Islam Assanov
Software engineer and AI researcher studying at Minerva University (B.S. in Artificial Intelligence and Computer Science, San Francisco, expected May 2027). Research interests: AI safety, interpretability, and alignment. Experience building GenAI-powered products and replicating AI safety research. Currently working on an open-source MCP server for AI coding agents and a jailbreaking replication study for a senior capstone.
Experience
- Founder, Cogni (Apr. 2026 – Present) — Open-core MCP server (Apache 2.0, Homebrew) giving AI coding agents structured recall over a Python repo via a tree-sitter–parsed SQLite code graph and 5 scoped MCP tools; benchmark harness showing 24.3% token reduction at parity pass-rate. Go, TypeScript, tree-sitter, SQLite, MCP. getcogni.dev, GitHub
- Research Assistant, Minerva University Psychology Lab (Sep. 2023 – Present) — Reviewed 20+ papers, designed 6 studies, peer-reviewed 10+ studies on Children Helping Science.
- Software Engineering Intern, Jubo Health Technologies (Sep. 2024 – Dec. 2024)
- Technology Intern, GlobalArt.AI (Feb. 2024 – Apr. 2024)
Research
Jailbreaking Modern LLMs (PAIR Replication), Jan. 2026 – Apr. 2026. Ran the PAIR jailbreaking algorithm (Chao et al., 2023) unmodified against Llama-3.1-8B, Llama-4-Scout, Qwen3-32B and Kimi-K2. Jailbreak rates ranged 86–98%; over 77% of successes came on the first attempt; roleplaying was the most effective strategy. Python, Groq API, Llama Guard 4. Code, write-up.
Projects
Achievements
- Calhacks 10.0 Winner (2023)
- UC Berkeley Hack For Impact Winner (2024)
- Seoul AI Hackathon Winner (2025)
Writing
- Blog — How I Jailbroke LLMs in 120 Hours and $10; How Qwen3's Thinking Mode Broke My Experiment; Jailbreak Is in the Process; Why Evals Matter in AI Safety; My Capstone: AI Safety Research; Story Behind LiveCaps; Building an Othello AI in Prolog.
Contact
San Francisco, California · islam@uni.minerva.edu · islam@getcogni.dev · GitHub · LinkedIn