PitchPilot

Preparing a product for a sales or investor conversation involves more than having the software working. PitchPilot automates the first pass at decks, written pitches and demonstration videos from product information, then keeps the output available for review and revision. It is used internally at Kuzushi Labs.

Status: Live internal tool at Kuzushi Labs
Purpose: Automate the production of pitch materials and product demonstrations

Turn product context into something people can see

PitchPilot is a live internal tool used to turn product information into presentation assets: pitch decks, videos, outreach copy and landing-page content. The project began as a hackathon build and continued into practical internal use.

The recurring problem was the work between a functioning product and a clear demonstration of it. A website, a feature list and a working application contain useful material, but each has to be reshaped for the audience. Repeating that work across a deck, a video and a short written pitch creates substantial room for inconsistency.

PitchPilot connects product discovery, structured content generation and media assembly into a reusable workflow. The output remains reviewable, with feedback feeding the next version.

Sample output: an automated product walkthrough

The ELANO Studio video below was produced through the automated demonstration workflow. It follows a working product through catalogue browsing, natural-language search, room visualization and product-information enrichment.

Open or download the MP4 sample.

Text description of the sample

1. The catalogue presents tile collections, material filters and product cards, including an English/French interface.
2. Natural-language search connects a descriptive request to relevant products.
3. A room photograph and a selected material are used to produce a visual concept, followed by a before-and-after comparison.
4. A product page shows how a French voice note supports reviewed, structured product information and searchable attributes.

This sample shows the screen-capture path. PitchPilot's generated pitch-kit path also supports narrated slide videos; the walkthrough above intentionally contains no audio track.

Two useful production modes

Agent-guided pitch generation starts with a URL. A browser agent gathers product information, extracts brand and positioning context, and produces a structured plan for the assets. Image and speech services supply media, while deterministic assembly produces the shareable files.

Scenario-driven product capture starts with a defined walkthrough of a running application. Browser automation performs the interactions, waits for the required UI state and records the experience. A reusable scenario supplies the sequence and accompanying descriptions. The same capture workflow can support both a video and a screen-by-screen document.

The two modes serve different needs. A pitch kit explains the product to a new audience. A recorded walkthrough shows the audience what the product actually does.

Product discovery becomes a structured brief

The browser workflow uses Playwright tools to navigate, extract page text and inspect brand information. GLM-5.1 coordinates the discovery loop and returns a structured product brief: product name, target user, value proposition, differentiation, features and visual context.

That brief is the common input to the rest of the pipeline. Slides carry titles, body copy, image prompts and narration scripts. The deck and video can therefore be assembled from related material, while outreach and landing copy are adapted for their own formats.

The pipeline separates Browse → Extract → Write → Generate → Assemble. This gives the UI useful progress states and gives failures a stage to belong to. A media-generation failure can be distinguished from a problem gathering the source information.

Generation and assembly have different responsibilities

Image generation and text-to-speech introduce provider latency and variable outputs. The application handles media generation as asynchronous work and collects the resulting assets before assembly.

The original generation stack uses Gemini image generation and MiniMax speech through provider adapters. Narration language and voice are selected as part of the request. The content model supports multiple output languages, while the asset record associates each slide with its text, image, audio and duration.

FFmpeg assembles the video from those assets. Narration duration influences segment timing; a stitching plan connects the content plan to the final sequence. PDF decks are rendered from branded HTML templates using Jinja2 and browser-based PDF generation.

For application walkthroughs, Playwright records the actual interactions and FFmpeg packages the recording into a shareable H.264 MP4. The supplied ELANO Studio sample is 1440 × 900 at 25 fps, making it a concrete output that can be inspected directly on this page.

A visible workflow and a revision loop

The Browser Theatre interface displays browser screenshots alongside the workflow's activity. Server-Sent Events keep progress and generated artifacts visible while a run is in progress. Model-call records capture purpose, token use and latency so the team can investigate where time and cost are spent.

The output is associated with a session and version. Natural-language feedback is recorded and used during regeneration, allowing the product story to be revised without treating every attempt as an unrelated artifact.

This matters for internal use: the first generated deck or video is a starting point for review. A useful tool needs an iteration path as well as a Generate button.

Why scripted capture remains valuable

An agent is useful when it must discover an unfamiliar product. Once a demonstration flow is known, a scripted scenario provides a repeatable way to show it. Navigation, input, waiting, scrolling and comparison interactions can be specified deliberately.

Waiting for a result to appear is different from assuming the result will appear after a fixed delay. The scenario can express both readiness conditions and pauses for the viewer to read a screen. This connects the needs of browser automation with the pacing of a product demonstration.

The ELANO example moves through a catalogue, multilingual search, a generated room concept and product enrichment. That makes the output a demonstration of a complete customer journey rather than a collection of disconnected screenshots.

Internal value

PitchPilot packages repeated presentation work into a reusable internal capability: gather context, draft the story, generate or capture the assets, assemble the output and revise it. A video can be embedded in a proposal or portfolio; a deck can support the same product conversation; written copy can reuse the underlying brief.

The live internal workflow and the automated walkthrough make the result tangible: working software can be turned into a presentation asset that is ready to review, share and reuse.

Contribution

My contribution covered the product workflow, browser automation, agent orchestration and media-assembly path, along with the interface for observing a run and iterating on its output. The work developed PitchPilot from an initial hackathon build into an internal tool for producing and reviewing presentation assets and product demonstrations.