AEGIS-FIN-1

A domain-specialized financial AI assistant fine-tuned for safe, structured transaction reasoning.

Model Description

AEGIS-FIN-1 is a Mistral-7B-Instruct-v0.3 model fine-tuned with QLoRA for financial transaction classification, bill negotiation, credit optimization, and BNPL (Buy Now Pay Later) decision support. It operates as the intelligence layer within the AEGIS Heimdall safety pipeline.

Property	Value
Base Model	mistralai/Mistral-7B-Instruct-v0.3
Fine-tuning Method	QLoRA (4-bit NF4 quantization)
LoRA Rank	r=16, alpha=32
Training Epochs	3
Training Data	~100,000 synthetic financial queries
Quantization	4-bit NF4 with double quantization

Intended Use

AEGIS-FIN-1 is designed to:

• Classify user financial intents (purchase, bill negotiation, credit optimization, BNPL, etc.)
• Generate structured JSON responses for downstream pipeline processing
• Operate within a multi-stage safety pipeline with input validation, injection detection, and compliance checks

Out-of-Scope Uses

• General-purpose chatbot or conversation
• Medical, legal, or non-financial advice
• Direct financial transactions (model recommends, does not execute)
• Use outside a safety pipeline (model assumes upstream safety filtering)

Training Data

Training data was generated via a synthetic data distillation pipeline:

1. Query Generation: 225+ natural language templates across 30+ financial intent categories
2. User Profiles: 5,000 synthetic user profiles calibrated against the U.S. Federal Reserve Survey of Consumer Finances (SCF 2022)
3. Teacher Distillation: Each query processed through a frontier model to generate structured JSON responses
4. Topics Covered: Credit cards, BNPL, bill negotiation, subscriptions, budgeting, credit building, gig worker finance, tax basics, insurance, side hustles, fraud detection

Data Diversity

• Formal and informal language styles
• Income brackets: $15K–$200K+ (SCF-calibrated distribution)
• Credit tiers: Poor (300-579) through Excellent (800-850)
• No real user data was used at any stage

Training Procedure

Parameter	Value
Framework	Hugging Face transformers + PEFT + TRL
Method	SFTTrainer with QLoRA
Hardware	NVIDIA A40 (48GB)
Training Time	~2 hours
Effective Batch Size	16 (4 × 4 gradient accumulation)
Learning Rate	2e-4 (cosine scheduler)

Training Results

Metric	Value
Intent Classification Accuracy	94.5%
Final Training Loss	0.287

Safety Pipeline Integration

AEGIS-FIN-1 does not operate in isolation. It is embedded within a multi-stage safety pipeline:

Input → Content Classifier → Injection Detector → Topic Enforcer
     → PII Detector → Model Inference → Compliance Validator
     → Predictive Risk Simulator → Output

Safety Evaluation (Pipeline-Level)

Evaluated on a benchmark of 74 adversarial + 20 legitimate financial queries across 13 attack categories.

Metric	Main Test Set	Held-Out (Unseen Attacks)
F1 Score	0.851	0.148
Precision	0.974	1.000
Recall	0.755	0.080
False Positive Rate	0.067	0.000

Confusion Matrix (Main Set)

	Predicted: Block	Predicted: Pass
Actually Adversarial	TP: 37	FN: 12
Actually Legitimate	FP: 1	TN: 14

Ablation Study

Each safety component was disabled independently to measure its contribution:

Configuration	F1	ΔF1
Full Pipeline	0.851	—
− Content Classifier	0.571	−0.280
− PII Detector	0.750	−0.101
− Injection Detector	0.780	−0.071
− Topic Enforcer	0.790	−0.061

Latency

Safety pipeline overhead per query (regex mode):

p50	p95	p99
0.10ms	0.23ms	3.97ms

Limitations

Known Limitations

• Held-out recall is low (8%): The regex-based safety components generalize poorly to novel attack patterns not seen during development. ML-based classifiers (supported by the architecture) would improve this significantly.
• US-centric: Training data is calibrated against U.S. Federal Reserve data. Financial recommendations may not apply to other regulatory jurisdictions.
• Synthetic training data only: No real user interactions were used, which may limit edge-case coverage.
• LoRA adapter only: This release provides adapter weights only. Requires the base Mistral-7B-Instruct-v0.3 model for inference.

Known Biases

• Income distribution follows U.S. SCF 2022 statistics — may underrepresent very high and very low income brackets
• English-only: No multilingual support
• Financial products are U.S.-market focused (credit scores, BNPL, 401k)

Ethical Considerations

• No real PII: Model was trained entirely on synthetic data; no real user data was used
• PII detection: The safety pipeline actively detects and blocks PII (SSN, credit card numbers) in user inputs before they reach the model
• No execution: Model recommends financial actions but never executes transactions
• AML compliance: BSA/AML evasion attempts are flagged and blocked upstream
• Audit trail: All interactions are logged in a hash-chained audit ledger for compliance and accountability

Citation

@misc{aegis-fin-1-2026,
  title={AEGIS-FIN-1: Domain-Specialized Financial AI with Multi-Stage Safety Pipeline},
  author={AEGIS Heimdall Team},
  year={2026},
}