Abu Bakar Siddik
AI engineer and researcher. I build agents that do real work, and study how they fail.

I lead the AI team at an early-stage legal-AI company, building agentic products for lawyers: autonomous workflows with a human in the loop where it counts.
Before that I led the Core RAG & AI team at AskTuring.ai, where I owned the platform's architecture while usage grew from 100 to 10,000+ users. I wrote up how we rebuilt the database layer to hold 1,500 concurrent AI requests. Earlier, at Sazim Tech, I shipped private LLM systems for enterprise clients, including ones that run with no internet connection.
I'm also an MSc researcher at Rajshahi University of Engineering & Technology (RUET), working on agent security and alignment. My review of cyber-capable AI agents is on arXiv, and my thesis looks at emergent misalignment: how narrow fine-tuning can make a model behave badly outside the task it was trained on.
I spend a lot of my time measuring whether agents work. At AskTuring I built the internal benchmark that cut evaluation time by 99%. I also made Jontro, a Bangla tool-use benchmark where every tool call can be replayed exactly, and Agent Bench, which runs the same tasks across agent frameworks.
In the open, I maintain Golem, a Go framework for AI agents, and claude-skills-collection, which has 1,134 stars on GitHub, and I publish fine-tuned models and datasets on Hugging Face. My agent Quicksilver won the Pokémon VGC Doubles track at UCLA's AltruAgent tournament in 2026.
Selected work
Essays
Golem: A Type-Safe AI Agent Framework for GoCan We Really Make AI Safe?Where Should the AI Actually Go?The Ship You Can't Dock: Architectural Debt in the AI EraAll essays · Guides · Newsletter · Guide: zero data retention across LLM providers
Agent Notes: a short recap every Saturday of what changed in AI that week, plus my thoughts on LLMs, agents and evaluation, and findings from my research now and then.
Research
Emergent misalignment after fine-tuning · MSc thesisCyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response · arXivJontro: a multi-turn Bangla and Banglish tool-use corpus and benchmark · Hugging Face datasetGet in touch
I read every email and reply within a working day. People usually write to me about:
- agents, RAG or LLM infrastructure you're working on
- research collaboration on agent security and alignment
- talks, podcasts, or something I wrote
My older address, abubakar1808031@gmail.com, still works.


