Adarsh Arun

AI Safeguards Engineer  ·  Bengaluru, India

I work where ML systems meet adversaries.

$ whoami

operator     Adarsh Arun
role         AI Safeguards Engineering Intern @ XelerAIT
location     Bengaluru, IN
education    BE CSE, BNM Institute of Technology
focus        AI guardrails / LLM security / detection engineering
shipping     Wardline — Rust guardrail library (pre-alpha)
open source  roostorg/osprey — PR #451 merged
uptime       since Aug 2023

Most of what I build answers one question: what happens when this model is given something it should refuse? At XelerAIT I design guardrails for production agentic AI in healthcare — safety, reliability, and compliance enforced at the framework layer rather than bolted on after an incident.

---

▸ Currently building

   PROMPT   ──▶   BROKER    ──▶   SCHEMA    ──▶   MODEL     ──▶   AUDIT
  untrusted      regex +        structured      policy-         verdict
    input       metaprompt         json        bound call         log

Wardline  

An embeddable guardrail evaluation library for Rust. Guards run inline in the request path — before an action is taken or an LLM response is released — and return a blocking verdict: allow, block, or modify. A library you call in-process, not a proxy you deploy: no separate service, no network hop, no async runtime. Guards are ordinary Rust implementing one trait, so a guard is testable like any other code.

---

▸ Selected work

Project	What it is
Wardline Rust tower/axum	Embeddable, synchronous, trait-based guardrail library. Four-crate workspace with panic isolation and a bounded audit trace.
ISTVON Prompt Engine Python Streamlit Gemini	Security broker for LLM prompts — dual-layer regex + metaprompt screening, JSON structuring that isolates malicious payloads, audited verdicts.
Vendor Risk Assessor Google ADK MCP FastAPI Docker	Hierarchical agent pipeline over an MCP tool server. Risk score computed deterministically in Python — the LLM writes prose, never the number. Kaggle Agentic AI Competition
Telecom SOC Simulation Python RAG AutoAI	DFIR automation with RAG-backed MITRE ATT&CK mapping over MDE investigation packages.
Adaptive Triage Engine Contextual Bandits scikit-learn	RL policy that scores and triages SOC alerts from forensic feature data, evaluated on reward distributions and precision-recall trade-offs.
Satellite Land-Use TensorFlow ResNet50 VGG16	Custom CNN vs. ResNet50 vs. VGG16, benchmarked head-to-head on EuroSAT imagery.

More projects

• AI-Powered Educational Feedback System — NLP analysis of student writing and code. Sentiment via VADER and RoBERTa, readability metrics, style checks, and a Python path that flags syntax, logic, efficiency, and PEP 8 issues with fixes.
• Retail Analytics Dashboard — Apriori market-basket rules and RFM customer segmentation over a year of UCI Online Retail transactions.
• Multi-Client Group Chat — Low-latency messaging on raw Java TCP/IP sockets; multi-threaded server with synchronized broadcasting.

---

▸ Open source

roostorg/osprey #451 — merged upstream into Roost's trust & safety rules engine.

docker compose up failed on Windows because Git's default core.autocrlf=true checks out *.sh with CRLF, turning the shebang into #!/bin/bash\r — surfacing as a misleading no such file or directory. The repo's pre-commit hook fixed line endings at commit time; nothing enforced LF at checkout time. Two lines of .gitattributes, zero content churn.

---

▸ Stack

_{Security — LLM guardrails · prompt-injection defense · MITRE ATT&CK mapping · DFIR pipelines · alert triage · compliance architecture
AI — RAG · contextual bandits · Model Context Protocol · multi-agent orchestration}

---

▸ How I think about it

Threat-model before you optimize.

A guardrail you cannot audit is a guess.

Latency and cost are safety properties too.

_{adarsh04arun.in  ·  the site has a terminal — type help}