PABPAT/TCI_Shield

1# AI Trade Credit Insurance Agent -- Project Metrics Tracker2# Last Updated: 21 Feb 20263# Status: Work in Progress -- figures to be replaced with real measurements on completion4 5---6 7## HOW TO USE THIS FILE8- Each section tracks metrics for one module9- [ESTIMATED] = placeholder, replace with real figure after testing10- [MEASURED] = confirmed real figure from actual testing11- Final resume bullet points at the bottom -- update as modules complete12 13---14 15## MODULE 1 -- UNDERWRITING ENGINE16# Status: Complete17# File: underwriting_engine.py18 19### Confirmed Figures (count from code)20- Countries assessed              : 238  [MEASURED -- count of COUNTRY_RISK keys]21- Risk factors (Stream 1)         : 7    [MEASURED -- industry, trade, country, payment, concentration, loss ratio, maturity]22- Financial ratios (Stream 2)     : 7    [MEASURED -- current ratio, TOL/TNW, bad debt%, TNW%, debtor days, creditor days, capital adequacy]23- Industry sectors                : 11   [MEASURED -- count of INDUSTRY_RISK keys]24- Risk tiers                      : 4    [MEASURED -- Standard, Enhanced, High Risk, Declined]25- Decline rules                   : 3    [MEASURED -- score>=75, negative TNW, industry off-cover breach]26- Premium range                   : 1% to 6% of credit sales  [MEASURED]27- Weighted streams                : 3    [MEASURED -- Business 30%, Financials 40%, Buyers 30%]28 29### Performance Figures (to be measured)30- Average underwriting time       : 0.013 ms (avg over 1000 runs)  [MEASURED -- 21 Feb 2026]31- Accuracy vs manual underwriting : [TO BE MEASURED] -- compare 10 sample cases with expert decision32 33### Business Baseline (industry standard)34- Traditional underwriting time   : 5-7 business days  [INDUSTRY STANDARD -- cite Euler Hermes]35- Manual financial analysis time  : 2-3 hours per application  [INDUSTRY STANDARD]36- Market size                     : $12B+ globally  [CITE -- ICISA Annual Report]37 38---39 40## MODULE 2 -- DATABASE41# Status: Complete42# File: database.py43 44### Confirmed Figures45- DynamoDB tables                 : 2    [MEASURED -- tci_customers, tci_buyers]46- Data entities tracked           : 4    [MEASURED -- customers, buyers, policies, claims]47- ID formats defined              : 4    [MEASURED -- CUST, POL, CLM, REG]48 49---50 51## MODULE 3 -- STRANDS AGENT52# Status: Complete53# File: tci_agent.py, models.py54 55### Confirmed Figures56- Number of agents                : 1 orchestrator  [MEASURED]57- Number of tools                 : 8  [MEASURED -- get_progress, set_buyer_count, collect_business_info, collect_buyer_info, collect_financial_data, run_underwriting, generate_policy_options, issue_policy]58- Conversation steps              : 20  [MEASURED]59- Pydantic models                 : 4  [MEASURED -- BusinessInfo, Buyer, BuyerInfo, FinancialData]60- Validation fields               : 30+ across all models  [MEASURED]61- Tool guards implemented         : all 8 tools have validation  [MEASURED]62- Progress tracking flags         : 5  [MEASURED -- business, buyers, financials, underwriting, policy]63- Issues resolved during testing  : 23 issues (Issues 008-031)  [MEASURED]64 65### Performance Figures66- Underwriting engine time        : 0.013ms  [MEASURED]67- Nova API round trip             : 2-5 seconds per call  [MEASURED]68- Full conversation end-to-end    : [TO BE MEASURED] -- time full test run69- Data collection reliability     : All fields collected correctly with Pydantic validation  [MEASURED]70- Conversation order flexibility  : Agent collects all data correctly regardless of order  [MEASURED]71 72---73 74## MODULE 4 -- NOVA SONIC VOICE75# Status: Not Started76 77### Target Figures78- Voice to policy time            : [ESTIMATED under 10 minutes] -- measure end to end79- Speech recognition accuracy     : [TO BE MEASURED] -- test with 10 sample conversations80- Languages supported             : [TARGET 1 -- English] -- expand post hackathon81 82---83 84## MODULE 5 -- NOVA MULTIMODAL DOCUMENT ANALYSIS85# Status: Complete86# File: document_extractor.py87 88### Confirmed Figures89- Financial ratios extracted      : 11   [MEASURED]90- Document types supported        : 5    [MEASURED -- PDF, DOCX, XLSX, CSV, TXT]91- Extraction accuracy             : 100% on test file (11/11 fields correct)  [MEASURED]92- Extraction time                 : 1.05 seconds  [MEASURED -- 21 Feb 2026]93- Pydantic validation             : Applied after extraction  [MEASURED]94 95### Performance96- Document size tested            : 483 bytes (TXT)  [MEASURED]97- Nova response size              : 331 characters  [MEASURED]98- Fields extracted correctly      : 11 of 11  [MEASURED]99- Validation passed               : Yes  [MEASURED]100 101---102 103## MODULE 6 -- POLICY COMPARISON UI104# Status: Not Started105 106### Target Figures107- Policy tiers displayed          : [TARGET 3 -- Standard, Enhanced, High Risk]108- Data points per policy          : [TO BE MEASURED] -- count fields shown in UI109 110---111 112## MODULE 7 -- PDF POLICY ISSUANCE113# Status: Not Started114 115### Target Figures116- Policy document sections        : [TO BE MEASURED] -- count sections in template117- Time to generate PDF            : [TO BE MEASURED] -- measure in testing118 119---120 121## FINAL COMBINED METRICS122# To be completed when all modules are done123 124### Core Performance125- End-to-end time (voice to PDF)  : [TO BE MEASURED]126- Traditional process time        : 5-7 business days  [INDUSTRY STANDARD]127- Time reduction                  : [CALCULATE -- ((traditional - new) / traditional) x 100]%128 129### Technical Depth130- Total risk factors              : [SUM all factors across all modules]131- Countries covered               : 238132- Financial ratios automated      : 7133- Nova models used                : [TARGET 3 -- Sonic, Lite, Multimodal]134- AWS services used               : [COUNT on completion]135 136### Business Impact137- Target market                   : 5.5M SMEs in UK  [CITE -- Companies House stats]138- Market size                     : $12B+ globally  [CITE -- ICISA]139- Cost of traditional broker      : 0.25-1.5% of turnover  [INDUSTRY STANDARD]140 141---142 143## RESUME BULLET POINTS144# Update versions as modules complete145 146### Current Version (Underwriting Engine complete -- 21 Feb 2026)147"Built a 3-stream weighted underwriting engine for trade credit insurance,148scoring 14 risk factors across 238 countries with automated financial ratio149analysis in 0.013ms average -- compared to 5-7 business day traditional process.150Foundation of an AI agent targeting a $12B+ global market."151 152### Target Version (All modules complete -- replace [X] with real figures)153"Built an end-to-end AI trade credit insurance platform reducing policy154issuance from 5-7 days to [X] minutes, using Amazon Nova Sonic for voice155intake, Nova Multimodal for automated financial analysis ([X] ratios extracted),156and a 3-stream underwriting engine covering 238 countries and [X] risk factors157-- deployed on AWS using Strands Agents, Lambda, DynamoDB and S3."158 159---160 161## HOW TO MEASURE KEY FIGURES162 163### Measure underwriting engine speed:164    import time165    start = time.time()166    result = calculate_risk_score(profile)167    elapsed = time.time() - start168    logging.info(f"Underwriting time: {elapsed:.3f} seconds")169 170### Measure end-to-end time:171    Record time from first voice input to PDF download link generated.172    Run 5 test cases and take average.173 174### Measure extraction accuracy:175    Take 10 real financial statements.176    Extract manually -- record figures.177    Run through Nova Multimodal -- record figures.178    Accuracy = (matching fields / total fields) x 100