A proof-of-concept Streamlit application that generates structured financial credit memos from loan application data using the Anthropic Claude API.
A portfolio project demonstrating AI-assisted credit analysis in financial services workflows.
Credit analysts spend hours per application pulling data from CRM systems, structuring it, and writing a memo by hand into a Word template.
This tool cuts that to under 30 seconds:
- Paste an email — AI extracts all application fields automatically, or fill the form manually
- Click Generate Credit Memo
- Get a structured, analyst-quality memo ready for review
- Refine with natural language — "Upgrade the risk rating to High", "Add an FX risk bullet"
- Export to PDF — branded, formatted, ready for credit committee
The memo covers: Executive Summary · Borrower Profile · Loan Structure · Financial Analysis · Risk Assessment · Recommendation.
Or load the built-in sample (Kowalski Logistics) to see a full output in one click.
# 1. Clone
git clone https://github.com/MKorp7/AI-Credit-Memo-Generator
cd AI-Credit-Memo-Generator
# 2. Install dependencies
pip install -r requirements.txt
# 3. Set your API key
cp .env.example .env
# Edit .env and add your Anthropic API key
# 4. Run
streamlit run app.pyOpen http://localhost:8501 in your browser.
credit-memo-generator/
├── app.py # Streamlit UI (manual entry + email intake tabs)
├── generator.py # Claude API — memo generation, refinement, email extraction
├── pdf_export.py # ReportLab PDF generation
├── test_agents.py # Automated quality test suite (see below)
├── requirements.txt
├── .env.example
└── sample_data/
├── sample_application.json
└── sample_raport.pdf
generator.py builds a structured prompt from application data and sends it to claude-opus-4-5 with a strict system prompt enforcing a 6-section credit memo format. The model is instructed to:
- Follow a strict 6-section structure with no creative deviation
- Flag missing data explicitly rather than invent numbers
- Never produce specific numerical estimates (LTV, recovery rates) unless calculable from inputs
- Calculate DSCR from available inputs
- Give a clear
APPROVED / DECLINED / APPROVED WITH CONDITIONSrecommendation
app.py provides a split-pane Streamlit interface with two input modes:
- 📋 Manual Entry — structured form with a one-click demo loader
- 📧 Paste Email — paste a raw application email, AI extracts all fields automatically
After generation, an analyst refinement chat lets you update any section using plain English without re-filling the form. Output can be downloaded as .txt or formatted PDF.
The project includes a two-layer automated test suite (test_agents.py):
- Layer 1 — Deterministic: regex and math checks (all 6 sections present, DSCR in range, key input values echoed). Fast and free.
- Layer 2 — AI Judge: one Claude call evaluating hallucinations and internal consistency, returning structured JSON with specific quotes for any issues found.
- Layer 3 — Refinement smoke test: verifies that a follow-up instruction is applied correctly and nothing else breaks.
Latest test run — 18/18 checks passed:
==========================================================
Credit Memo Generator — Quality Test Suite
==========================================================
Generating memo...
· Generated in 24.4s (4,327 chars)
[ Layer 1 / Deterministic ]
✓ Section: EXECUTIVE SUMMARY
✓ Section: BORROWER PROFILE
✓ Section: LOAN REQUEST
✓ Section: FINANCIAL ANALYSIS
✓ Section: RISK ASSESSMENT
✓ Section: RECOMMENDATION
✓ Recommendation verdict is explicit
✓ DSCR in reasonable range → memo=1.57x expected≈1.54x
✓ Input echoed: Loan amount
✓ Input echoed: Company name
✓ Input echoed: Registration no.
[ Layer 2 / AI Judge ]
· Sending memo to AI judge for hallucination + consistency check...
· AI judge responded in 2.8s
✓ No hallucinations detected → 0 found
✓ No consistency issues → 0 found
✓ Overall quality: Good
· Judge summary: The memo accurately reflects all application data with no
hallucinations or contradictions; all calculations are mathematically derived
from inputs and all risk assessments are appropriately qualified.
[ Layer 3 / Refinement ]
· Instruction: "Change the risk rating to High and add one bullet about client concentration risk."
· Refined in 15.7s
✓ All 6 sections survive refinement
✓ Risk rating updated to High
✓ Loan amount preserved after refinement
✓ Company name preserved after refinement
==========================================================
18 passed 0 failed / 18 checks
✓ All checks passed.
==========================================================
Run it yourself:
python test_agents.py| Extension | Effort | Value |
|---|---|---|
| CRM integration (HubSpot / Salesforce) | Medium | High |
| Analyst feedback loop for prompt fine-tuning | Medium | High |
| Batch processing from Excel / CSV upload | Low | Medium |
| Multi-language output (EN / PL) | Low | Medium |
| Audit trail & memo versioning | High | High |
| Component | Technology |
|---|---|
| UI | Streamlit |
| LLM | Claude claude-opus-4-5 (Anthropic) |
| PDF generation | ReportLab |
| API client | anthropic Python SDK |
| Config | python-dotenv |
- Python 3.10+
- Anthropic API key — get one here
- No database or cloud infrastructure required — runs fully locally
MIT