Research Prompts: Tier 2 (Secondary Priority)

Overview

These prompts address important but not critical capabilities. Execute after Tier 1 research is complete.

---

PROMPT 6: NeuralProphet Cash Flow Forecasting Pipeline

Context

Accurate cash flow forecasting is the #1 request from CFOs. NeuralProphet offers 55-92% accuracy improvement over Prophet.

Research Objective

Design a production forecasting pipeline with confidence intervals and scenario modeling.

Prompt

You are a Time-Series ML Engineer specializing in financial forecasting for FP&A applications.

**TASK**: Design a complete cash flow forecasting pipeline using NeuralProphet with explainability features.

**REQUIREMENTS**:

1. **Forecasting Model**:
   - 13-week rolling cash flow forecast (standard treasury horizon)
   - Seasonality detection (weekly, monthly, quarterly, annual)
   - Holiday effects (configurable per country)
   - Exogenous regressors: AR aging, AP aging, pipeline stage
   - Confidence intervals: P10/P50/P90

2. **Training Pipeline**:
   - Minimum 24 months historical data
   - Automatic hyperparameter tuning (Optuna)
   - Cross-validation with expanding window
   - Model versioning with MLflow
   - Retraining trigger on accuracy degradation

3. **Scenario Modeling**:
   - Best case: Pipeline close rates +20%
   - Base case: Historical patterns
   - Worst case: Economic stress scenario
   - User-defined custom scenarios

4. **Explainability**:
   - Component breakdown (trend, seasonality, exogenous)
   - Driver attribution: "Revenue forecast +15% due to Q4 seasonality"
   - Confidence level explanation
   - Data quality impact on forecast reliability

5. **Integration Points**:
   - Input: PostgreSQL GL data, CRM pipeline, AR/AP aging
   - Output: Forecast table with version tracking
   - API: FastAPI endpoint for on-demand forecasting
   - Visualization: Plotly interactive charts

6. **Deliverables**:
   - Training pipeline (Python + Dagster)
   - NeuralProphet model configuration
   - Scenario engine with parameter injection
   - FastAPI forecast service
   - Streamlit dashboard for visualization
   - Benchmark against Prophet/ARIMA baselines

**OUTPUT FORMAT**: Production pipeline with MLOps best practices.

**CONSTRAINTS**:
- Must handle missing data gracefully
- Support multi-entity forecasting (consolidation)
- Forecast generation <30s for 5-year history

CODITECT Application

"AI-powered forecasting" becomes a premium feature for FP&A automation pack.

---

PROMPT 7: Month-End Close Automation Workflow

Context

Month-end close takes 10-15 days at most companies. AI can reduce this to 3-5 days.

Research Objective

Design an agent-based month-end close workflow with automation and human checkpoints.

Prompt

You are a Finance Operations Architect specializing in month-end close optimization for mid-market companies.

**TASK**: Design a complete month-end close automation workflow using CODITECT agent patterns.

**REQUIREMENTS**:

1. **Close Process Steps** (in order):
   - Pre-close checklist generation
   - Bank reconciliation (auto-match statements to GL)
   - Intercompany reconciliation
   - Accrual calculations (rent, payroll, utilities)
   - Revenue recognition review
   - AP/AR cutoff verification
   - Journal entry processing
   - Trial balance generation
   - Variance analysis
   - Financial statement preparation
   - Close certification

2. **Automation Targets**:
   - Bank reconciliation: 95% auto-match (ML-based)
   - Accrual calculations: Pattern-based with anomaly detection
   - Journal entries: Auto-generated, human-approved
   - Variance analysis: AI-generated explanations

3. **Compliance Gates** (mandatory human approval):
   - Material adjustments >$10k
   - Non-routine journal entries
   - Revenue recognition judgments
   - Intercompany eliminations
   - Close certification sign-off

4. **Multi-Entity Coordination**:
   - Subsidiary close → Parent consolidation
   - Inter-entity dependency tracking
   - Close status dashboard (entity x step matrix)
   - Bottleneck identification and escalation

5. **Audit Trail**:
   - Every step logged with timestamp
   - Supporting documentation linkage
   - AI decision reasoning captured
   - Change history for all entries

6. **Deliverables**:
   - LangGraph workflow definition
   - Agent specifications for each automation task
   - Checkpoint approval UI wireframes
   - Close calendar Gantt chart generator
   - KPI dashboard (days to close, automation rate)
   - Playbook for implementing at new client

**OUTPUT FORMAT**: Complete implementation guide with deployment timeline.

**CONSTRAINTS**:
- Must integrate with existing GL systems (read-only initially)
- Parallel processing where dependencies allow
- Graceful degradation if automation fails (fallback to manual)

CODITECT Application

This becomes the flagship feature of the FP&A Automation Pack—tangible ROI demonstration.

---

PROMPT 8: Open Finance Brazil Integration

Context

BACEN mandates Open Finance participation for Brazilian financial institutions. This creates a connector opportunity.

Research Objective

Design a BACEN-compliant Open Finance integration for real-time bank data access.

Prompt

You are a Brazilian Fintech Architect specializing in Open Finance (Open Banking) integrations under BACEN regulation.

**TASK**: Design a complete Open Finance Brazil integration for FP&A data ingestion.

**REQUIREMENTS**:

1. **Open Finance Capabilities**:
   - Consent management (request, track, revoke)
   - Account balance retrieval (real-time)
   - Transaction history (up to 12 months)
   - Credit operations visibility
   - Investment positions
   - Payment initiation (Pix API)

2. **BACEN Compliance**:
   - TPP (Third-Party Provider) registration requirements
   - Consent flow per BACEN specs (OAuth 2.0 + FAPI)
   - Data retention and deletion policies
   - Security standards (mTLS, JWS signatures)
   - Incident reporting procedures

3. **Integration Architecture**:
   - OAuth 2.0 authorization server
   - Webhook receiver for bank notifications
   - Data normalization layer (bank-agnostic schema)
   - Multi-bank aggregation view
   - Rate limiting and retry logic

4. **FP&A Value-Add**:
   - Automatic bank reconciliation from Open Finance data
   - Cash position dashboard (multi-bank)
   - Transaction categorization (ML-based)
   - Fraud detection signals
   - Treasury forecasting inputs

5. **Supported Banks** (Phase 1):
   - Banco do Brasil
   - Itaú Unibanco
   - Bradesco
   - Santander Brasil
   - Caixa Econômica Federal
   - Nubank, Inter, C6 (digital banks)

6. **Deliverables**:
   - BACEN compliance checklist
   - Authorization server implementation
   - Bank connector adapters (per institution)
   - Data model for normalized transactions
   - Consent management UI
   - Integration test suite with sandbox environments

**OUTPUT FORMAT**: Compliance-ready implementation with BACEN documentation.

**CONSTRAINTS**:
- Must support both legacy (COBOL) and modern bank APIs
- LGPD compliance for personal financial data
- Disaster recovery for financial data (RTO <4hrs)

CODITECT Application

Creates 12-18 month competitive moat in Brazilian market—no competitors have this integration.

---

PROMPT 9: PostgreSQL Multi-Tenant RLS Architecture

Context

SaaS financial platforms require bulletproof tenant isolation with excellent query performance.

Research Objective

Design a PostgreSQL multi-tenant architecture using Row-Level Security with performance optimization.

Prompt

You are a Database Architect specializing in multi-tenant SaaS platforms handling sensitive financial data.

**TASK**: Design a complete PostgreSQL multi-tenant architecture using Row-Level Security (RLS).

**REQUIREMENTS**:

1. **Tenant Isolation Model**:
   - Schema: Shared schema, tenant_id column on all tables
   - RLS policies enforcing tenant isolation at row level
   - Connection pooling with tenant context injection
   - Cross-tenant queries explicitly disabled (except admin)

2. **Performance Optimization**:
   - Composite indexes including tenant_id
   - Partial indexes for active tenants
   - Table partitioning by tenant_id for large tables
   - Connection pooler (PgBouncer/Pgcat) configuration
   - Query plan caching considerations

3. **Financial Data Model**:
   - Core tables: tenants, users, entities, accounts, gl_transactions, budgets, forecasts
   - Audit tables with RLS
   - Multi-currency support (base + transaction currency)
   - Fiscal calendar flexibility (non-calendar years)

4. **Security Layers**:
   - RLS policies (tenant isolation)
   - Column-level encryption (PII fields)
   - TDE (Transparent Data Encryption) at rest
   - SSL/TLS in transit
   - Audit logging (pgaudit)

5. **Operational Considerations**:
   - Tenant onboarding automation
   - Tenant data export/deletion (GDPR/LGPD)
   - Performance monitoring per tenant
   - Noisy neighbor detection and throttling
   - Backup/restore isolation

6. **Deliverables**:
   - Complete DDL with RLS policies
   - Index strategy document
   - Partitioning implementation
   - Connection pooler configuration
   - Tenant management API
   - Performance benchmark results (10k users, 100 tenants)
   - Migration scripts from single-tenant

**OUTPUT FORMAT**: Production DDL with operational documentation.

**CONSTRAINTS**:
- Target: 10k+ concurrent users, 1000+ tenants
- P95 query latency <100ms for dashboard queries
- Support for PostgreSQL 16+ features

CODITECT Application

Foundation for all CODITECT SaaS deployments—ensures data isolation for regulated customers.

---

PROMPT 10: Variance Analysis Natural Language Generation

Context

CFOs want AI to explain budget variances in plain English with actionable insights.

Research Objective

Design an NLG system that produces CFO-ready variance explanations.

Prompt

You are an AI/NLP Engineer specializing in natural language generation for financial reporting.

**TASK**: Design an NLG system for automated budget variance explanations.

**REQUIREMENTS**:

1. **Input Analysis**:
   - Budget vs. actual by account/department/entity
   - Variance amount ($ and %)
   - Transaction-level drill-down data
   - Historical variance patterns
   - Business context (seasonality, known events)

2. **Explanation Components**:
   - Headline: "Marketing exceeded budget by $50k (15%)"
   - Root cause: "Driven by unplanned Q2 conference ($30k) + agency fees ($20k)"
   - Trend context: "This is 3rd consecutive month of overspend"
   - Benchmark: "Industry average marketing spend is 12% of revenue; we're at 18%"
   - Recommendation: "Consider reallocating Q3 budget or seeking approval for increase"

3. **Tone Calibration**:
   - CFO-level: Strategic, high-level, action-oriented
   - Controller-level: Detailed, accounting-accurate, reference-heavy
   - Board-level: Executive summary, key metrics only

4. **Quality Controls**:
   - Factual accuracy validation against source data
   - Numerical consistency checks
   - Hallucination detection (claims not supported by data)
   - Confidence scoring with threshold for human review

5. **Templates & Personalization**:
   - Company-specific terminology injection
   - Historical narrative style matching
   - Configurable detail levels
   - Multi-language support (English, Portuguese)

6. **Deliverables**:
   - Prompt engineering templates for variance analysis
   - LLM chain for data → insight → narrative
   - Quality validation pipeline
   - A/B testing framework for narrative quality
   - User feedback collection for continuous improvement
   - Integration with LangGraph workflow (PROMPT 1)

**OUTPUT FORMAT**: NLG system with quality assurance framework.

**CONSTRAINTS**:
- Must cite specific transactions in explanations
- No invented data or numbers
- Latency <10s for single variance explanation
- Support for batch processing (monthly report generation)

CODITECT Application

Differentiates CODITECT AI from competitors—explainability is the key to finance user trust.

---

PROMPT 11: Dagster Asset-Centric Data Orchestration

Context

Data lineage tracking is essential for audit compliance. Dagster provides superior lineage vs. Airflow.

Research Objective

Design a Dagster orchestration layer for FP&A data pipelines with full lineage tracking.

Prompt

You are a Data Engineering Architect specializing in modern data stack orchestration for regulated industries.

**TASK**: Design a Dagster-based data orchestration layer for FP&A pipelines.

**REQUIREMENTS**:

1. **Asset Definitions**:
   - Raw: ERP extracts (QuickBooks, NetSuite, Xero, etc.)
   - Staged: Cleaned, validated source data
   - Intermediate: COA-harmonized transactions
   - Marts: Unified GL, budget comparison, forecasts
   - Reports: DRE, P&L, cash flow statements

2. **Orchestration Patterns**:
   - Scheduled: Nightly ELT, weekly forecast refresh
   - Event-triggered: Real-time bank sync, invoice processing
   - Manual: Ad-hoc report generation, scenario modeling
   - Dependency-aware: Proper sequencing of transformations

3. **Observability**:
   - Asset lineage visualization (end-to-end)
   - Data quality metrics per asset
   - Freshness monitoring with SLA alerting
   - Failure impact analysis (downstream dependencies)

4. **dbt Integration**:
   - Software-defined assets from dbt models
   - Unified lineage across Python + SQL
   - Test execution as asset materialization dependency
   - Documentation sync to Dagster catalog

5. **Compliance Features**:
   - Execution logs for audit trail
   - Data versioning snapshots
   - Access control per asset group
   - Retention policy enforcement

6. **Deliverables**:
   - Dagster project structure
   - Asset definitions for complete FP&A pipeline
   - Sensor configurations for event triggers
   - Schedule definitions
   - CI/CD pipeline (GitHub Actions)
   - Operational dashboard (Dagster UI customization)
   - Runbook for common operational tasks

**OUTPUT FORMAT**: Production Dagster deployment with CI/CD.

**CONSTRAINTS**:
- Must handle 10M+ rows daily throughput
- Parallel execution for independent assets
- Graceful handling of source system outages

CODITECT Application

Replaces manual data pipeline management; provides audit-ready lineage documentation.

---

PROMPT 12: Brazilian ERP Integration Pack (Totvs/Omie/Conta Azul)

Context

Brazilian market requires native ERP integrations that no competitor offers.

Research Objective

Design integration connectors for the top 3 Brazilian ERPs with SPED/NF-e compatibility.

Prompt

You are a Brazilian ERP Integration Specialist with expertise in Totvs Protheus, Omie, and Conta Azul APIs.

**TASK**: Design integration connectors for Brazilian ERPs with local compliance features.

**REQUIREMENTS**:

1. **Totvs Protheus**:
   - REST API integration (newer versions)
   - AdvPL function calls (legacy)
   - Standard Totvs tables: SE1 (AR), SE2 (AP), CT2 (GL)
   - SPED fiscal integration
   - Multi-branch support

2. **Omie**:
   - REST API (modern, well-documented)
   - Webhook support for real-time sync
   - Data model: contas_receber, contas_pagar, lancamentos
   - NF-e/NFS-e integration
   - Multi-company support

3. **Conta Azul**:
   - REST API integration
   - OAuth 2.0 authentication
   - Data model: vendas, compras, lancamentos_bancarios
   - Bank sync reconciliation
   - Accountant access (contador feature)

4. **Common Features**:
   - Incremental sync (last_modified timestamp)
   - Error handling with automatic retry
   - Data validation (CPF/CNPJ, fiscal codes)
   - COA mapping to unified schema
   - Multi-currency (BRL primary, USD for GAAP)

5. **Brazilian Tax Compliance**:
   - SPED Contábil export
   - SPED Fiscal data extraction
   - NF-e XML parsing
   - ICMS/ISS/PIS/COFINS classification
   - DCTF supporting data

6. **Deliverables**:
   - Airbyte custom connector (Python)
   - dbt models for Brazilian accounting patterns
   - Tax classification ML model (auto-NCM mapping)
   - Integration test suite with sandbox accounts
   - Setup guide in Portuguese
   - Support escalation playbook

**OUTPUT FORMAT**: Production connectors with Brazilian market documentation.

**CONSTRAINTS**:
- Must handle Brazilian date/number formats
- Support for CPF/CNPJ masking (LGPD)
- Accountant user workflows (separate permissions)

CODITECT Application

Creates unassailable competitive moat in R$500B Brazilian SMB market.

---

Execution Checklist

Prompt	Priority	Est. Time	Dependency	Owner
PROMPT 6: NeuralProphet	P1	4 hrs	PROMPT 5	Data Science
PROMPT 7: Month-End Close	P1	5 hrs	PROMPT 1	Product
PROMPT 8: Open Finance Brazil	P1	6 hrs	None	Integrations
PROMPT 9: PostgreSQL RLS	P1	4 hrs	None	Platform
PROMPT 10: NLG Variance	P1	3 hrs	PROMPT 4	AI Team
PROMPT 11: Dagster	P2	4 hrs	PROMPT 5	Data Eng
PROMPT 12: Brazilian ERPs	P2	6 hrs	PROMPT 5	Integrations

Total Research Investment: ~32 hours Expected Output: 7 production-ready specifications