TCI Shield

PASSWORD TO ACCESS: AmazNova@2026tci_PAB

Inspiration

With extensive experience in trade credit insurance underwriting, I have seen first-hand how slow, paper-heavy, and inaccessible the traditional underwriting process can be. A small business owner seeking cover for their buyers must fill in lengthy forms, wait days for a decision, and often speak to multiple people before a policy is issued. The Amazon Nova AI Hackathon gave me the opportunity to ask a simple question: what if a business owner could simply speak or chat naturally and walk away with a trade credit insurance policy - all within minutes? That question became TCI Shield.

What it does

TCI Shield is an AI-powered trade credit insurance underwriting agent that conducts a real-time conversation with a customer - by voice or text - collects all necessary underwriting information, scores the risk automatically, and issues a policy in minutes. Customers can speak to Alex via the Voice Agent powered by Amazon Nova 2 Sonic, or chat with Alex via the Text Chat Agent powered by Amazon Nova Lite and AWS Strands Agents. Both agents ask the right questions, understand the answers, and hand off to the underwriting engine which calculates a risk score and generates a policy in real time. Customers can also upload financial statements which Alex reads automatically using Amazon Nova Multimodal, extracting all 11 required financial figures instantly - eliminating manual data entry entirely. No forms. No waiting. No back and forth.

How I built it

The system is built in three layers:

1. Conversation Layer— Customers can apply through two AI-powered interfaces. The Voice Agent uses Amazon Nova 2 Sonic for real-time bidirectional voice conversation via the aws_sdk_bedrock_runtime streaming SDK - the customer speaks naturally and Alex responds with voice. The Text Chat Agent uses Amazon Nova Lite via AWS Strands Agents - the customer types naturally and Alex responds with text. Both agents collect the same underwriting information and issue a policy at the end.

2. Intelligence Layer — AWS Strands Agents with Amazon Nova Lite orchestrates the underwriting workflow. Six tools are called automatically based on the conversation: business intake, buyer assessment, financial data collection, risk scoring, policy generation, and policy issuance.

3. Underwriting Layer — A purpose-built risk scoring engine applies real trade credit underwriting logic across six dimensions: industry risk, buyer country risk, financial health, payment terms, portfolio concentration, and loss history. The final weighted score maps to a policy tier and premium rate. $${Risk Score} = \lfloor w_1 \cdot S_{\text{industry}} + w_2 \cdot S_{\text{country}} + w_3 \cdot S_{\text{financial}} + w_4 \cdot S_{\text{terms}} + w_5 \cdot S_{\text{concentration}} + w_6 \cdot S_{\text{experience}} \rfloor$$

Policies and customer data are stored in AWS DynamoDB. The entire stack runs on AWS.

Challenges I ran into

1. Experimental SDK Compatibility The aws_sdk_bedrock_runtime package for Nova Sonic is still experimental. Between versions, module paths changed, parameter names changed, and constructor signatures changed - none of which was reflected in the documentation. Each issue required inspecting the installed package directly using pkgutil and help().

2. Echo Feedback Loop Nova Sonic's microphone was picking up Alex's own voice through the speaker, causing Alex to respond to himself in a continuous loop. The fix was to track an alex_is_speaking flag and mute microphone capture while audio output was playing.

3. Encoding Domain Knowledge The hardest challenge was not technical - it was translating real underwriting judgement into explicit, computable rules. What makes a buyer high risk? When should a policy be declined versus rated up? These decisions that happen intuitively in a human underwriter had to become weighted scores, thresholds, and validation rules.

Accomplishments that I am proud of

1. A full text chat underwriting agent powered by Amazon Nova Lite and AWS Strands Agents with 9 tools and Pydantic validation across 30+ fields
2. A risk scoring engine built on genuine trade credit insurance underwriting logic - not a generic model - covering 238 countries and 14 risk factors
3. Automated financial document extraction using Amazon Nova Multimodal supporting 5 file formats with 100% accuracy across 11 financial figures
4. A 4-page Streamlit UI deployed live on Hugging Face Spaces
5. Resolving multiple undocumented SDK breaking changes to get Nova Sonic running in production
6. A system that makes trade credit insurance accessible to any business owner - no forms, no jargon, just a conversation

What I learned

1. How to build multi-tool agents using AWS Strands Agents that reason over a live conversation context
2. How Amazon Nova 2 Sonic's bidirectional streaming session management and event-driven architecture works
3. How system prompt structure directly controls agent behaviour and conversation flow
4. How to manage concurrent asyncio tasks for simultaneous audio capture, processing, and playback
5. How to debug and navigate undocumented experimental AWS SDK packages using pkgutil and help()
6. How to encode domain expertise into deterministic, auditable underwriting logic

What's next for TCI Shield

1. Connect the voice agent directly to the Strands underwriting engine for full end-to-end policy issuance by voice
2. Integrate live Companies House and OpenCorporates APIs for real-time buyer credit assessment
3. Build a React frontend with FastAPI backend for a production-grade web application
4. Expand to export credit insurance and surety bond underwriting
5. Add Nova Act for broker portal UI automation and workflow management

Built With

• amazon-nova-2-sonic
• amazon-nova-lite
• amazon-nova-multimodal-(document-extraction)
• asyncio
• aws-bedrock
• aws-cli
• aws-dynamodb
• aws-strands-agents
• boto3
• pyaudio
• pycharm
• pydantic
• python