What is AEGIS-FIN-1?

AEGIS-FIN-1 is a specialized financial AI assistant developed by aegisheimdall, built upon the mistralai/Mistral-7B-Instruct-v0.3 base model. It has been fine-tuned using QLoRA (4-bit NF4 quantization) on approximately 100,000 synthetic financial queries to provide safe and structured reasoning for various financial tasks.

Key Capabilities

• Financial Intent Classification: Accurately identifies user intents related to purchases, bill negotiation, credit optimization, and Buy Now Pay Later (BNPL).
• Structured Output: Generates structured JSON responses, making it suitable for integration into automated financial pipelines.
• Safety-Oriented: Designed to operate within a multi-stage safety pipeline that includes content classification, injection detection, PII detection, and compliance validation.
• Domain-Specific Knowledge: Covers topics such as credit cards, BNPL, budgeting, credit building, gig worker finance, and basic tax information.

What makes THIS different from other models?

Unlike general-purpose LLMs, AEGIS-FIN-1 is purpose-built for the financial domain with a strong emphasis on safety and structured output. Its fine-tuning on synthetic financial queries, calibrated against U.S. Federal Reserve data, allows it to understand and process complex financial scenarios with high accuracy (94.5% intent classification accuracy). The model's integration into a robust safety pipeline is a core differentiator, ensuring that financial advice is compliant and secure, actively blocking PII and potential fraud attempts.

Should I use this for my use case?

AEGIS-FIN-1 is ideal for applications requiring a specialized AI for financial transaction classification, negotiation, and optimization within a controlled, safety-first environment. It is particularly suited for:

• Financial AI assistants: Providing structured recommendations for personal finance management.
• Automated financial processing: Classifying and routing financial queries for backend systems.
• Compliance-focused applications: Where input validation, PII detection, and audit trails are critical.

However, it is not intended for:

• General-purpose conversations or chatbots.
• Medical, legal, or non-financial advice.
• Direct execution of financial transactions (it only recommends).
• Use outside of a safety pipeline, as it assumes upstream filtering and validation.

Limitations to consider:

• US-centric: Training data is calibrated for the U.S. market, limiting applicability to other regulatory jurisdictions.
• Synthetic data: May have limitations in covering all real-world edge cases.
• LoRA adapter only: Requires the base Mistral-7B-Instruct-v0.3 model for inference.