Demoly utilizes a multi-model architecture: - DeepSeek: Powers large language model (LLM) reasoning, text processing, transcript understanding, and conversational QA. - Google Gemini: Powers vision processing, multimodal image understanding, and visual frame analysis. (Note: Underlying AI model providers are internal architectural details and not marketed publicly).

1. Speech audio transcripts (when audio is present) 2. DOM structure, code hierarchies, and text nodes 3. OCR / on-screen rendered text 4. Mouse clicks, movements, and user interactions 5. Timestamped layout state changes 6. Video metadata 7. High-resolution visual screenshots captured at specific event timestamps

Demoly does not process heavy video pixel streams frame-by-frame. Instead, it continuously indexes textual and layout changes from the DOM tree. When a user asks a question that requires visual spatial verification (e.g., verifying a color, diagram, or unlabelled graphic), the system captures and analyzes a targeted screenshot from that exact DOM timestamp. This hybrid approach provides full visual intelligence while keeping processing fast and efficient.

A creator records a user creation workflow and fills out a form on screen while talking about general settings, never mentioning the word "email." - Transcript-only tool: Cannot answer "What fields are required to invite a user?" - Demoly: Inspects the captured DOM/visual state of the form, detects the Email Address (*required) field, and provides the correct answer with the exact timestamp.

A creator demonstrates a scrollable dashboard containing 100 users, but only 15 are visible on screen. If the creator verbally explains "We have active clients like Acme Corp and Globex on this tier," Demoly uses the speech transcript to answer questions about Acme Corp even though the name was scrolled out of view.

Demoly relies entirely on its DOM extraction engine. It analyzes buttons, alt text, headings, input labels, division structures, navigation events, and visual state changes to index the entire recording.

Demoly recognizes buttons, navigation bars, dropdown menus, input forms, modals, tables, dashboards, analytical charts, images, embedded video players, and custom web elements.

Yes. Demoly tracks sequential events (e.g., 1. Open Settings -> 2. Click Team -> 3. Select Invite -> 4. Assign Admin Role). When both DOM events and audio transcripts exist, sequence recognition is highly reliable; with DOM-only silent recordings, the LLM reconstructs sequences with ~50–60% baseline confidence.

Yes. Because DOM states are tied to precise timestamps, the AI can describe the exact state of the UI at any point in the recording.

Timestamps are currently 80% to 85% accurate. When a feature is discussed or displayed multiple times across a video, Demoly provides a primary match along with secondary candidate timestamps.

Demoly evaluates confidence scores. If confidence is low, it avoids asserting false facts, states that the information could not be verified with certainty, and suggests the closest matching timestamp or related workflow found in the recording.

- Causes: Hallucinations typically arise when a topic is mentioned repeatedly across different timestamps with slight variations, or when questions are overly vague and lack context. - Mitigation: Demoly consolidates ambiguous matches into a structured response presenting multiple verified timestamp links for the user to choose from.