# About — Abu Bakar Siddik

> AI engineer and researcher working on production agent systems, open-source agent infrastructure, and agent security and alignment

Abu Bakar Siddik is an AI engineer and researcher based in Rajshahi, Bangladesh. He is Lead AI Engineer at a stealth legal-AI company, building agentic software for lawyers: agent orchestration, real-world tool integration, and low-level LLM control in production. He is the author of the open-source agent framework Golem and the self-hosted AI platform Herbie. As an MSc researcher at RUET he studies agent security and alignment: how agents fail at their boundaries, and how fine-tuning can shift model behaviour.

Previously he led the Core RAG & AI Team at AskTuring.ai, building a production RAG platform without vendor lock-in that scaled from 100 to 10,000+ concurrent users with ChatGPT-level latency. He cut hallucinations by 98% with hybrid search, reranking, and strict source-grounding; designed multi-agent workflows with explicit state management; and built a time-aware RAG system with short-term, long-term, and semantic memory layers.

Before AskTuring he shipped private on-premise air-gapped AI for enterprise clients at Sazim Tech, designed hexagonal multi-provider LLM architectures, and built a 3M+ sample RAG application end-to-end. Earlier work included Bengali ASR research, Rasa-based conversational AI, and Bengali NLP datasets at Intelsense AI.

In October 2026 his autonomous agent Quicksilver competed in three games at the UCLA AltruAgent AI Agent Gaming Tournament: it won Pokémon VGC Doubles, reached the Werewolf finals, and reached the Honor of Kings 5v5 semi-finals. In 2025 he won the Google Nano Banana Hackathon with MagicSpin 360° (single-image-to-3D rotation pipeline using Gemini Pro Vision + Stability AI + Segment Anything).

## Current role

**Lead AI Engineer** at [Stealth](https://abubakar-siddik.com) — since Apr 2026, Rajshahi, Bangladesh.

## Research

- **Alignment drift after fine-tuning** (Ongoing, MSc thesis, RUET) — How narrow fine-tuning can shift a language model's behaviour well outside the task it was trained on (emergent misalignment), and which mitigations work. The work starts by reproducing a published result on a controlled pipeline before testing anything new.
- **Cyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response** (arXiv preprint, 2026) — A structured review of what goes wrong when cyber-capable models are given autonomy: multi-step attack chains, privilege separation, provenance, and who can reach a responder when an evaluation escapes its sandbox. https://arxiv.org/abs/2607.25379
- **Jontro: agentic data for Bangla** (Ongoing, Dataset) — A tool-use dataset for Bangla-language agents. Tool results come from deterministic simulators, so every trajectory can be replayed and checked, and a first-call diagnostic tests whether a model picks the right tool in Bangla at all.

## Open source contributions

- **open-code-review** (Alibaba): Refactored the telemetry summary path into a typed struct. https://github.com/alibaba/open-code-review/pull/909
- **Atlas** (pacifio): Fixed agents not inheriting the configured permission mode, and polished git-status indicators in the file explorer. https://github.com/pacifio/atlas/pulls?q=is%3Apr+author%3Aabubakarsiddik31+is%3Amerged

Also: claude-skills-collection (1.1k stars); zdr, a guide to zero-data-retention LLM providers; whisperx-transcribe, a Claude Code skill; the-truth-serum, an AI fact-checker; releasebot-ai.

## Education

- **MSc in Computer Science & Engineering** — Rajshahi University of Engineering & Technology (RUET) (Nov 2025 – Present) — Ongoing
  - Thesis: Alignment drift in language models after domain-specific fine-tuning (in progress)
  - Research: Agent security and LLM alignment
  - Supervisor: Dr. Md. Rabiul Islam
  - Thesis studies how narrow fine-tuning can shift a model's behaviour outside its target domain, and which mitigations hold
  - Published a review of cyber-capable AI agents on arXiv (2026): vulnerabilities, evaluation containment, and defensive response
  - Building Jontro, a replayable tool-use dataset for Bangla-language agents
- **BSc in Mechatronics Engineering** — Rajshahi University of Engineering & Technology (RUET) (Jan 2019 – Jan 2024)
  - Thesis: Deep learning for medical applications
  - Three applied machine learning papers (2023–2024), on genetic-disorder classification, detecting AI-generated images, and patient monitoring
  - Competed in ML challenges, including Mayo Clinic STRIP AI on Kaggle and a top 0.5% finish in Analytics Vidhya's Dataverse Hack
  - Relevant coursework: Artificial Intelligence, Machine Learning Algorithms, Software Engineering, Robotics, Numerical Analysis & Statistics, Digital Signal Processing & Machine Vision, Automation

## Location

Rajshahi, Bangladesh

## Availability

Selective about new work: strategic AI builds with a small number of companies at a time.

## Contact

- Email: [abubakar1808031@gmail.com](mailto:abubakar1808031@gmail.com)
- GitHub: https://github.com/abubakarsiddik31
- LinkedIn: https://linkedin.com/in/abu-bakar-siddik31
- X (Twitter): https://x.com/abubakar_AIE
- ORCID: https://orcid.org/0009-0003-7834-7469
- Website: https://abubakar-siddik.com
