← gpu-pilot · json · source: CONTINUE_6_27_2026.md

▶ CONTINUE 6/27 (YC re-review → clean A · mortgage POS/LOS live w/ self-hosted Qwen AI · deck re-themed to v3 — newest handoff)

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). · updated 2026-06-27 15:08

TL;DR

  • Deck (gpu-pilot/gtm/deck/build_certihomes_deck.py): reconciled to the financial model (Y1 450/$6.01M, CAC $600, ROI/CAC 20.5x->59.2x, break-even M6, EBITDA 35->60%, 9-agent Y1 bridge), 94% relabeled + EBITDA-led, one wedge + Now/Next/Later + appendix, honest claims (~23x never 1000x; published patent APPLICATION; AUS sandbox). Re-themed to the v3 style (black + KW-red #CF142B, sans, investor-memo footer). V3_RECONCILIATION.md = before->after memo for the founder's Google deck+sheet. Commits a4c2c3b/d5ab475/b89d85c. Live PDF: losingground.certihomes.com/certihomes-investor-deck-d286894e.pdf.
  • Mortgage POS/LOS LIVE on geo2 (NEW repo tlcengine/certihomes-los; los.tlcengine.com): full 1003/URLA -> AI-assembled validated URLA -> loan pipeline + auto-conditions -> AUS labeled sandbox. AI-1003 assembly runs on self-hosted qwen2.5-coder:7b via Ollama (geo2 localhost:11434) as the primary provider — quota-free, no per-call cost; provider chain ollama,claude->deterministic; assembly_source honest. certihomes-los main=a726e8f. posSubmit.js POS->LOS bridge committed (bitbucket dev_stevens_mortgage main_dev) but not yet imported into the live POS.
  • Tech A-gates (engine): hexmap-api/kernel_gate.py = bounded GPU FIFO queue replacing the 429 single-lock (concurrent reqs queue; live-validated 3->all-200) on :8098; honest ~23x benchmark (raptor-design/bench_cpu_raptor.py + BENCHMARK.md, bit-exact GPU==CPU) retiring the 1000x. Commit a0d0214.
  • GOTCHAS: auto-formatter strips unused imports on save (crashed the live :8098 engine + the LOS — add usage in the same edit or # noqa); geo2 certihomes-los venv has a broken pip shebang (use python -m pip); fleet Claude proxy :8080 ignores output_config/json_schema + is weekly-quota-limited (use self-hosted Qwen); geo's Ollama :11434 is firewalled from geo2 (use geo2-localhost); no Google Drive rclone remote so the deck can't auto-upload to Drive (host static on the fleet). gpu-pilot main=b89d85c.

CONTINUE — 2026-06-27 — YC re-review → A · POS/LOS live (self-hosted Qwen AI) · deck re-themed to v3

Host: geocoder (RTX 5080). gpu-pilot main = b89d85c (pushed via geo bundle). certihomes-los main = a726e8f (NEW repo, pushed from geo2). Read first: memory certihomes-poslos-deck-pipeline, certihomes-raise-stack-live, public-copy-no-gpu-raptor.

DONE this session

  • Tech A-gates (engine): hexmap-api/kernel_gate.py — bounded GPU FIFO queue replaces the 429 single-lock (concurrent reqs QUEUE, validated 3→all-200; /healthz.kernel_gate); deployed live on :8098 (a0d0214). Honest ~23× benchmark (raptor-design/bench_cpu_raptor.py + BENCHMARK.md, bit-exact GPU==CPU) — retires the "1000×".
  • 4-seat YC funding review redone on the founder's v3 deck (Drive 1sHktF…) + the 60-mo financial model + the 100-repo tlcengine org → B/B+, NOT A: number contradictions (deck "60–100 listings" vs model 450; CAC $218 vs $600; break-even M5 vs M6; traction 10/$340K→7/$390K), 94% "gross margin" mislabel, ~25-product/100-repo sprawl, mortgage-AI vaporware (GitHub certihomes-pos was boilerplate), hockey stick.
  • Deck reconciled → clean A across all 4 seats (a4c2c3b): every number tied to the model (Y1 450/$6.01M, CAC $600, ROI/CAC 20.5×→59.2×, break-even M6, EBITDA 35→60%), 94% relabeled "transaction GM (pre-personnel)" + EBITDA-led, new 9-agent Y1 bridge slide, ONE wedge + Now/Next/Later + appendix expansion vectors, honest mortgage/IP/tech, hero = Collingswood $663K deal. gtm/deck/V3_RECONCILIATION.md = slide-by-slide before→after memo to apply to the founder's Google deck + sheet.
  • Deck re-themed to the v3 visual style (d5ab475): black + KW-red #CF142B, heavy sans (Carlito-bold) headlines, red ALL-CAPS eyebrows, dark cards, green-check competition matrix, footer CertiHomes · Pre-Seed Investor Memo · Confidential N/17. Content 100% preserved.
  • POS/LOS over the finish line + LIVE on geo2 → new repo tlcengine/certihomes-los (/DataDrive/krish/krish/certihomes-los/; 3 pm2 svcs certihomes-{losapi:5053,los-ui:5052,posapi:5051}; nginx {pos,posapi,los,losapi}.tlcengine.com). Demo https://los.tlcengine.com: full 1003/URLA → AI-assembled validated URLA → loan in a stage pipeline + auto-conditions → AUS = labeled sandbox. posSubmit.js POS→LOS bridge committed to bitbucket dev_stevens_mortgage main_dev (295119b) but NOT imported into the live POS yet.
  • AI-1003 assembly LIVE on self-hosted qwen2.5-coder:7b via Ollama (geo2 localhost:11434) — primary provider, quota-free, proven (2 stable assembled via ai:qwen2.5-coder:7b@ollama runs); /health shows the active model; assembly_source honest (only ai:* when all 7 URLA sections parse). Provider chain CERTIHOMES_LLM_PROVIDER=ollama,claude → deterministic (a726e8f).
  • Deck S8 bumped to "live self-hosted Qwen AI-1003 — no per-call API cost or quota" (b89d85c) + published.

DECK URL (live, shareable)

https://losingground.certihomes.com/certihomes-investor-deck-d286894e.pdf — static on geo /var/www/losing-ground/, unguessable token (Confidential). Overwrite that exact file to refresh; URL stays stable. Source: gtm/deck/build_certihomes_deck.py → CERTIHOMES-seed-deck.pptx/.pdf.

PENDING (precise)

  • Deck → Google Drive: not automated — geo's rclone has only aws/minio (no Drive remote) and the Drive create_file MCP needs the PDF as inline base64 (~340K tokens for 1 MB → infeasible in-context). Options: user "Save to Drive" from the fleet URL, OR set up a Drive rclone remote (one-time OAuth) for programmatic push. (Founder's deck folder parentId 0ACrTvpAqgn9_Uk9PVA; my reconciled v4 already there as 1lpL9nCf….)
  • POS→LOS live wiring: import posSubmit.js into AFLConfirmationForm on pos.tlcengine.com (1 import + 1 call) so the borrower POS submits into the LOS (today the full flow runs via los.tlcengine.com's own intake tab). Staged, not imported.
  • Real AUS (Fannie DU / Freddie LPA): needs a sponsoring-lender + AUS credentials — sandbox until then (honestly labeled).
  • Durable-A (real-world, not paper): ≥1 signed pilot; visibly park the 100-repo sprawl; bottoms-up closes behind the 450 Y1. The DECK is A; the business must keep matching it.
  • #61 LG-native commutingcost.com page; #62 multiuser N-person FE collapse — still open.

DECISIONS (user chose + why)

  • Self-hosted Qwen (Ollama) for the LOS AI, not the Claude proxy (quota-limited) or a paid Anthropic key — always-on, zero per-call cost, and reinforces the "runs on our own GPU fleet" capital-efficiency story.
  • Deck reconciled to the financial MODEL (one source of truth) and re-themed to the founder's v3 visual style (content unchanged). Honest claims always: ~23× (never 1000×), "published patent APPLICATION (not issued)", AUS sandbox.
  • Deck hosted static on the fleet (losingground.certihomes.com static root) since Drive auto-upload isn't wired.

GOTCHAS (don't relearn)

  • Auto-formatter strips "unused" imports on save → add the usage in the SAME edit or pin # noqa; py_compile won't catch the resulting runtime NameError. Bit me on app.py (kernel_gate import → live engine crash-loop) AND assemble.py (urllib/Optional).
  • geo2 certihomes-los/venv has a broken pip shebang (relocated venv) → use venv/bin/python -m pip, never venv/bin/pip. anthropic wasn't installed at all (the real reason AI fell back) — now in via python -m pip.
  • Fleet Claude proxy (:8080, pm2 claude-proxy) serves claude-opus-4-6-thinking + claude-sonnet-4-6 (NOT claude-opus-4-8), IGNORES output_config/json_schema (returns Markdown → use a strict-JSON prompt + _extract_json), and is weekly-quota-limited (RESOURCE_EXHAUSTED, ~167h reset). Hence self-hosted Qwen.
  • Ollama on the fleet: geo2/geocoder/geo all run :11434; geo's :11434 is firewalled from geo2 → geo2 services must use geo2-localhost Ollama. qwen2.5-coder:7b pulled on geo2 (RTX 5090, ~6.9 GB resident beside ComfyUI/kokoro).
  • Push routes: gpu-pilot via geo bundle (geocoder→172.26.1.45→GitHub; git bundle origin/main..main → scp → fetch+push refs/temp → git update-ref refs/remotes/origin/main HEAD). certihomes-los pushes direct from geo2 (https GitHub auth). Bitbucket needs GIT_SSH_COMMAND="ssh -i ~/.ssh/bitbucket_phase4".
  • ssh geo/geo2 hostnames don't resolve from geocoder → use IPs 172.26.1.45 / 172.26.1.151.

Comments