Zenit

AI-assisted customer service for retail banks: an AI agent answers customers and hands off to human advisors with full context.

Features

• AI customer agent (David) in Spanish and Portuguese: answers credit card and savings account questions, registers complaints and checks their status.
• Intent and guardrail classification with Jev; keyword rules as fallback.
• Policy outside the model: handoff rules are deterministic code, the LLM only reads language and writes the reply.
• Secure by design: the customer's identity comes from the session, never from the chat text, and every data read is fixed SQL filtered by that customer.
• Human handoff with context: David collects and verifies the case first, then stops replying until an advisor returns the chat.
• Advisor console: an inbox of human cases and a separate view of the chats David is handling.
• Turn queue in the back: messages that arrive while the agent is answering wait in Lakebase and are processed in order.
• Governed data: a bronze → silver → gold pipeline on Databricks, copied to Lakebase as read-only synced tables.

Development

The platform has three components that run independently:

Agent is a LangGraph graph served at POST /invocations.

cd agent && uv run start-server                 # → :8000 (needs agent/.env)
cd agent && uv run --group dev pytest tests -q  # tests

Back is an Express server: API, demo login, chat history and the turn queue. It serves the built front.

cd front && npm install && npm run build
cd back && npm install && npm run build && npm run start   # → :3000

Front is the React chat UI and advisor console. For hot reload, run npm run dev in back (port 3001) and in front (port 3000).

back/.env needs DATABRICKS_CONFIG_PROFILE and API_PROXY=http://localhost:8000/invocations.

Deployment

The back (with the front built in) and the agent run on AWS App Runner (us-west-2). Chats and bank data stay in Lakebase, and the LLM runs on Databricks Model Serving.

Data (once, or after new files)

Requirements: Databricks CLI.

python data/generate_dummy_data.py
databricks fs cp -r --overwrite data/dummy_output dbfs:/Volumes/workspace/bank_bronze/landing
cd data && databricks bundle deploy && databricks bundle run bank_data_pipeline

The gold and silver tables are copied to Lakebase with the scripts in back/scripts/bank-ro/.

Back and agent on AWS

Requirements: an AWS session, Docker, jq, openssl.

cd back && scripts/aws/deploy.sh    # build, push and roll out the back
cd agent && scripts/aws/deploy.sh   # build, push and roll out the agent

First-time setup (ECR, IAM roles, secrets, services): back and agent.

Docs

Grouped by the hackathon's evaluation dimensions. The evidence for each one, with its numbers, is in Evaluation criteria.

Technical judgment

Architecture, trade-offs, reliability, safety and production readiness.

• Architecture on AWS + Databricks: flow, services and decisions.
• Path to production: what changes for a real bank, and why (CloudFront, LLM capacity, identities, login).
• Agent and back limits: who stores what.
• Agent architecture: shared core, playbooks and human review.

AI engineering

Backend, frontend, system integration and deployment.

• Agent: use cases and technical decisions.
• Attention flow: conversation stages and handoff to advisors.
• Evaluation of David: 60 test cases × 3 runs, before and after, local and production.
• Deployment with CDK: the AWS stack.
• Demo runbook.

Data engineering

Data quality, pipelines, preparation and reproducibility.

• Data pipeline: data contract and bronze/silver/gold.
• Pipeline validation.
• Bank data in Lakebase: synced tables and access.

Machine learning

Modeling approach, evaluation, baselines and performance.

• Fraud model: features, training and evaluation.
• Fraud model proposal.
• Fraud data reference.

Data analytics

Metrics, insights, visualization and decision support.

• Admin fraud intelligence dashboard: data insights, complaint prediction coverage and model evaluation at /fraud.
• Admin retention dashboard: customer follow-up signals at /retention.
• Analytics serving: how the dashboards read and cache their data.

License

Built for the Factored AI & Data Hackathon 2026. Derived from the Databricks banking-agent-accelerator and modified by the team. See LICENSE.md and NOTICE.md.