devinblaze/laya-multidomain-calibrated

Laya Multi-Domain Calibrated Decision Engine

Pre-trained and Platt-calibrated decision model weights for Laya (System 1 non-autoregressive decision engine).

Operational Performance

• Inference Latency: 0.07ms – 0.25ms on CPU (Pure NumPy/SciPy engine)
• Validation Accuracy: 100.00%
• Expected Calibration Error (ECE): 0.2263
• Scoring Rule: Strictly Proper Scoring (proper_reward: log + spherical - RPS)

Supported Domains

1. Technoverse B2B ICP Rubric: 1-5 ordinal score (Dubai/UK/China stocking distributors vs excluded OEMs & India leads)
2. B2B Buying Intent: rfq (Immediate RFQ/PO), project (Planned), info (Catalog), none (Spam)
3. Hardware Catalog Match: exact, high, partial, none for Cisco, Fortinet, Siemens, ABB, Schneider
4. Binance USD-M Futures Signals: Quantitative 33ms decisions (fade on overextended pump, bounce, breakout, hold)
5. Agentic Browser DOM Actions: Navigation primitives (click, type, select, scroll, wait)

Usage

import json
from laya import LayaTrainer

trainer = LayaTrainer(dim=256)
# Load calibrated checkpoint
with open("laya_multidomain_calibrated.json", "r") as f:
    ckpt = json.load(f)
trainer.weights = {k: np.array(v) for k, v in ckpt["weights"].items()}
trainer.biases = {k: np.array(v) for k, v in ckpt["biases"].items()}
trainer.temperatures = ckpt["temperatures"]

res = trainer.predict("Company: Gulf Tech Stockist, Dubai UAE. Urgently procuring Cisco Catalyst switches.", question)
print(res["choice"], res["confidence"])