Neuralchemy

AI Security • Prompt Defense • LLM Safety

Building secure, reliable AI systems focused on prompt security, adversarial robustness, and practical safety tooling.

Featured Project — PromptShield & Threat Matrix

A comprehensive prompt injection and adversarial intent detection framework, classifying malicious jailbreak patterns across real-world and massive synthetic attack typologies.

Core Resources

• SOTA Datasets: neuralchemy/prompt-injection-Threat-Matrix A highly curated, leakage-free classification dataset mapping 32,000+ entries across a 5-dimensional security ontology (Intent, Technique, Severity).

  neuralchemy/prompt-injection-dataset 6000+ prompt injection and benign samples collected from realistic attack scenarios.

• DeBERTa Fine-Tuned Model: neuralchemy/prompt-injection-deberta Transformer-based prompt injection classifier.

• DistilBERT Base Model: neuralchemy/distilbert-base-threat-matrix A 99.4% F1-scoring Transformer defense gateway, optimized for high-speed, accurate prompt intent gating.

• Classical ML Models: neuralchemy/prompt-injection-detector Ultra-lightweight machine learning classifiers (RF, LR) for legacy/offline prompt risk detection.

• Live Demo Space: Prompt-injection-DeBERTa Interactive inference demo for prompt safety classification.

Research & Architecture

• AI In The Loop (AITL): Pioneering an inherently secure, multi-agent orchestration loop designed strictly to mitigate Prompt Injection (PI) bypass methodologies, enforce JSON-structured constraints, and evaluate autonomous systemic risks.

Mission

Advancing AI security through enterprise open-source datasets, robust model deployment, and adversarial safety research.

Connectivity

• Website: https://www.neuralchemy.in

Building safer AI systems through open security research. 🚀