Markdown source

Reconstructed Slides: Dream Machine: Scaling to 1m users in 4 days — Keegan McCallum, Luma AI

Source Video

Dream Machine: Scaling to 1m users in 4 days — Keegan McCallum, Luma AI

Method

This deck is reconstructed from the existing video frame captures by detecting likely slide regions with OpenCV, cropping/upscaling those regions, deduplicating similar crops, and OCRing the cropped slide images locally. It is a cleaner companion to the full-stage frame deck.

Reconstructed Slides

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.94
• Text source: agent_vision.

Slide text:

AIE

Luma

Microsoft

smol.ai

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.93
• Text source: none.
• OCR decision: ready — dense embedded social post screenshot with small text

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.94
• Text source: none.
• OCR decision: ready — dense chat screenshot collage with small text

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: agent_vision.

Slide text:

Luma's mission is to build multimodal general intelligence that can generate, understand, and operate in the physical world

• Recreated text/layout view: open HTML recreation
• AI slide classifier: title_card confidence 0.98
• Text source: agent_vision.

Slide text:

Public API

Check it out at:

https://lumalabs.ai/api/pricing

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.92
• Text source: advanced OCR rapidocr-live/bright-screen/contrast reconciled by agent.
• OCR decision: ready — Dense repeated diagram labels are better handled by OCR than direct transcription.

Slide text:

AIE

CPU Worker

triton-inference-server

CPU Worker

triton-inference-server

CPU Worker

triton-inference-server

CPU Worker

triton-inference-server

CPU Worker

triton-inference-server

CPU Worker

triton-inference-server

Luma

Microsoft

smol ai

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: advanced OCR rapidocr-live/bright-screen/contrast reconciled by agent.
• OCR decision: ready — Body bullets are small and dense enough that OCR is the safer extraction path.

Slide text:

Challenges

- Brittle, need to coordinate between both CPU and Triton being up at the same time

- Triton not built for multi-gpu/multi-node

- Push model not ideal for multi-node (which node has rank 0?)

- No/limited support for non-nvidia chipsets with Triton

- Very difficult to develop against

- Need to have every piece everywhere, hard to bring in disparate compute (i.e. from our training cluster :kekw:)

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.91
• Text source: advanced OCR rapidocr-live/border-trim/contrast reconciled by agent.
• OCR decision: ready — Dense architecture diagram labels and connectors are better suited to OCR than direct transcription.

Slide text:

API

Redis (standby)

Redis

Redis (standby)

GPU Workers

CPU Workers

Seaweedfs

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: agent_vision.

Slide text:

Challenges

- Backpressure

- Priorities/fair scheduling

- Handling many different models

- Handling Bursts

• Recreated text/layout view: open HTML recreation
• AI slide classifier: title_card confidence 0.96
• Text source: agent_vision.

Slide text:

Queues, Queues, Queues

• Recreated text/layout view: open HTML recreation
• AI slide classifier: title_card confidence 0.98
• Text source: agent_vision.

Slide text:

Model Management

• Recreated text/layout view: open HTML recreation
• AI slide classifier: title_card confidence 0.97
• Text source: agent_vision.

Slide text:

THANK YOU

Hidden Non-Slide Evidence

• `slide-001.jpg` — sponsor_logo confidence 0.99; sponsor logo wall, no presentation content
• `slide-002.jpg` — speaker_stage confidence 0.98; speaker on stage, not a slide
• `slide-007.jpg` — demo_video confidence 0.99; embedded video footage, not a presentation slide
• `slide-008.jpg` — demo_video confidence 0.99; embedded video footage, not a presentation slide
• `slide-017.jpg` — speaker_stage confidence 0.99; Camera shot of speaker, audience, and projected screen; not a presentation slide.

Classification audit: raw/sources/slide-ai-classification/reconstructed/EY4O9M6AsWI/audit.json