🧠 Why a Fraud Defense Decision Engine?
Standard conversational LLMs are prone to hallucinations, verbosity, and latency—rendering them unsuitable for mission-critical financial safety audits.
connectailab-rl-CHOCO-BRAIN_GUARD-v1 is an External Prefrontal Cortex Decision Engine for Solopreneurs & E-Commerce — powered by the JEV (Judgment-Evaluation-Verification) protocol and calibrated via Deep Reinforcement Learning (PPO), this on-device executive cortex neutralizes malicious fraud attacks, high-risk refund exploitation, and compliance crises in 0.05 seconds with zero latency.
🎯 Model Persona: "The FinTech Risk Guardian"
While standard generic models maintain balanced weights (1.0 : 1.0 : 1.0), CHOCO-BRAIN_GUARD-v1 is mathematically tuned with an aggressive risk penalty:
Reward=0.8⋅10Profit+1.0⋅(ΔSatisfaction×100)−3.0⋅(Incidents×5)
🛡️ Core Guard Protocols:
- Zero-Tolerance Security Filtering: Hard blocks triggered on transaction fraud score ≥0.70 (Industry standard is 0.95).
- Suspicious Refund Isolation: Automatic human review routing for high-risk refund claims (score ≥0.60).
- Legal & Compliance Sentinel: Immediate escalation on critical CS signals (litigation, compliance, defamation threats).
📊 Benchmark & Real-World Validation
Tested against 713 unseen real-world credit card transactions (177 confirmed fraud attempts) and 40 months of stochastic enterprise operations:
Table with columns: Evaluation Metric, Rule-Based Baseline, CHOCO-BRAIN_GUARD-v1, Net Advantage| Evaluation Metric | Rule-Based Baseline | CHOCO-BRAIN_GUARD-v1 | Net Advantage |
|---|
| Fraud Attacks Detected | 138 / 177 cases | 158 / 177 cases | +20 Additional Frauds Blocked |
| Financial Loss Prevented | 68% | 88% Loss Reduction | +20%p Net Capital Saved |
| Monthly Operating Profit | 3.05M KRW / mo | 10.62M KRW / mo | +248% Net Profitability |
| Operational Governance | Baseline Static Rules | Prefrontal Risk Governance | Resilient Capital Protection |
🔄 Dual-Cortex Decision Architecture
This repository forms the natural language reasoning cortex of Hwihwa Lab's two-tier autonomous decision pipeline:
┌────────────────────────────────────────────────────────┐│ CONNECT TYCOON Enterprise Simulation Environment │└───────────────────────────┬────────────────────────────┘ │ Multi-objective PPO ▼┌────────────────────────────────────────────────────────┐│ [Policy Brain] CHOCO-BRAIN_GUARD-v1-teacher ││ - File: brain.json (142 KB Policy Matrix) ││ - 11.11M KRW/mo | 96% Satisfaction | 2.0 Incidents │└───────────────────────────┬────────────────────────────┘ │ JEV Trajectory Ground-Truth ▼┌────────────────────────────────────────────────────────┐│ [Reasoning LLM] CHOCO-BRAIN_GUARD-v1 (Current Model) ││ - Qwen3.5-0.8B LoRA Adapter (Natural Language JEV) ││ - 0.05s Real-time Fraud Adjudication & Action Routing │└───────────────────────────┬────────────────────────────┘ │ Quantization (Q8_0) ▼┌────────────────────────────────────────────────────────┐│ [Edge Package] CHOCO-BRAIN_GUARD-v1-GGUF ││ - 811 MB Standalone GGUF Binary + Modelfile Bundle ││ - Zero-latency local execution in Ollama & LM Studio │└────────────────────────────────────────────────────────┘
🚀 Quick Start & JEV Protocol
1. High-Level Decision Engine (connect-tycoon)
import connect_tycoon as ct # Load directly from Hugging Face Hubbrain = ct.CompanyBrain.from_hub("hwihwalab/connectailab-rl-CHOCO-BRAIN_GUARD-v1") # Make real-time decisionresponse = brain({ "model": "typesafe/jev-1.13", "state": { "department": "refund", "reason": "defect", "amount": 45.0, "fraud_score": 0.82 }, "questions": {"decision": {"type": "choice"}}})print("Guardian Decision:", response["answers"]["decision"]["choice"])
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_model_id = "Qwen/Qwen3.5-0.8B"adapter_id = "hwihwalab/connectailab-rl-CHOCO-BRAIN_GUARD-v1" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base_model, adapter_id)
# Run simulator with CHOCO-BRAIN_GUARD as CEOconnect-tycoon play --brain hf:hwihwalab/connectailab-rl-CHOCO-BRAIN_GUARD-v1
⚠️ Honest Limitations & Operational Scope
- Not a general chatbot. CHOCO-BRAIN_GUARD is an ultra-specialized 0.8B reasoning engine that outputs structured JSON fraud verdicts in 0.05s. It will not write marketing copy, handle open-ended customer support chats, or answer general knowledge queries.
- 0.8B parameter boundaries. While extraordinarily fast and zero-cost on edge hardware, complex multi-hop forensic reasoning is bounded by its 0.8B scale. Highly convoluted criminal rings should be analyzed alongside upstream telemetry rules.
- Korean e-commerce & solopreneur primary. Fine-tuned on Korean-primary transaction disputes, refund abuse vectors, and chargeback scenarios. International scenarios are supported but Korean business contexts exhibit the highest fidelity.
- Human-in-the-loop for high exposure. Designed to filter 80-90% of routine fraud attempts autonomously. Decisions outputting
ESCALATE, AUDIT, or borderline confidence (< 0.85) should always route to human operators.
- v1.0 research release. Thresholds are calibrated for solopreneur risk tolerance. Enterprise deployments should benchmark false positive vs false negative rates against historical transaction logs before live gating.
👤 About Hwihwa Lab
“AI × Automation × Aesthetics — Crafting tools you'll love ✨”
Hwihwa Lab is an independent AI research & engineering studio dedicated to blending practical on-device intelligence, autonomous operations, and refined aesthetics. We engineer resilient systems that free solopreneurs and indie creators from operational overhead—giving people their time and creative joy back.
🔬 Core Research & Engineering Pillars
- Zero-Cost Edge Operations: Private, on-device AI pipelines for autonomous business workflows without ongoing cloud API overhead.
- Autonomous Decision OS & Micro-ERP: Deterministic, prefrontal-cortex-style reasoning architectures designed to orchestrate daily business decisions and operations.
- Embodied AI & Domain-Specific Robotics: Simulation-driven reinforcement learning, physical control brains, and specialized domain data engineering for real-world environments.
- Interactive Creative Aesthetics: Fusing functional automation with refined visual craft and modern design philosophy to build delightful tools.
🌐 Connect & Research Ecosystem
Part of the CONNECT AI LAB Collaborative Research Network · CONNECT TYCOON Decision Foundation
📄 License
This project and fine-tuned weights are distributed under the Apache 2.0 License — see the repository LICENSE file for full details.
Base model weights adhere to the original Qwen License.
📖 모델 개요
connectailab-rl-CHOCO-BRAIN_GUARD-v1은 1인 기업 및 이커머스 운영자를 위한 초경량 온디바이스 사기 방어 언어 모델입니다. Qwen3.5-0.8B에 강화학습 정책을 전수하여, 결제 사기·환불 악용·법적 리스크 상황을 환각 없이 0.05초 만에 판단하고 정형화된 행동 판결을 내립니다.
🚀 핵심 활용
- Python (JEV 프로토콜):
connect-tycoon 라이브러리로 사전 학습된 모델을 즉시 로드하여 자동 의사결정 수행
- Hugging Face PEFT:
AutoModelForCausalLM 및 PeftModel로 어댑터 직접 로드 및 파이프라인 통합
- 시뮬레이터 연동:
connect-tycoon play --brain hf:hwihwalab/connectailab-rl-CHOCO-BRAIN_GUARD-v1 명령어로 가상 기업 CEO 자리에 앉혀 실시간 경영
📘 한국어 공식 가이드
전체 아키텍처 및 상세 사용법은 README_KR.md 파일에서 확인하실 수 있습니다.
Built and maintained by HWIHWA LAB in collaboration with CONNECT AI LAB.