Validation snapshot
What is verified today, what is still prototype, and what still needs public media.
This page is the trust layer behind the portfolio. It separates real validation from prototype framing so a recruiter, hiring manager, or engineer can see what was actually run on this machine on April 11, 2026.
Four things that are true right now.
The portfolio is stronger when the claims are narrow, current, and inspectable. This is the current state of the package.
The portfolio, film-proof page, role-fit page, and demo-script page are live and aligned around the filmmaker-first workflow story.
splat-web-viewer now has a real local Node toolchain pass: install, typecheck, lint, test, and build all completed successfully.
video-gen-eval and cloud-inference-platform both have passing local suites, and the video evaluation CLI was exercised end to end.
Prototype research layers are now explicitly presented as architecture explorations rather than trained production systems.
Validation matrix.
The point of this table is simple: show what passed, what command or flow was exercised, and what caveat still matters.
| Repo | Verified on April 11, 2026 | Result | Important caveat |
|---|---|---|---|
| creative-ai-workflows | Public pages, docs, and asset slots checked locally and published | Published and consistent | The strongest score bump still depends on one approved public still, clip, or reel. |
| splat-web-viewer | npm ci, npm run typecheck, npm run lint, npm run test, npm run build |
All green on Node 24.14.1 | The renderer is still a prototype viewer; real scene proof improves when a public splat asset or screen capture is added. |
| video-gen-eval | pytest plus a CLI merge of local physics checks with sample VBench, IVEBench, and TiViBench JSON |
28 passed, 1 skipped, unified score report generated | Adapter hooks are real, but the upstream official benchmark runners still live in their own native environments. |
| cloud-inference-platform | pytest across router and integration tests |
22 passed, 2 skipped | Serving backends are still partly prototype-grade; the strongest artifact is the routing logic plus the Triton kernel work. |
| cinematic-controlnet | Lightweight environment-aware test pass | 1 passed, 2 skipped | Full training and generation credibility still depends on real checkpoints, data, and a GPU PyTorch environment. |
What is real today.
These are the parts of the package that are defensible right now without stretching the claim.
The site now consistently leads from S4 and AI-assisted post into control, review, evaluation, and deployment.
The video evaluation repo can now merge official-style benchmark outputs with local heuristic scoring into one report.
The browser viewer now has a verified TypeScript build, lint, and test path instead of an unverified JS toolchain.
The cloud repo has working router and integration tests that make the enterprise story more than a README-only pitch.
What is still intentionally labeled prototype.
This is the part of the package that stays explicitly scoped so the strongest claims remain easy to trust.
The repo demonstrates interface design, data flow, and control ideas. It does not claim a trained physics or video backbone.
The deprioritized AV prototype remains architecture exploration only. It does not claim real-data training or production-level generation.
The portfolio still needs one approved public still, timeline screenshot, GIF, or short reel to fully close the creative-proof gap.
The viewer can build and test locally now, but a public screen capture or asset-backed clip will make the demo more memorable.
"The goal is not to look bigger than the work. The goal is to make the strongest parts of the work easy to trust."