Stringline AI Finance Pipeline — Phase 0

A Python pipeline that processes construction company month-end financial data through six sequential AI-assisted steps, producing a Word document pack and an Excel review checklist for human sign-off.

Built as a proof-of-concept for the Stringline Phase 0 brief. Runs against synthetic Business Central data that mirrors real OData response shapes.

---

Pipeline Architecture

BC Data (OData / JSON)
        │
        ▼
┌─────────────────┐
│ Step 1          │  Ingestion — read and normalise BC data
│ Confidence: det.│  Deterministic confidence based on data completeness
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Step 2          │  Classification — tag each transaction by cost type
│ Confidence: hyb.│  Deterministic (account codes) + Claude (ambiguous)
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Step 3          │  Reconciliation — transaction-level bank vs ledger
│ Confidence: det.│  Deterministic match scoring + Claude commentary
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Step 4          │  Anomaly Flagging — unusual amounts, CIS, margins
│ Confidence: AI  │  Claude self-assessed against prior month comparatives
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Step 5          │  Commentary — plain-English MD narrative
│ Confidence: AI  │  Claude self-assessed, structured 4-section output
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Step 6          │  Sector Intelligence — web search, subsector filtered
│ Confidence: AI  │  Claude web search tool, relevance-scored output
└────────┬────────┘
         │
         ▼
┌──────────────────────────────┐
│ output/month_end_pack.docx   │  Word document for SharePoint
│ output/review_checklist.xlsx │  Excel checklist for human reviewer
└──────────────────────────────┘

---

Confidence Scoring Approach

Each step returns a confidence score (0–100):

Step	Scoring Method	Rationale
1 — Ingestion	Deterministic — presence and population of expected data sets	We can measure completeness directly
2 — Classification	Hybrid — deterministic for account codes (95%), Claude self-assessed for ambiguous (0–100)	Account codes are reliable; descriptions need LLM judgement
3 — Reconciliation	Deterministic — weighted: exact match=100, partial=70, unmatched=0	Match quality is measurable
4 — Anomaly	Claude self-assessed — per anomaly, averaged for step score	We can't measure what we don't know; Claude estimates its own certainty
5 — Commentary	Claude self-assessed — extracted from end of response	Narrative quality requires self-evaluation
6 — Sector Intel	Relevance score — Claude-assessed per item, averaged	Source quality varies; relevance is LLM judgement

Items below LOW_CONFIDENCE_THRESHOLD (60%) are flagged for mandatory human review. Items below MEDIUM_CONFIDENCE_THRESHOLD (80%) are flagged as advisory.

---

Quick Start

# Install dependencies
pip install -r requirements.txt

# Set your Anthropic API key
export ANTHROPIC_API_KEY=your_key_here

# Run full pipeline against synthetic data
python pipeline.py

# Run only through Step 3 (ingestion, classification, reconciliation)
python pipeline.py --step 3

# Run against a custom data file
python pipeline.py --data path/to/your_bc_data.json

# Run individual steps for testing
python steps/step1_ingestion.py
python steps/step2_classification.py
python steps/step3_reconciliation.py
python steps/step4_5_anomaly_commentary.py
python steps/step6_sector_intelligence.py

---

Project Structure

stringline-pipeline/
├── pipeline.py                          # Main orchestrator
├── config.py                            # All configuration (client, thresholds, paths)
├── document_generator.py                # Word document (.docx) generation
├── review_checklist.py                  # Excel review checklist generation
├── requirements.txt
├── steps/
│   ├── step1_ingestion.py               # BC data ingestion and normalisation
│   ├── step2_classification.py          # Transaction classification (hybrid)
│   ├── step3_reconciliation.py          # Bank vs ledger reconciliation
│   ├── step4_5_anomaly_commentary.py    # Anomaly flagging + management commentary
│   └── step6_sector_intelligence.py    # Sector intelligence (web search)
├── data/
│   └── synthetic_bc_data.json          # Realistic synthetic BC data (Phase 0)
├── output/                              # Generated documents (gitignored)
└── logs/                               # Step-by-step output logs (gitignored)

---

Configuration

All client-specific configuration is in config.py:

CLIENT_NAME = "Stringline Testbed Construction Ltd"
REPORTING_MONTH = "May 2025"
SUBSECTOR = "modular/offsite manufacturer"   # Change per client in Phase 1

LOW_CONFIDENCE_THRESHOLD = 60    # Below this = mandatory review
MEDIUM_CONFIDENCE_THRESHOLD = 80 # Below this = advisory review

For Phase 1 multi-client support, each client gets its own config file. No pipeline code changes required — only config changes.

---

What This PoC Demonstrates

• Pipeline architecture: 6 independent, chainable steps with consistent interface
• Hybrid confidence scoring: deterministic where measurable, Claude self-assessed where not
• Transaction-level reconciliation: not balance-level — surfaces individual unmatched items
• Structured outputs: every step returns {"result": ..., "confidence": 0-100, "flags": [...]}
• Failure logging: every step output persisted to logs/ for Phase 0 learning
• Document generation: complete Word pack and Excel reviewer checklist
• Independently runnable steps: each step can be tested in isolation

---

Phase 0 Boundaries (per brief)

Not built and not intended:

• ❌ Web application or API
• ❌ Database or persistent data store
• ❌ Email or notification sending
• ❌ SharePoint integration (checklist is generated as .xlsx for manual upload)
• ❌ Automated scheduling
• ❌ Write operations to Business Central
• ❌ Multi-tenant architecture

These are Phase 1 decisions, made with data from Phase 0 failure logs.

---

Synthetic Data

data/synthetic_bc_data.json contains realistic construction finance data:

• 13 bank feed transactions (May 2025)
• 12 ledger entries with BC account codes
• 4 active projects with WIP and retention data
• Aged debtors and creditors
• Prior month comparatives (April 2025)
• Deliberate anomalies for pipeline testing:
  • 2 bank transactions cleared but not in ledger (BF012, BF013)
  • 1 ledger entry posted but not cleared in bank (GL012)
  • Missing CIS deduction on one subcontractor payment
  • Retention held but bank receipt outstanding

---

Phase 0 — June 2026 | Built by Virendra Vyas / Yug Solutions Ltd