# Projects: Abu Bakar Siddik

## Golem

*Author & Maintainer · 2026 - Present · Open Source · Go*

[Docs](https://abubakarsiddik31.github.io/golem/) · [GitHub](https://github.com/abubakarsiddik31/golem) · [Details](https://abubakar-siddik.com/projects/golem)

A Go framework for building AI agents with typed dependencies and outputs, explicit tools, and inspectable runs, verified by a 19-check live harness.

**Stack:** Go, MCP, LLM Adapters, Streaming

---

## Portolan

*Author & Maintainer · 2026 - Present · Open Source · TypeScript*

[Docs](https://abubakarsiddik31.github.io/portolan/) · [GitHub](https://github.com/abubakarsiddik31/portolan) · [Details](https://abubakar-siddik.com/projects/portolan)

A site-map runtime for agents. Portolan learns a website's states and typed operations, such as search(query) or add_to_cart(product_id, size), from real runs, verifies them by replay, and serves them to any model over MCP, so an agent calls operations instead of clicking through screens.

**Stack:** TypeScript, MCP, Agent Skills, Browser Automation

---

## Herbie

*Author & Maintainer · 2026 - Present · Open Source · Go + React*

[GitHub](https://github.com/abubakarsiddik31/herbie) · [Details](https://abubakar-siddik.com/projects/herbie)

An open-source, self-hosted AI platform: multi-model chat, project workspaces, visual workflows, one-click MCP apps, custom API tools, and production RAG, on a Go backend built with Golem. The long-term goal is a fully open-source desktop super app.

**Stack:** Go, React, MCP, RAG

---

## Axiom Wiki

*Author & Maintainer · 2026 - Present · Open Source · TypeScript*

[Docs](https://abubakarsiddik31.github.io/axiom-wiki) · [GitHub](https://github.com/abubakarsiddik31/axiom-wiki) · [Details](https://abubakar-siddik.com/projects/axiom-wiki)

Compiles documents into a linked markdown knowledge base that an LLM keeps up to date and agents can query over MCP.

**Stack:** TypeScript, MCP, Hybrid Search, CLI

---

## Skill Manager

*Author & Maintainer · 2026 - Present · Open Source · Rust + Tauri*

[Website](https://abubakarsiddik31.github.io/skill-manager/) · [GitHub](https://github.com/abubakarsiddik31/skill-manager) · [Details](https://abubakar-siddik.com/projects/skill-manager)

One desktop dashboard to discover, install, enable, disable, edit, and delete AI agent skills across Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and more.

**Stack:** Rust, Tauri, React, TypeScript

---

## LLM From Scratch

*Author · Open Book · 2026 - Present · Open Source · PyTorch*

[Read the book](https://abubakarsiddik31.github.io/llm-from-scratch/) · [GitHub](https://github.com/abubakarsiddik31/llm-from-scratch) · [Details](https://abubakar-siddik.com/projects/llm-from-scratch)

A free, implementation-first book on training LLMs end to end. You build the tokenizer, the model, the pre-training loop, and the fine-tuning yourself, in PyTorch, and read the code behind each step.

**Stack:** PyTorch, Python, Transformers, LoRA

---

## Agent Bench

*Author & Maintainer · 2026 - Present · Open Source · Python*

[GitHub](https://github.com/abubakarsiddik31/agent-bench) · [Details](https://abubakar-siddik.com/projects/agent-bench)

A benchmark that runs the same tasks across agent frameworks and compares tool-calling reliability, orchestration control, and observability, with every run published.

**Stack:** Python, Evaluation, Benchmarks

---

## Quicksilver (Won Pokémon VGC Doubles · Werewolf finalist · Honor of Kings semi-finalist, UCLA AltruAgent Tournament 2026)

*Tournament Agent · 2026 · UCLA AltruAgent Tournament*

[Details](https://abubakar-siddik.com/projects/quicksilver)

A fully autonomous agent that played three games at the UCLA AltruAgent AI Agent Gaming Tournament: it won Pokémon VGC Doubles, reached the Werewolf final, and reached the Honor of Kings 5v5 semi-final. No human in the loop during matches.

**Stack:** Python, MCP, Game AI, Damage Calculation, Belief Tracking

---

## ShowOnce

*Hackathon Build · 2026 · Meta Global AI Developer Hackathon · Python*

[GitHub](https://github.com/abubakarsiddik31/showonce) · [Details](https://abubakar-siddik.com/projects/showonce)

Film a tabletop task once. ShowOnce turns the clip into a layout-independent program, asks you when it is unsure, then performs the task on a shuffled table in a MuJoCo robot simulation and checks its own work.

**Stack:** Python, MuJoCo, SAM 3.1, Muse Spark 1.3, FastAPI, React

---

## CareerKor

*Founder & Product Lead · 2024 - Present · Remote*

[Open careerkor.com](https://careerkor.com) · [Details](https://abubakar-siddik.com/projects/careerkor)

A career platform where one profile generates role-specific resumes, cover letters, and SOPs, with a voice-based Interview Coach and a hiring workspace for organizations.

**Stack:** Next.js, NestJS, Voice AI, PostgreSQL

---

## MagicSpin 360° (Winner, Google Nano Banana Hackathon 2025)

*Hackathon Build · 2025 · Google Nano Banana Hackathon*

[Read the write-up](https://www.kaggle.com/competitions/banana/writeups/magicspin-360) · [GitHub](https://github.com/abubakarsiddik31/MagicSpin360) · [Details](https://abubakar-siddik.com/projects/magicspin)

Winner, Google Nano Banana Hackathon 2025. Turns a single 2D image into an interactive 360° rotation through a multi-model generation pipeline.

**Stack:** Gemini, Stability AI, Segment Anything, Next.js

---

## AI Virtual Try-On

*Computer Vision R&D · 2024 · Private engagement*

[Details](https://abubakar-siddik.com/projects/tryon)

A diffusion-based virtual try-on pipeline that transfers a garment onto a person photo while preserving pose and fabric detail, built for e-commerce.

**Stack:** Diffusion, Human Parsing, Pose Estimation, PyTorch


---

## Models and datasets (Hugging Face)

- [Qwen3.5-4B-fable5-sft](https://huggingface.co/abubakar-siddik/Qwen3.5-4B-fable5-sft) (Model, QLoRA SFT on Qwen/Qwen3.5-4B-Base): QLoRA fine-tune of Qwen3.5-4B-Base on real Claude Code agent sessions: reasons before it answers, and 97.5% of held-out agent turns produce a valid tool call. 80.5% on GSM8K; fits on one 16 GB GPU.
- [jontro-qwen3.5-2b-lora](https://huggingface.co/abubakar-siddik/jontro-qwen3.5-2b-lora) (Model, LoRA on Qwen/Qwen3.5-2B): LoRA adapter that teaches Qwen3.5-2B to call tools in Bangla and Banglish conversations: right tools at 84.6% of decision points against 41.5% for the base model.
- [functiongemma-270m-it-mobile-actions](https://huggingface.co/abubakar-siddik/functiongemma-270m-it-mobile-actions) (Model, SFT with TRL on google/functiongemma-270m-it): A 270M-parameter FunctionGemma fine-tuned for on-device function calling across seven phone actions.
- [Jontro](https://huggingface.co/datasets/abubakar-siddik/Jontro) (Dataset): 2,000 multi-turn Bangla and Banglish tool-use conversations with replayable, simulator-backed tool results.
- [bangla-alpaca](https://huggingface.co/datasets/abubakar-siddik/bangla-alpaca) (Dataset): About 52K Alpaca instructions adapted into natural Bangla rather than translated word for word.
