Autonomous product feedback monitoring agent. Ingests signals from GitHub, reviews, Slack. Learns what matters to you. Calls you when it's critical. Gets smarter with every interaction.
npx @senso-ai/shipables install UJameel/SiftAn autonomous, self-improving AI agent that monitors what people are saying about your product across multiple channels and only bothers you when it truly matters.
The key differentiator: after each voice call, the founder's response (approve/dismiss/create issue) is stored as feedback. An LLM then generates a new heuristic rule:
"GitHub issues mentioning 'data loss' from production repos should always be escalated" "Feature requests from new contributors are low priority regardless of tone"
These rules are injected into every future analysis prompt. The agent literally gets smarter with every interaction.
pip install -r requirements.txt
# Set up Ghost DB (or use local Postgres)
ghost db create sift
# Configure environment
cp .env.example .env
# Fill in: DATABASE_URL, GITHUB_TOKEN, BLAND_API_KEY, ANTHROPIC_API_KEY, ALERT_PHONE_NUMBER
# Run
cd sift
uvicorn backend.main:app --reload
# Open dashboard
open http://localhost:8000
| Method | Path | Description |
|---|---|---|
| GET | /signals | List all ingested signals |
| POST | /signals | Create signal manually |
| POST | /agent/scan | Run full analysis loop |
| POST | /agent/ingest | Pull fresh GitHub issues |
| GET | /agent/accuracy | Accuracy time series |
| GET | /agent/learned-rules | Current learned heuristics |
| POST | /feedback/{id} | Submit founder feedback |
| POST | /webhooks/bland-complete | Bland AI call callback |
Airbyte GitHub → signals table → LLM Analyzer (Overmind traced)
↓ severity > 7?
Bland AI Voice Call
↓ founder response
Learning Service
↓ generate rule
learned_rules table
↓ fed back into next analysis
Improved accuracy