223Harsh/alfred-signal-0.8b
bank email signal classifier for autopilot
2 min read
The model is built on Qwen/Qwen3.5-0.8B and is designed to analyze incoming customer emails to a bank. It outputs a specific signal: hardship, dispute, deceased_notice, complaint, payment_promise, change_of_details, or no_signal. This allows the system to stop the overdue loan reminder autopilot in time and hand the conversation over to a human operator. The model is distilled from a larger 9B assistant from the ALFRED project and focuses on automating support in the financial sector.
Key features
- Based on the compact Qwen3.5-0.8B architecture.
- Accurately identifies 6 key types of customer signals.
- Distilled on synthetic data in just 3.1 minutes of training.
- Helps safely route complex or sensitive queries to live operators.
What you can do
- Integrate into a banking CRM system or email bot to automatically filter incoming emails from debtors.
- Use it to protect vulnerable customers from erroneous automated reminders.
- Test fast text classification in Russian and English for corporate automation tasks.
Example
The model takes the text of a customer email and returns one of the fixed tags (for example, hardship or dispute), preventing automated payment demands from being sent to customers in difficult life situations.
Read more
DetailsAccuracy on a held-out set (280 phrases): Qwen3.5-0.8B base — 35.0%, this model — 85.7%, 9B assistant (zero-shot) — 95.4%.
Trained on synthetic data (Faker) without using real customer information.
Created during the Agentic AI in companies hackathon (Turin, 2026) for the ALFRED ecosystem.










Comments