Devlog #1: MVP Scope and Classification Pipeline
Building Jetson's core classification engine to automatically turn Help Scout conversations into GitHub issues. This week: MVP decisions, 77% accuracy baseline, deduplication system, and real-world test with 40 broken sessions. Technical decisions from our bi-weekly dev log.
This Week's Focus
We made five major decisions about MVP scope and built out the core classification pipeline. The big trade-off: shipping a working classifier now versus waiting for perfect accuracy.
Key Decisions This Week
MVP Scope Shipping the classifier and GitHub integration first. That's the core value. Documentation generation stays but as drafts requiring human review. Auto-publishing and auto-updating docs moves to v2 to avoid accuracy issues.
Classification Accuracy Set up a labeled test set with weekly accuracy checks measuring precision and recall per category. Current baseline: 77% accuracy. Target: 90% through improved prompts and embeddings.
Deduplication Strategy Implementing embeddings-based clustering to group similar issues. Instead of 37 copies of "app won't load," you get one grouped issue with all affected users linked.
GitHub as Source of Truth Close an issue in GitHub, Jetson marks it resolved. Adding project column mapping so non-technical team members can track progress without GitHub access.
Closing the Loop (maybe) When issues ship, Jetson identifies all reporters and drafts a single reply for bulk sending. Maintains your tone while eliminating repetitive work.
What We Built This Week
Classification Pipeline Implemented the core Help Scout to GitHub pipeline:
- Webhook triggers on new conversation
- Content analysis via OpenAI API
- Classification into bug/feature/docs (77% accuracy currently)
- GitHub issue creation with original context
- Bidirectional sync maintains state
Deduplication System Built embeddings-based clustering to group similar issues. Instead of 37 copies of "app won't load," you get one grouped issue with all affected users linked.
Mobile Interface Every action works on mobile. Approve issues, merge duplicates, send replies. No desktop required for core workflows. Potential future update: Telegram/SMS/Whatsapp messages for busy founders who want quick updates/actions.
Live Test: Recipe Kit React Migration
Perfect timing for a real test. This week Patrick shipped Recipe Kit's biggest update ever: a full stack upgrade from Vue/Express to React. It didn't go great - lots of errors, broken workflows, and "we want this feature back" emails almost immediately. Over 45 emails received within a few hours - chaos for my support team.
Thankfully, Jetson handled it:
- Grouped all similar bug reports into unified issues
- Created GitHub issue with affected user count and relevant information/conversation URLs
- Patrick and his developer hopped on, fixed the Github issues immediately, and weren't bogged down by tracking down conversations / context.
Manual triage: ~4 hours. With Jetson: 15 minutes.
Progress This Sprint
- Built core classification pipeline (77% accuracy)
- GitHub issue creation working
- Deduplication system functional
- MVP 75% complete
Next Week's Focus
- Improve classification accuracy to 90% (prompt engineering + better embeddings)
- Build GitHub project column sync / Kanban board layout
- Add webhook support for Helpscout
- Rebuild UI with Shadcn components
Alpha Testing
If your team needs better signal from support noise, we'll connect Jetson to your Help Scout for free during alpha. We need honest feedback and messy data.