Newest handoff: fixed the two map defects the user spotted (DC/Arlington/Fairfax had no hexes; Edison→Philly cut off) and generalized the fix. Multi-state OSM merge (osmium merge) repaired DC (District-only→DC+MD+VA: walk nodes 312K→3.47M, coverage 4.7%→54.2%) + St. Louis (MO→MO+IL). Built 2 corridor MEGA-REGIONS (socal LA↔San Diego, midatlantic Philly↔Baltimore↔DC) with a priority region-picker + 7 bbox WIDENINGS to rail termini + Québec City — engine registry 51→54, all long-distance regional rail now reaches. Shipped a searchable MULTI-SELECT transit-system dropdown on the /map. Fixed drive/P&R hexes painted on open WATER via an OSRM snap-distance filter. 3 commits pushed (0edfe7a, b4f6c8a, b927240).

Newest handoff: took the multimodal system national end-to-end — pre-warmed the isochrone cache for 151,781 national POI dest cells (87,997 hexmaps, mirrored gzipped into commute.iso_precompute so map clicks serve GPU-free), grew engine coverage 31→50 metros (registry 51), shipped a Google-Maps-style national transit-LINE overlay (all 935 golden feeds, viewport-loaded, dynamic legend + State dropdown + Train/Bus toggles + DC cross-border filter), and CLOSED the keyed-feed set (WMATA + AC Transit + 6 SF Bay 511.org ops → every US+CA catalog feed now in golden). LOS AI-1003 restored to qwen2.5-coder:14b on gtfs3.

Newest handoff: moved the mortgage LOS AI-1003 assembly off geo2's shared GPU onto gtfs3's new dedicated RTX 5060 Ti (16GB) running qwen2.5-coder:14b (an upgrade over the 7b) — always-on, quota-free, warm ~0.6s. Fixed the new GPU with no reboot / no commute-DB downtime (595-open driver hot-swap), installed Ollama LAN-bound, and freed geo2. macstudio was investigated + rejected (contended); geo<->Mac passwordless SSH staged.

Newest handoff: redid the 4-seat YC funding review on the founder's v3 deck + 60-mo financials + the 100-repo org (B/B+, not A — number contradictions, 94% GM mislabel, ~25-product sprawl, mortgage-AI vaporware); reconciled the deck to a clean A across all 4 seats and re-themed it to the v3 visual style; took the mortgage POS/LOS over the finish line LIVE on geo2 with AI-1003 assembly on a self-hosted Qwen coder (Ollama, quota-free).

Newest handoff: the consumer surface got a full brand pass — commutingcost.com wears the Losing Ground theme site-wide, the new losingground-cities app (own repo, gated /admin social studio with avatar Shorts) is live under losingground.certihomes.com, plus cleanups (GPU/RAPTOR scrubbed at the generator source, VIA San Antonio fares, #45 map-res contract, SF fare edge-cases).

Newest handoff: the Certihomes $750K financial-OS raise stack went live — commutingcost.com redesign + multiuser /map + Losing Ground 2026 (losingground.certihomes.com) + investor portal + TLC widget + *.tlcengine.com wildcard renewed fleet-wide. True Lifestyle Cost (H+T+U) is the moat.

Newest handoff: the GPU engine — NYC-only behind a national-DB facade — is now MULTI-REGION across 32 US+CA metros. Per-metro walk/bike SSSP + transit RAPTOR networks built, engine flipped (NYC byte-identical), precompute 32/32, API all-9-modes national, traffic Google-calibrated, walkability/transit/best-commute rankings published off 497K POIs.

Newest handoff: national cost data (gtfs_fares 44k/522 feeds/36k OD, toll_rates 194/40 states, cbd_parking 1,515 garages) all wired LIVE into commutingcost.com (OD fares, national parking, DB-correct tolls) + a unified monthly GTFS Golden-Source ETL.

Newest handoff: the OUTBOUND isochrone feature (direction flag, P&R/K&R/bike+transit = capped access-to-station + transit, page controls, best-departure tuning, route/origin-aware tolls + tiered parking) and the React2Shell reinfection contained.

Newest handoff: drive ~10x faster (res-8), drive serialization fix (res-8 output), P&R multi-seed routes via NEC, OTP-NE oracle + 19-pair validation. Two ⭐ TODOs queued.

Newest handoff: MultiModal NYC map done; the drive-isochrone fix, 9-icon revert, cost model, and fuel-on-frontend pending.

The national roadmap — current status, data provenance, and the ordered next-pass plan.

True Lifestyle Cost Index — Housing + Utilities + Commute: v1 formula, primary sources, the five resolved decisions, and the CNT H+T® benchmark protocol.

11 counties × 3 income bases × 2 U scopes — HUD-vs-median gap quantified (+47% avg), LEAD county energy live, CNT r=0.90 holds.

commutingcost.com/TLCcosts — spelled-out naming, income-basis radio picker, dual-U display, tier badges, SEO/GEO templates, CNT attribution block.

The citable methodology for journalists and answer engines: framework, every source with vintage, validation r=0.90 — no model coefficients.

USPTO + common-law availability search: TRUE LIFESTYLE COST green (0 federal hits ever), first-use evidence 2013-07-23, TLC INDEX yellow (LISEP collision).

Positioning, segment priority (MLS warm channel first), pricing hypotheses, 30-day pilot motion, competitive table, 12-week plan.

Office-hours, engineering DD (incl traffic + OSRM), CEO, and pitch reviews synthesized into a fundability verdict + top-5 fixes for the Certihomes raise.

The moat as capabilities buyers lack; one page per target buyer; license-vs-acquire decision tree.

8-week dated per-platform calendar (LinkedIn/YouTube/TikTok/Facebook), KPI funnel, 3 budget tiers.

Avatar Short scripts, long-form outlines, blog drafts, podcast outlines (rendered on the geo2 Avatar Studio).

14-outlet journalist target list with per-outlet angle, press-kit one-pager, embargo strategy.

Ready-to-post LinkedIn, X (single + thread), Facebook, Instagram, and stat-card captions; brand rule: 'our proprietary algorithm', never the tech words. Emailed to krishna@certihomes.com.

Our recreation of CNT's Losing Ground with 2026 data: moderate-income households' Housing+Transportation+Utilities burden, national + metro.

Kiss&Ride, Uber, Citi Bike, Park&Ride — first/last-mile legs + a cost model around the unchanged kernel.

30 commuter towns (NJ/LI/Westchester/CT) — rail vs drive with REAL DB fares + peak traffic.

How the GPU RAPTOR kernel works and why every major decision was made.

The working checklist. A and B-core are done; C (walk matrix), D (tri-state), E (cost engine) are open.

Independent PhD-grade review: what's solid, what was wrong, what the product still needs.

Dated log of every working session — newest first. The narrative of how the numbers happened.

Self-contained context: hardware, data inventory, constraints, validation pairs, do-nots.

EIA-861 utility rates + territories, URDB tariffs, heating oil/propane, retail-choice registry, and the cross-fuel + savings views in the commute DB.

Programmatic SEO/GEO architecture for the utility-rate layer (spec only — domain/DNS is Krishna's call); NJ prototype at energy/site/index.html.

RECS-2020-calibrated per-home usage model: the 5 dominant variables, NJ attribute backbone, and the validation table.

FastAPI service: address in → monthly electric/gas/oil/propane out, with per-input provenance and provider alternatives.

TIGER 2022 geometries → 22M-cell crosswalk → hex-level CNT H+T view in the commute DB — closes the long-flagged blkgrp↔H3 item.