Published Research

Paper 1 — AI In The Loop (AITL)

AI In The Loop: A Systems Taxonomy for Closed-Loop Autonomous Evaluation Sanskar Jajoo, Neuralchemy Labs, 2026

Establishes a formal taxonomy for autonomous AI evaluation systems, defining the layered architecture (Coder, Reviewer, Meta-Controller) that enables closed-loop ML engineering without human intervention.

Read on Zenodo | Code on GitHub

Paper 2 — The Autonomous Sunk-Cost Fallacy

The Autonomous Sunk-Cost Fallacy: Stopping Failures and Meta-Reasoning in LLMs Deployed within AEOS Sanskar Jajoo, Neuralchemy Labs, 2026

Discovers that LLM agents exhibit a computational analog of the human sunk-cost fallacy — continuing to invest compute into failing strategies rather than stopping. Introduces the AEOS (Autonomous Empirical Optimization System) framework and demonstrates that dual-agent architectures with asymmetric reviewer-coder roles eliminate this failure mode.

Read on Zenodo | Code on GitHub

Datasets

Prompt Injection Dataset

Curated samples for prompt injection detection with real-world attack scenarios.

neuralchemy/prompt-injection-dataset

Live Demo

Try our prompt injection classifiers directly in the browser:

Prompt Injection DeBERTa Space

Research Frameworks

AEOS — Autonomous Empirical Optimization System

A multi-agent framework where LLMs autonomously write, evaluate, and iterate on ML models. AEOS implements a Reviewer-Coder architecture where a critic agent with different weights oversees a coding agent, eliminating the computational sunk-cost fallacy.

github.com/m4vic/AEOS

Complete Model Inventory

Our HuggingFace Hub currently hosts the 5-Dimensional Threat Matrix Specialists, along with our legacy binary and DeBERTa baselines.