Orbit: Client Discovery Notes

Orbit: Turn Client Discovery Notes into a Project You Can Review This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle What I Built A client discovery call can leave a freelance developer with a page of notes and several unanswered questions. Before building anything, the developer has to separate deliverables from assumptions, identify missing decisions, and turn the conversation into a plan. I built Orbit, a spatial notebook that helps with that step. Ideas live in circles. Open a circle to find another layer of notes, tasks, and questions. Add text boxes, images, and color to organize the project in a way that makes sense to you. The main workflow is client notes → AI-generated draft → human review → organized project notebook. Import a text or Markdown brief, review the suggested deliverables and tasks, inspect their source excerpts, and choose what to keep. Orbit preserves the original brief alongside the resulting plan. My intended customer is a solo freelance web developer preparing client scope. The potential paid value is help turning discovery notes into a reviewable plan, with unresolved questions kept visible. This is a working prototype; I have not yet validated willingness to pay. Demo Try Orbit — public demo Two walkthroughs show different sides of Orbit: Client planning: import the synthetic Cedar Studio brief, review its proposed scope and source excerpts, then create a notebook containing Deliverables, Features & tasks, Client constraints, and Questions to clarify. Depth and navigation: explore the 24-space “Milkyway galaxy” notebook through Our Solar System → Earth → Continents → any of the seven continents. You can also ask a question about the active notebook and follow the answer’s source buttons back to the notes. Spoken questions are transcribed into editable text before submission. Circle images can appear with or without text labels. The public demo runs on Render and opens with the Milkyway notebook. Hosted Atlas saving and reload persistence have been verified. Each browser has an anonymous cloud workspace; clearing cookies removes access to that cloud copy. Use sample content and export anything important. The free service may take about a minute to wake after inactivity. Code Orbit on GitHub The project uses a browser interface and a Node.js/TypeScript backend. The repository includes the application, automated tests, workflow code, and a dated evidence log that separates completed integrations from planned work. How I Built It Gemma 3 4B generates note-organization suggestions, draft client plans, and notebook answers. Orbit supports local inference through Ollama and hosted inference through OpenRouter. Both paths have been exercised with synthetic notes. Mastra coordinates three workflows: organize notes, import a client brief, and answer notebook questions. Each workflow prepares the input, requests structured output from Gemma, and validates the result before showing it to the user. A useful lesson came from checking source references. A valid line number does not prove that an AI interpreted a requirement correctly. Orbit therefore distinguishes exact source quotations from suggestions, preserves source text, and requires review before adding generated work. Notebook answers reject unknown sources and unmatched quotations. These checks help catch errors; they do not guarantee a correct interpretation. MongoDB Atlas provides optional workspace persistence. Notes save in the browser first, with explicit cloud-sync choices, recovery copies, and revision checks to prevent stale saves from silently overwriting newer work. Live checks verified saving, loading, image persistence, and conflict rejection. ElevenLabs Scribe v2, accessed through OpenRouter, transcribes recorded questions. The transcript remains editable and is never automatically submitted to Gemma. Synthetic WAV transcription has been verified; actual microphone recording still needs owner testing. The current automated suite has 35 passing tests, covering workflow validation, notebook isolation, storage conflicts, recovery, transcription behavior, image limits, and reconnection after a failed Atlas connection. Why Does Open Innovation Matter? Client briefs can contain information a freelancer wants to keep on their own machine. Running Gemma through Ollama gives Orbit a local inference option without requiring a hosted model API for those workflows. Atlas backup is optional; enabling it sends notebook content to the cloud. Hosted inference and transcription have their own processing disclosures. Using an open-weight model and an open-source workflow framework also lets me inspect and change the process. When the first generated plans included generic tasks the client had never requested, I could tighten the prompt, adjust validation, and test the revised behavior. The provider adapter lets me choose between local and hosted Gemma while keeping the review workflow. For this project, that control matters: I want a tool whose handling of client scope I can examine and improve as I learn where it makes mistakes. My Agent Session I used Codex for implementation and GitHub Copilot CLI for a separate, read-only review of client-brief import. Copilot found two problems: partial quotation matching could lose a negation, and removing blank lines could shift source-reference numbering. I used Codex to implement the fixes and add three regression tests. Read the saved Copilot review and the findings and fixes. Prize Categories Best Use of Gemma: Gemma powers planning, organization, and notebook questions. Best Use of Mastra: three workflows coordinate preparation, model generation, and validation. Best Use of MongoDB Atlas: optional notebook persistence, recovery support, and revision-checked saves. Best Use of GitHub Copilot: a documented review found issues that led to fixes and regression tests. Best Use of ElevenLabs: Scribe transcription turns spoken questions into editable input.