Skip to contentAbu Bakar Siddik

Projects

Things I've built in the open and for competitions. Work for employers and clients is on the About page. Everything else is on GitHub.

Quicksilver screenshot

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.

2026 · UCLA AltruAgent tournament · Python, MCP · Details

Axiom Wiki screenshot

Axiom Wiki

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

2026 · TypeScript · 112 stars · Details · Source

ShowOnce screenshot

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.

2026 · Meta Global AI Developer Hackathon · MuJoCo, SAM 3.1 · Details · Source

Open source

claude-skills-collection · a curated collection of Claude skills1,134 ★Golem · Go framework for typed, inspectable AI agentsGoPortolan · typed operations for websites, served to agents over MCPTypeScriptHerbie · self-hosted AI platform built on GolemGo, ReactLLM From Scratch · a free book on training LLMs, seven chapters inPyTorchSkill Manager · one desktop app for agent skills across tools31 ★zdr · zero-data-retention guide for LLM providers30 ★whisperx-transcribe · Claude Code skill for speaker-labelled meeting transcripts11 ★Agent Bench · the same tasks run across agent frameworksPythonPAGANS · adapts a prompt to each model family; on PyPIPythonworktrunk-vscode · VS Code sidebar for git worktrees and their pull requestsTypeScript

Contributions: merged pull requests to alibaba/open-code-review and pacifio/atlas.

Models and datasets

Fine-tuned models and datasets on Hugging Face.

Qwen3.5-4B-fable5-sftQLoRA 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.QLoRA SFTjontro-qwen3.5-2b-loraLoRA 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.LoRAfunctiongemma-270m-it-mobile-actionsA 270M-parameter FunctionGemma fine-tuned for on-device function calling across seven phone actions.SFT with TRLJontro2,000 multi-turn Bangla and Banglish tool-use conversations with replayable, simulator-backed tool results.Datasetbangla-alpacaAbout 52K Alpaca instructions adapted into natural Bangla rather than translated word for word.Dataset

Earlier

MagicSpin 360° · Winner, Google Nano Banana Hackathon 20252025CareerKor · AI career platform for candidates and hiring teams2024AI Virtual Try-On · Diffusion-based virtual try-on2024