Dealroom pressure test, primary. Models and agent, parallel tracks.
End-to-end. Solo. Singapore. 2026.
RAZER-355M finished at step 100,000 with final loss 0.2952. The 1.4B GPT-NeoX run is a planned next phase, pending AWS vCPU quota approval for g4dn.12xlarge.
22.7M records. 87GB raw.
The primary product. A seller uploads their pitch material, the app assembles a simulated buying committee for the target customer, each exec raises objections against the material, and a gap report synthesises what the material is missing. Output is a results page, an optional rebuttal doc, and an optional read-only share link.
Session history and share links are stored as JSON on disk under data/, with sensitive fields encrypted at rest using AES-256-GCM: customer name, deal size, material text, committee, objections, gap report, rebuttals. Writes are atomic — temp file plus rename — and serialised through a single-slot Promise queue so concurrent appends never tear. History entries expire 30 days after creation and are swept on an interval and lazily on load. Corrupt JSON files are preserved to a timestamped sidecar, never deleted.
Tokens are minted with crypto.randomBytes(32) and base64url-encoded. Each record stores only a SHA-256 hash of the owner's auth cookie, so revoke can be scoped to the original owner without keeping the raw cookie on disk. Links expire after 7 days. Expired or revoked tokens return 410 Gone with a no-store cache header. The public /shared/:token route is rate-limited to 60 requests per minute per IP, sets X-Robots-Tag: noindex, nofollow, and writes an access log with SHA-256-hashed IP and truncated user-agent. The log is capped at 10,000 rows and trimmed on each append.
Tests run with Jest against an in-process supertest harness. 451 tests across 15 suites cover route contracts, rate limiting, CSRF enforcement, the full share-link lifecycle (create, expire, revoke, rate-limit, rebuttal redaction), history encryption round-trip, and shared-view field parity. The sentinel convention — one unique string per rendered field, asserted to appear in the response body — was added in Phase 18 after a field-name mismatch shipped to prod where objection text rendered blank because the renderer read the wrong field names. Positive-content assertions now guard every surfaced field.
The 1.4B GPT-NeoX run is a planned next phase, pending AWS vCPU quota approval for g4dn.12xlarge. The 355M is complete and hosted on HuggingFace. Neither claims frontier-model performance. They exist to demonstrate end-to-end ownership of the training pipeline — corpus assembly, tokenizer training, distributed training, evaluation, and deployment.
The dealroom app does not use either of these local models. It uses Anthropic Claude Sonnet for committee generation, objection generation, and gap-report synthesis. The local models and the dealroom app are parallel demonstrations, not a single integrated system.