# Sparkie: product landscape and presentation evidence

Reviewed **September 20, 2026**, against official public product pages and this repository. These are product descriptions, not independent performance benchmarks. Website positioning and branding can change. No latency, accuracy, revenue, or time-saving statistics are used in the presentation.

## The positioning

**A teammate inside the meeting, with hands on the project.**

Sparkie combines shared meeting context, explicit spoken delegation, asynchronous project work, and a spoken return path. The story is about that interaction loop. It is not a claim that competitors only take notes, cannot use tools, or cannot join meetings.

Three useful, overlapping centers of gravity:

- **Remember:** meeting knowledge, notes, synthesis, preparation, and follow-through.
- **Converse:** voice interactions that advance a business workflow.
- **Execute:** workplace context connected to tools and deliverables.

Sparkie explores the combination in a multi-person Zoom meeting: discuss → address the agent → work in a project → inspect a result → explicitly request a revision.

## Official product research

| Product and official source | Published use cases observed | Relevance to Sparkie |
| --- | --- | --- |
| [Granola](https://www.granola.ai/) | Meeting notes, context across meetings, pre-meeting briefs, follow-ups and project-plan drafts. The page describes computer-audio capture without a meeting bot. | Strong adjacent experience for turning discussion into useful context. Sparkie explicitly occupies a participant seat and responds by voice. |
| [Otter](https://otter.ai/) | Meeting transcripts, summaries, decisions/action items, CRM connections, and sales/recruiting agent use cases. | Meeting intelligence already extends into workflows. Do not present Otter as transcription alone. |
| [Fireflies](https://fireflies.ai/) | Meeting capture and analysis, follow-up drafts, CRM updates, project-task creation, and MCP access to meeting knowledge. | Capture and downstream execution overlap. Sparkie emphasizes iterative spoken delegation into the active project. |
| [Vapi](https://vapi.ai/) | A builder platform for voice agents; the page exposes support, lead qualification, and appointment scheduling examples. | Voice can carry out work, rather than merely answer questions. Sparkie's current product surface is a shared meeting. |
| [Retell](https://www.retellai.com/) | Phone/contact-center agents, appointment booking, qualification, system updates, and human escalation. | Strong precedent for voice-to-action workflows. Sparkie explores team collaboration rather than a packaged phone workflow. |
| [ElevenLabs Agents](https://elevenlabs.io/agents) | Conversational agents, tools and workflows for support, sales and operations across voice and other channels. | Voice orchestration and tool use are established categories. Sparkie's differentiating focus is how they compose inside a working team meeting. |
| [Zoom AI-assistant product page](https://www.zoom.com/en/products/ai-assistant/) | The retrieved page is branded **ZoomMate**, with research, documents/decks, task delegation, routines, memory and connected workplace workflows. | Direct overlap with the vision of conversations becoming deliverables. Sparkie is an inspectable prototype in a developer-controlled project; no claim of unique agentic execution. |

Short excerpts supporting the categories:

- Granola: “Notes, actions and memory. Without a meeting bot.”
- Otter: “Automate the action items.”
- Vapi: “Customer Support”, “Lead Qualification”, “Appointment Scheduling”.
- Retell describes “booking appointments or updating systems”.
- Zoom: “Turn inspired discussions into impressive deliverables.”

All links were fetched during research. The comparison is deliberately qualitative. It does not assert that omitted features are absent, nor that vendor marketing claims have been independently validated.

## What the repository supports

The review covered the application and provider adapters, native Zoom bridge, workspace frontend/server, task execution, simulations, tests and operational documentation. The core product narrative follows the documented macOS Zoom path, not a mock session or a proposed integration.

| Presentation claim | Repository evidence | Boundary |
| --- | --- | --- |
| Sparkie joins a real Zoom meeting | [README](../README.md), [macOS Zoom guide](../docs/zoom-macos.md), [platform validation](../docs/zoom-setup.md), native bridge under [native](../native/) | The demonstrated path is macOS Zoom. Linux, local audio, browser audio and simulations existing in the repo do not prove equivalent real-session behavior. |
| Separate human tracks provide context | [Participant STT](../src/sparkie/participant_stt.py), [Zoom input](../src/sparkie/zoom_audio.py), [Realtime Zoom audio](../src/sparkie/realtime_zoom_audio.py) | Excludes the bot's SDK track. Sound played on human speakers can still reenter through a human microphone. |
| A pause need not end a request | [Semantic turn detection](../docs/semantic-turn-detection.md), [semantic turns adapter](../src/sparkie/semantic_turns.py) | Separate per-participant Realtime semantic detectors; medium eagerness in the current setup. Probabilistic, not a guaranteed endpoint. |
| Completed turns are routed semantically | [Wake-router research](../docs/interfaces.md), [wake router](../src/sparkie/wake_router.py), [Realtime session](../src/sparkie/realtime_session.py) | The current configured experiment uses Gemini 3.5 Flash Minimal through a separate tool-free Devin ACP process. Calling the agent by name is not alone proof it should reply. |
| Human speech can interrupt playback | [Realtime session](../src/sparkie/realtime_session.py), [Realtime Zoom audio](../src/sparkie/realtime_zoom_audio.py), [semantic turn notes](../docs/semantic-turn-detection.md) | Speech activity can pause playback; absent confirming text, a 350ms candidate window allows buffered audio to resume. Confirmed speech cancels the current reply. No perfect speaker-echo rejection claim. |
| Voice and tools run separately | [Realtime](../src/sparkie/realtime.py), [task center](../src/sparkie/task_center.py), [task workers](../src/sparkie/task_workers.py), [Devin ACP](../src/sparkie/devin_acp.py) | GPT Realtime is foreground voice. Devin SWE 1.6 Fast is the current background worker, with Codex selectable. Files, shell, network and configured tools run in the project. |
| The team can refine a saved result | [README](../README.md), [demo shooting script](sparkie-demo-script.html), [task center](../src/sparkie/task_center.py) | The task starts with finalized context at delegation. Later decisions need an explicit update or follow-up. An acknowledgement is not a successful task result. Inspect the file. |
| Providers can change without changing the meeting contract | [Contracts](../src/sparkie/contracts.py), [interfaces](../docs/interfaces.md) | A separation of implementation responsibilities, not a promise of identical behavior across providers. |

In this mode, complete assembled text reaches the foreground voice session. Audio is also sent to separate per-participant Realtime sessions for semantic turn completion. Avoid a diagram that implies a single undifferentiated model receives all audio and independently runs all tools.

## Broader use-case map

The six interactive scenes are **possibilities**, not six additional validated integrations. “Imagine” is deliberately visible beside the possible result.

| Room | Moment worth acting on | Possible delegated work | Review or integration boundary |
| --- | --- | --- | --- |
| Product review | The team chooses a direction | Turn the discussion into a brief, alternatives table or first spec | Team checks scope and decisions. No invented owners or deadlines. |
| Engineering | A failure or tradeoff becomes concrete | Inspect the repository, compare designs, draft a fix or test plan | Repository access and human code review; no implied production deployment. |
| Research | Two claims need evidence | Find sources, compare findings, draft a sourced memo | Verify citations and source quality. Web access/tool configuration is required. |
| Customer call | Priorities become clear | Draft a tailored follow-up, proposal outline or requirements summary | A draft, not an email silently sent to a customer. CRM integration requires actual tools/access. |
| Hiring debrief | Interviewers compare observations | Organize notes against an agreed rubric and identify missing evidence | Human review; not automated hiring decisions or inferred sensitive traits. |
| Launch planning | A dependency or plan changes | Update a project checklist or draft a rollout brief | Explicit updates; no automatic assignment of people, dates or external project tickets. |

Additional rooms for Q&A: a design critique leading to a revision brief; an incident review producing a draft timeline; a workshop synthesizing questions into a research agenda; a weekly review updating a project document. These extend the same loop. They are not new native connectors or independently tested workflows.

## Evidence for the demo slot

The actual product loop is documented in the README and the existing shooting script: three teammates plan a demo film, ask Sparkie for an outline, keep discussing filming, open the real output, and ask it to incorporate the agreed roles.

The presentation contains **illustrated placeholders**, not a recording or a fabricated result. Add a real session excerpt and a real artifact screenshot before presenting it as evidence. If an edit shortens a wait, label that edit in the recording. A human opens and shares the file; Sparkie is not claimed to understand video or autonomously share its screen.

User feedback accepted the current medium semantic-turn setting and audible responses in real sessions. That is qualitative feedback, not a latency benchmark. This presentation makes no universal accuracy, long-session reliability, multilingual voice, enterprise deployment, or unattended external-action claim.

## Visual and interaction references

- [Chrome: Smooth transitions with the View Transition API](https://developer.chrome.com/docs/web-platform/view-transitions). Shared element snapshots support continuity between distinct layouts. Used for circles, the Sparkie mark and artifact panels across slides, with a reduced-motion and unsupported-browser fallback.
- [Codrops: Vivid — Turning a Visual Experiment Into an Interactive Webflow Experience](https://tympanus.net/codrops/2026/09/15/vivid-turning-a-visual-experiment-into-an-interactive-webflow-experience/). The article explains state-to-state visual navigation and transitions as part of the experience. The deck uses explicit scene navigation; the homepage preserves ordinary document scrolling.
- [Awwwards: Animation websites](https://www.awwwards.com/websites/animation/). A surveyed inspiration index for expressive web motion, not a source for product claims.

The art, SVGs, layout and animation implementation here are original repository assets. No third-party site assets, fonts, animation libraries or tracking scripts are loaded. The product icon is reused from this repository.

## Deeper comparison used in the revised slide 3

The slide now stages three representative comparisons and then Sparkie itself. “Tradeoff” is a consequence of the chosen product surface relative to **an audible participant with direct project execution**, not a general quality score.

| Product | Strength | Tradeoff for this specific use case | What follows for Sparkie |
| --- | --- | --- | --- |
| Granola | Computer-audio meeting memory without a bot to admit | The notepad is not itself a speaking participant in that capture model | Give the shared meeting a participant that can be addressed and answer aloud |
| Vapi | Voice infrastructure connected to tools and business workflows | A developer still supplies the meeting transport, shared context and task application | Package the Zoom/context/project-work loop together |
| ZoomMate | Native meeting/workplace context with agentic document and workflow capabilities | Its packaged experience follows Zoom product surfaces and connected tools | Explore a project-local workbench with selectable CLI workers and inspectable artifacts |
| Sparkie | Spoken delegation, direct project tools, and explicit in-meeting revisions | Prototype setup, macOS Zoom validation scope, broad worker access and human review | A focused interaction experiment, not proven superiority or feature exclusivity |

The Vapi and Zoom tradeoffs are architectural/product-surface interpretations, derived from the official positioning, not tested claims about absent features. In particular, Zoom already executes work; this presentation makes no claim that it cannot create artifacts or connect external tools. Granola's own “without a meeting bot” description supports the distinction about its capture model. The wider seven-product survey above remains available in the sources drawer and speaker preparation.

Official site screenshots, actual product icons and links are now present on the slide. Asset provenance is recorded in [assets/CREDITS.md](assets/CREDITS.md).

## Practical worked scenarios

Revised slide 4 and slide 9 have different jobs:

- **Slide 4: inspiration.** Two scripted meeting vignettes reveal three dialogue turns each. The product-review dialogue establishes a choice and a disagreement before delegation. The customer-call dialogue establishes a changed priority before asking for a draft. These are imagined scenes, not transcripts.
- **Slide 9: practical playbooks.** Each case names the context, concrete instruction, expected artifact and acceptance check. This is how a team can judge work, not another list of possible uses.

The demo-film case is grounded in the real-session workflow documented in the repository README. The displayed outline structure is explicitly illustrative, because an actual generated outline has not been supplied for the presentation. The real recording and screenshot slot remains the place for direct evidence.

The bug-triage case is a worked instruction inspired by the reported multi-track input-queue failure: human participant transcription continued while an unused mixed queue accumulated frames. It describes a reproduction, producer/consumer investigation and regression expectation. It does not claim that Sparkie autonomously performed this repair by voice in a recorded meeting.

The customer-follow-up case is a worked example with an explicitly unconfirmed date. Its artifact is a draft, with human review before sending. No real customer, promised deadline or successful external action is invented.

The conversation example on slide 7 incorporates turn completion and interruption into participation: indirect mention → complete addressed request → correction. The two staged system diagrams show concurrency and the technical signal paths. Their motion illustrates ordering, not measured latency.

## Slide 3 pitch: the handoff gap

The live presentation now argues for Sparkie, rather than conducting a neutral product tour. Each comparison follows **They deliver → Still missing → Sparkie closes the gap**, followed by a positive synthesis of Sparkie's experience.

The gaps have deliberately specific scopes:

- Granola: its bot-free notepad capture experience does not put a speaking participant in the meeting. Sparkie adds a participant that can receive spoken delegation and respond.
- Vapi: voice infrastructure still needs an application that connects multi-party meeting input, shared context, participation decisions and project execution. Sparkie supplies that connected workflow.
- ZoomMate: the comparison is about the additional integration needed to reach a team's own local CLI/project environment. Sparkie directly integrates the project worker into the meeting loop. This is not a claim that Zoom cannot execute tasks, produce artifacts or connect developer tools.

The closing beat emphasizes the outcome: the team continues discussing, inspects the actual file, and asks for another edit in the same room. Evidence boundaries and prototype limitations stay in the research and Q&A rather than becoming the slide's main message.