HOSEIN ABDOLLAHI
Computer Science undergrad working where machine learning meets real systems.
I am an undergraduate researcher in Computer Science at Shahid Chamran University of Ahvaz, ranked first in my Computer Science program, working at the intersection of machine learning and large language models. My research investigates the reasoning and reliability of LLMs and agentic AI systems — how these models fail, why they fail, and how to design methods that remain sound where conventional approaches do not.
I am currently developing two papers in this area, including one on applying machine learning to industrial control systems (ICS). A single thread runs through this work: the edge cases and failure modes that standard algorithms quietly ignore. I am less interested in models that perform well on benchmarks than in understanding the conditions under which they break — and in building systems that hold up under exactly those conditions.
This research is grounded in hands-on engineering. I work across the ML stack (PyTorch, TensorFlow, scikit-learn, XGBoost) and the modern LLM toolkit (RAG, agentic frameworks, Ollama), and I build the end-to-end systems that carry models from experiment to deployment (Django, FastAPI, SQL/NoSQL). I have also served as a teaching assistant in my faculty.
⌨ LANGUAGES
✦ MACHINE LEARNING / AI
▦ DATA & BACKEND
⚙ TOOLS
B.Sc. Computer Science
Teaching Assistant
Top of the Class
Research YOU ARE HERE
Machine learning & large language models — including one paper applying ML to industrial control systems (ICS).
What's Next
Publishing the current research and pushing further into ML and network intelligence.
An emotion engine for literary text, plus the critique tool built on top of it. The core library maps a text's emotional content into color space using a 12-emotion taxonomy, rendering narrative arcs and keyword-level sentiment. The companion Claude Skill turns that engine into a developmental critic: bundled Python measures told-emotion density, pacing variance and prose metrics, while the model judges what those numbers mean — so critique is evidence-backed rather than generic.
A provenance-based defense against indirect prompt injection. Instead of pattern-matching malicious text — which an attacker can always rephrase — it tracks where each piece of context came from and constrains what untrusted data is permitted to do. Shipped in two forms: a standalone gateway, and an MCP server that drops the same protection into any Model Context Protocol tool chain.
Does telling a code model to think harder make it worse? A replication porting inverse scaling in test-time compute (Gema et al., TMLR 2025) from natural-language reasoning to code reasoning. The result was negative — the effect did not transfer — and the repository reports it as such, harness and findings intact.
VIEW REPO →A new quest is being prepared. Check back soon.
UNLOCK PENDING