Osprey Narrator

v0.1 Public

AI & Compliance Automation · Apache 2.0 Open Source

Fine-tuned domain model for SAR narrative generation & CEL rule synthesis.

Osprey Narrator is a specialized language model fine-tuned on AML typologies, Google CEL grammar, and FinCEN regulatory guidelines. Run locally on Ollama or in the cloud via HuggingFace.

View on GitHub Documentation & Specs Browse All Projects

Fine-Tuned LLMOllamaHuggingFaceGoogle CELFinCEN SARs

Model Architecture

8B Instruct

4-bit GGUF & FP16 weights

Context Window

8,192 Tokens

Supports extensive alert histories

Inference Latency

< 1.2s

Locally on Apple Silicon / CUDA

License

Apache 2.0

100% open weights & source

Overview & Engineering Philosophy

Compliance analysts spend over 60% of their investigative time writing narrative prose for Suspicious Activity Reports (SARs) and manually translating legal risk guidelines into engineering tickets.

Osprey Narrator automates these repetitive workflows. Given a sequence of flagged transactions and account histories, Narrator synthesizes formal, legally compliant SAR narratives citing relevant FATF typologies and 31 CFR regulations.

Furthermore, Narrator functions as a bi-directional compiler: compliance officers can describe a risk hypothesis in natural English, and Narrator synthesizes valid, compiled Google CEL expressions ready for Osprey.

Architecture Specifications

Specialized Fine-Tuning Corpus

Trained on hundreds of thousands of regulatory enforcement actions, FinCEN advisories, and FATF guidance papers.

Strict Hallucination Guardrails

Constrained decoding ensures synthesized CEL rules only access authorized transaction fields and standard CEL operators.

Zero Third-Party Egress

Deploy via Ollama directly inside your air-gapped banking VPC. No customer PII or transaction data ever leaves your perimeter.

Direct Osprey Engine Hook

Integrates with Osprey alert webhooks to automatically generate narrative drafts when transactions are flagged.

Ollama CLI Quickstart

# Pull and launch Osprey Narrator in Ollama
ollama pull josephgoksu/osprey-narrator
ollama run josephgoksu/osprey-narrator

# Example Python integration via HuggingFace
from transformers import pipeline

narrator = pipeline("text-generation", model="josephgoksu/osprey-narrator-v0.1")
result = narrator("Write a CEL rule to detect structuring transactions under $10,000")
print(result[0]['generated_text'])