Finance Copilot

The description states that Finance Copilot is a tool that uses:

• GPT-5 to plan workflows and write narratives
• Codex to generate Python scripts for exact calculations
• A sandboxed execution environment to run those scripts on real data files (CSVs, Excel, PDFs)
• Server-Sent Events (SSE) to stream the process live in a frontend UI

The system claims to avoid hallucination by:

1. Using GPT-5 for reasoning and narrative
2. Having Codex write code that computes exact figures from full datasets
3. Running that code in a sandboxed subprocess
4. Feeding verified outputs back into GPT-5 for final analysis

It also supports voice interaction via WebRTC, with an ephemeral key and injected context.

Not evidenced No evidence of actual revenue, customers, or real-world usage beyond the author's own account.

The author positions Finance Copilot as a solution to inefficiencies in finance teams who:

• Manually extract data from PDFs and spreadsheets
• Recompute ratios by hand
• Write repetitive analyses for different audiences

The core claim is that existing tools either force users to do the work themselves or rely on chatbots that hallucinate numbers. Finance Copilot aims to be a “trust loop” where:

1. GPT-5 plans and reasons
2. Codex computes exact figures
3. The results are verified before being fed back into GPT-5

This is described as a way to avoid the problem of LLMs hallucinating arithmetic.

Inference The positioning reflects an attempt to differentiate from generic AI financial assistants by emphasizing accuracy and auditability.

Not evidenced No evidence of market research, user interviews, or competitive analysis beyond the author’s own claims.

The description states that Finance Copilot is aimed at finance teams, particularly those who:

• Spend days analyzing data manually
• Need to produce reports for multiple audiences (e.g., board, investors)
• Work with large datasets in formats like PDFs and spreadsheets

It also targets users who want:

• Board-ready reports in multiple formats (PDF, Word, Excel, CSV)
• Voice interaction grounded in their own documents

Inference The ICP appears to be mid-to-large finance departments or individuals within them who need scalable, accurate financial reporting.

Not evidenced No evidence of specific customer personas, use cases, or adoption patterns beyond the author’s self-description.

The description does not provide any information about:

• Revenue model
• Pricing structure
• Monetization strategy
• Subscription tiers or usage-based billing

It only mentions that:

• One analysis generates four deterministic renders (PDF, Word, Excel, CSV)
• The expensive model calls happen once
• Renders cost nothing extra

Not evidenced No evidence of a business model or pricing plan beyond the author’s own account.

The system is built using:

• Backend: FastAPI with SSE streaming
• Frontend: React with Tailwind CSS and Framer Motion
• AI models: GPT-5, Codex, Realtime API
• Execution environment: Sandboxed subprocess with CPU/memory limits and timeout
• Data handling: CSV, Excel, PDF via pandas, openpyxl, python-docx, reportlab

Key features:

• Live streaming of the agent loop (plan → code → sandbox → answer)
• Voice interaction via WebRTC with injected context
• Support for multiple output formats from one analysis
• Pre-computed digests to improve prompt accuracy

Inference The architecture shows a focus on robustness and auditability, especially around avoiding hallucination.

Not evidenced No evidence of scalability, performance metrics, or production deployment details.

The project is described as:

• A hackathon submission
• Built by one person (Samuel Fotso)
• Submitted to the OpenAI 2026 hackathon on Devpost

There is no evidence of:

• Revenue
• Customers
• Product usage
• Market traction
• Product iteration or feedback loops

Not evidenced No signs of product-market fit, user adoption, or commercial viability.

The author states that existing tools either:

1. Make you do the work yourself
2. Trust a chatbot that confidently makes up numbers

They position Finance Copilot as solving the second issue by using a trust loop involving Codex and sandboxed execution.

Inference It competes with generic AI financial assistants or data analysis platforms that lack accuracy guarantees.

Not evidenced No evidence of competitors, market size, or competitive positioning beyond the author’s own claims.

1. Unproven commercial viability: The project is a hackathon submission by one person with no revenue or customer data.
2. High technical complexity for a solo developer: The system involves AI orchestration, sandboxed execution, and real-time voice interaction — all of which are non-trivial to implement and maintain.
3. No evidence of product-market fit: No users, feedback, or traction beyond the author’s own account.
4. Dependency on proprietary APIs: Relies heavily on OpenAI models (GPT-5, Codex, Realtime), which may not be stable or available long-term.
5. Limited scalability assumptions: The system is described as working with large datasets but lacks evidence of performance testing or scaling.

Not evidenced No evidence of risk mitigation strategies, team expansion plans, or long-term roadmap.

1. What was the original problem you were trying to solve? How did you validate that it was a real pain point?
2. Have you tested this with any actual finance teams or users?
3. What is your plan for monetization and scaling beyond the hackathon?
4. How do you intend to handle edge cases in data formats (e.g., corrupted files, missing columns)?
5. Are there any legal or compliance concerns around executing arbitrary code in a sandboxed environment?
6. What are your plans for integrating with existing financial tools or platforms (e.g., Excel, ERP systems)?
7. How do you plan to ensure the accuracy of Codex-generated code over time?