Skip to content
View muhib-karim's full-sized avatar

Highlights

  • Pro

Block or report muhib-karim

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
muhib-karim/README.md

muhib-karim: AI evaluation, verification and validation, developer tools. Test engineer turned AI builder. Software you can inspect.

muhib-karim

I work on AI engineering and software verification and validation (V&V). I came to software through automotive test engineering, so I build tools the way I was trained to test: written requirements, repeatable checks and evidence someone else can inspect.

Open to AI engineering and V&V roles in Australia.

Projects

  • agent-eval-harness: compares coding models and agent CLIs on the same tasks, with hidden pytest grading, unbiased pass@k and JSONL evidence with SHA-256 provenance. Site CI
  • api-vv-toolkit: generates API tests from an OpenAPI 3.x spec and written requirements, runs them and reports a requirements traceability matrix with a CI exit-code gate. Site CI
  • codegraph: indexes Python and Java into an embedded code graph and answers callers, inheritance, dead-code and change-impact queries with file:line evidence. Site CI
  • bangla-llm-eval: a small, reproducible Bangla (Bengali) evaluation of LLMs: translation, NLI and Bengali numerals, with raw outputs and a live leaderboard. Site CI
  • humming-keyboard: private keyboard with offline voice typing: a FlorisBoard-based Android keyboard plus iPhone, Windows, macOS and Linux apps; Bangla phonetic typing. Site CI
  • code-video-kit: 34 typed Remotion components for code and data videos, with a rendered gallery and measured render times. Site CI
  • Rambler-Bangla: Bengali-first voice typing for Gboard 18.3.1: a fail-closed, fingerprint-guarded smali patcher plus a Java policy runtime, with a device-free CI suite that builds, patches and verifies a synthetic APK. Source only, no APKs. Site CI

Extended open source

  • sonu: private, offline voice typing for Windows, macOS and Linux, transcribing into any app with local Parakeet/Whisper/Moonshine models; an independent fork of Handy. Site CI
  • Antigravity-Claude-Code-Proxy: local Anthropic API gateway so Claude Code can use Gemini, Claude (via Google Cloud Code) and Perplexity; extends antigravity-claude-proxy. Site CI

Developer tools and AI agents

  • OpenSH: natural-language shell: plain English in, one PowerShell/Bash/Zsh command out, with a safety guard that holds destructive commands for confirmation. Single-file, stdlib-only Python CLI. Site CI
  • NewTerminal-AI: PowerShell module that turns plain-English requests into commands with a local Ollama model, behind an AST-based destructive-command gate; Pester-tested on Ubuntu and Windows. CI
  • opencode-superapp: schema-validated registry of 50 MCP servers and 7 one-command stacks for the OpenCode agent, with a zero-dependency installer that merges configs safely and never writes secrets. CI
  • antigravity-workflow-kit: VS Code extension for the Antigravity editor: auto-accepts AI edits over the Chrome DevTools Protocol, with a banned-command deny-list and a loop guarded by a circuit breaker and rate limiter. Open VSX CI
  • Claude-Proxy-Extension: VS Code extension for a local Claude/Gemini proxy: per-window model switching, live per-account quota bars and a status-bar proxy toggle. Open VSX CI
  • bounty-radar: daily sweep of live, payable open-source bounties with repo-liveness checks and spam filtering; publishes HTML/RSS/JSON, an MCP server and a Worker API. Site CI
  • google-takeout-metadata-fixer: Python CLI that restores EXIF dates, GPS, descriptions and people tags to Google Photos Takeout exports by matching JSON sidecars and writing them with ExifTool. CI

Apps

  • openclaw-desktop: Tauri 2 + Rust + React desktop app for the OpenClaw assistant: installs the CLI, runs and monitors the local gateway, embeds its dashboard, streams logs and runs diagnostics. CI
  • codex-manager: prototype Tauri v2 desktop app for managing several OpenAI API accounts: encrypted local storage, usage tracking and a local OpenAI-compatible proxy with 4 routing strategies. CI
  • VoiceDicatationCodex: offline voice dictation for Windows: a .NET 8 WPF app that downloads SHA-256-verified whisper.cpp models and transcribes locally, no cloud API. CI
  • vehicle-selector-pro: multi-tenant Shopify app (Rails 7.2) for automotive stores: Year/Make/Model/Trim/Engine fitment, a cascading storefront selector, fit badges and OE-number search. Report CI
  • voip-saas-platform: full-stack VoIP SaaS reference: Express API with JWT auth, signed call-control webhooks, an atomic prepaid wallet and call records; React softphone and Expo app, 200 backend tests on Postgres. CI
  • HoliBooks: browser reader for the Quran, Bible, Bhagavad Gita, Torah, Guru Granth Sahib and Dhammapada: a vanilla JS PWA on public scripture APIs, with bookmarks, search and audio. Site CI
  • markdown-resume-micro: client-side Markdown resume builder: live preview, 3 print-ready themes, PDF via the browser and an offline CLI. Zero dependencies, no tracking. Site CI

How I work

  • Written requirements: tests start from them, and results are reported in a traceability matrix.
  • Independent checks: hidden pytest grading, a labelled retrieval evaluation, a spec validation gate.
  • Raw evidence kept: JSONL with SHA-256 provenance, raw model outputs, file:line references.
  • CI: runs on every repo that shows a badge here; each badge links to its workflow runs.

Built with: Python (pytest, pydantic, httpx), TypeScript (React, Remotion, zod), JavaScript (Node.js, Express), Rust (Tauri), Kotlin (Jetpack Compose), SwiftUI, C# (.NET, WPF), Ruby on Rails, PowerShell, tree-sitter, sherpa-onnx, MCP, GitHub Actions.

Pinned Loading

  1. agent-eval-harness agent-eval-harness Public

    Python CLI and library for comparing coding models and agent CLIs on the same tasks: hidden pytest grading in isolated subprocesses, unbiased pass@k, JSONL evidence with SHA-256 provenance, and HTM…

    Python