Neuralchemy Research

AI Security · Autonomous Systems · LLM Safety

Independent research lab building open datasets, models, and frameworks for LLM security and autonomous evaluation.

Research Papers

AI In The Loop (AITL): A Systems Taxonomy for Closed-Loop Autonomous Evaluation Sanskar Jajoo · Neuralchemy Labs · 2026 zenodo.org/records/19551173

The Autonomous Sunk-Cost Fallacy: Stopping Failures and Meta-Reasoning in LLMs Deployed within AEOS Sanskar Jajoo · Neuralchemy Labs · 2026 zenodo.org/records/19846960

Datasets

Prompt Injection Threat Matrix 32,320 samples · 7 intent classes · 10 severity levels · Full threat schema View Dataset

Prompt Injection Dataset 6,000+ samples · Benign vs malicious · Real-world attack scenarios View Dataset

Models

DistilBERT Threat Matrix Classifier 99.4% F1 · Prompt intent classification · High-speed inference View Model

DeBERTa Prompt Injection Classifier Transformer-based injection detection View Model

Classical ML Detector Lightweight RF/LR classifiers for offline/legacy deployment View Model

Live Demo

Try our prompt injection classifier: Prompt-injection-DeBERTa Space

About

NeuralAlchemy is an independent AI security research lab based in India. We build open datasets, train security models, and publish research on LLM behavioral failures and autonomous evaluation systems.

Website: neuralchemy.in GitHub: github.com/m4vic Contact: Via GitHub or neuralchemy.in