Sift

An 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.

How It Works

1. Ingest: Pulls signals from GitHub issues, app reviews, Slack, support tickets (via Airbyte connectors)
2. Analyse: LLM scores severity (0-10), classifies type, decides whether to escalate
3. Act: Calls you via Bland AI with a briefing, captures your verbal response
4. Learn: Stores your feedback, generates rules, improves accuracy over time
5. Repeat: Every cycle makes the agent smarter at knowing what matters to YOU

The Self-Improving Loop

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.

Sponsor Tools Used

• Airbyte Agent Connectors — data ingestion from GitHub (21+ sources available)
• Ghost — Postgres database for agent memory, signals, and learned rules
• Bland AI + Norm — voice alerts when severity > 7, captures founder's verbal response
• Overmind — traces every LLM decision, evaluates quality, continuous optimization
• Auth0 — secure dashboard access (React SPA)

Setup

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

API Endpoints

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

Architecture

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