Projects
19 projects. Built, not mocked up.
16 AI projects (products that run, image tools, models trained from scratch) and 3 channel projects. A project can sit in two sections. Each page gives the context, the measured numbers and the measurement setup.
The projects
AI products
Agents, assistants and tools that run, delivered with their tests, measurements and documentation.
- 0 / 14attacks succeeding with the gateway (14 / 14 without), even with an agent deliberately rigged for the test
- 0leaks: evaluation 32 / 32 on SQLite and PostgreSQL alike
- −21%error on 28-day views against the rolling median: log10 MAE 0.499 → 0.393 (102 test videos)
- 0.95 / 0.79abstention precision / recall on 90 frozen test questions (0.81 / 0.45 before)
Also relevant here: RELIEF Studio, Atelier Qwen, Studio Poche
Films and series
Series and formats released on my channel, in front of a real audience: 95 million views.
Also relevant here: Thumbnails as a system
Images and style
Style models, image tools and local studios, evaluated before they are used.
- 10thumbnails delivered for 10 real channel videos
Also relevant here: ComfyUI extension, Video lab
Models and research
Trained models, scheduling and benchmarks: the foundation of the other projects.
La Forge
In progressA small French language model (16.9M parameters), written and trained from scratch
- +76.7%less weighted tardiness than FIFO; ATC, the reference, stays ahead (+78.2%)
- ×7.8video distillation speed-up (4 steps vs 16)Qwen2.5-3B Q4 · tokens per second, every run
Also relevant here: RELAIS, The Oracle, “vctrstyle” LoRA
- Released
- Functional
- In progress
Self-initiated projects, in-house tools and a YouTube channel. RELAIS and Vigie are run in a client format with fictional clients: each page states its context.