Neuralchemy — AI Security

Sanskar Jajoo

AI Security Engineer · LLM Evaluation · Autonomous Red-Teaming

Self-taught, building since March 2025. I ship the tools that test whether LLM systems hold up under attack — prompt-injection firewalls, autonomous red-teaming loops, and CI/CD-style safety regression testing. Based in Raipur, India. Open to remote roles and freelance security work.

Get in touch See the work ↓

Projects

shipped · open source

detonategithub

Run untrusted AI tools in a sandbox and report what they actually do — not what their manifest claims.

9 starsgithub ↗

asrt-benchgithub

Fire a frozen attack pack at an AI agent, verify what lands, and diff safety across versions — a deterministic tool-call trace, not a judge's opinion.

clone & rungithub ↗

socraticgithub

A self-questioning skill for coding agents — Claude Code, Codex, and others — that interrogates a build before writing it.

106 starsgithub ↗

Datasets

Hugging Face

Prompt Injection Dataset

The original prompt-injection corpus — real and synthetic attack/benign pairs.

10,500+ downloadshf ↗

Videos

YouTube · teaching & journey

Learning AI, No Maths

Full roadmap and best resources for getting into AI without a math background.

roadmapwatch ↗

The College Project

A stock-market tool built with the Gemini API and Hugging Face — the project that started it all.

Gemini APIwatch ↗

Encoder from Scratch

5-part series: tokenization → positional encoding → attention → BERT-style encoder. Hindi + English.

5-part seriesplaylist ↗

Open Source

built from scratch

Autopilot-zero-to-herogithub

ML & deep-learning models from first principles — also the source code behind the AITL research paper.

from first principlesgithub ↗

TinyMobileLLMgithub

A small language model built and trained from scratch, sized to run on-device.

trained from scratchgithub ↗

Research

self-published · Zenodo

01zenodo

AI In The Loop (AITL) — a systems taxonomy for closed-loop autonomous evaluation. The research foundation behind ASRT.

Zenodo, 2026read ↗

02zenodo

The Autonomous Sunk-Cost Fallacy — stopping failures in agentic systems: why AI doesn't know when it's done.

Zenodo, 2026read ↗

03zenodo

The Modality Paradox in autonomous LLM engineering.

Zenodo, 2026read ↗

Writing

Medium & X

AI Can't Satisfy Itselfx

Six months of autonomous AI loops, and the hardest problem: teaching a model when it's done.

essayread ↗

Stuck in Its Own Loopmedium

Why agentic systems fail to stop themselves, and how to catch it — the Sunk-Cost Fallacy, explained.

Mediumread ↗

Stock-Analysis AI Toolmedium

A practical walkthrough with Python, the Gemini API, and FinBERT — the project that started it all.

Mediumread ↗

Skills

the stack

{

"security": ["adversarial attack generation", "direct + indirect prompt injection", "automated red-teaming", "OWASP LLM Top 10", "RAG poisoning", "LLM-as-judge", "threat taxonomy design"],

"ml_engineering": ["PyTorch", "Transformers", "PEFT", "scikit-learn", "AsyncIO", "FastAPI", "Docker"],

"llm_ecosystem": ["Anthropic API", "OpenAI API", "Hugging Face", "LiteLLM", "Ollama", "LangChain", "MCP"],

"practices": ["PyPI packaging", "CI/CD", "automated regression testing", "Bloom filters", "Aho-Corasick"]

}

contact.md

Let's talk about your model's blind spots.

Open to remote AI security roles and freelance red-teaming / evaluation work. Based in Raipur, India — happy to work across time zones.

sanskarmaheshwari062@gmail.com GitHub 🤗 Hugging Face X LinkedIn Medium