Character 01 — now playing
AI Engineer
AI Engineer building production multi-agent systems and AI-powered platforms. Experienced across Python, Rust and TypeScript — agent orchestration, RAG pipelines, LLM evaluation, fine-tuning of open models, and end-to-end automation. Active open-source contributor.
▸ objective Fine-tuned Gemma IBS health assistant with a disciplined, safety-gated eval.
QLoRA fine-tune of Gemma for the Kaggle Gemma hackathon. Data pipeline mines clinical guidelines, synthesizes Q&A under grounded prompts, and gates every sample through a medication whitelist + hallucination blocklist + LLM judge. 3-seed paired eval vs baseline.
▸ field metric +0.073 vs baseline · 4/4 red-flag handling
▸ objective Fine-tuned Gemma for modern Rust frameworks that mainstream code LLMs handle poorly.
Open pipeline building instruction datasets and fine-tuning Gemma for Leptos, Axum and Rig. ~4,200 train rows collected via tree-sitter + scraping; multi-judge eval (9 judges) tracking regressions per dataset version.
▸ field metric #1 overall on RustWebBench-15 (a trust-me-bro benchmark)
▸ objective AI calendar-intelligence platform — full-stack in a single Rust binary.
Leptos (SSR + WASM hydrate) + Axum + SQLx + PostgreSQL. AI assistant (Rig + Claude) with intent detection and a user-intelligence memory; Google Calendar, OAuth (PKCE) + JWT, Stripe, SSE streaming, per-IP/per-user rate limiting, circuit breaker, OpenTelemetry → Langfuse.
▸ objective Multi-agent orchestrator OpenCode plugin — delegate, verify, distill.
Orchestrator-only agent: the lead never writes code, it delegates each task to isolated workers, runs a mandatory verifier, and distills results to keep its context tiny across long multi-phase builds. Ships an npm package with runtime structured-output validation.
▸ objective Voice-to-text dictation for Linux — hotkey, speak, text at your cursor.
Multi-backend ASR (Groq, OpenAI Realtime, Deepgram, local whisper.cpp) with deep systems work: layout-aware key injection via uinput/evdev, cross-WM window tracking, and a Wayland OSD. Published to crates.io, docs.rs, AUR and Nix.
▸ field metric 50+ stars
AI Engineer · Remote, USA
AI Intern · Remote, USA
AI Software Engineering Intern · Remote
Trainable B-spline activations for chess NNUE evaluation, with custom CUDA kernels, int8 quantized LUT inference, and SPRT playing-strength validation.
Built and shipped Attio CRM and Linear integrations end to end — MCP tools, GraphQL/REST clients, OAuth 2.0 setup, tests and docs — plus a credential-decryption bug fix. 2 PRs merged.
RepoMerged a fix adding an async Tokio wakeup channel that resumes the parent agent loop the moment parallel sub-agents finish — a real concurrency fix in an agent runtime.
RepoCharacter 02 — now playing
Rust / Systems Engineer
Systems engineer who designs, builds and distributes production Rust software: a published Linux dictation tool, an end-to-end-encrypted remote-control daemon, and a full-stack platform in a single Rust binary. Comfortable from async runtimes and OS-level integration up through web services, packaging and multi-platform release engineering.
▸ objective Voice-to-text dictation for Linux — hotkey, speak, text at your cursor.
Multi-backend ASR (Groq, OpenAI Realtime, Deepgram, local whisper.cpp) with deep systems work: layout-aware key injection via uinput/evdev, cross-WM window tracking, and a Wayland OSD. Published to crates.io, docs.rs, AUR and Nix.
▸ field metric 50+ stars
▸ objective Remote control for Claude Code — drive sessions from your phone, end-to-end encrypted.
Rust workspace (daemon + relay + shared) on tokio/axum with X25519 + AES-256-GCM encryption; the relay only forwards opaque ciphertext. Flutter mobile app, QR pairing, self-hostable, distributed via curl, AUR and Homebrew.
▸ objective A browser engine built from scratch in Rust — HTML & CSS all the way to pixels.
Takes real HTML and CSS from tokenizer through parsing, style resolution, layout and painting to pixels in a native window. Built as a spec-driven engine with black-box conformance tests and a test runner.
▸ objective AI calendar-intelligence platform — full-stack in a single Rust binary.
Leptos (SSR + WASM hydrate) + Axum + SQLx + PostgreSQL. AI assistant (Rig + Claude) with intent detection and a user-intelligence memory; Google Calendar, OAuth (PKCE) + JWT, Stripe, SSE streaming, per-IP/per-user rate limiting, circuit breaker, OpenTelemetry → Langfuse.
▸ objective Adaptive TUI typing trainer — a faithful terminal port of keybr.com's algorithm.
Offline adaptive learning in the terminal; imports keybr.com history exports. Full release engineering: CI, Nix flake, justfile, multi-platform binaries, one-line install. Published to crates.io and docs.rs.
▸ objective KAN-based chess evaluation research — trainable activations inside a real engine.
Integrates Kolmogorov-Arnold Networks (trainable B-spline activations) into NNUE evaluation, validated with SPRT gauntlets. Custom B-spline IR op + int8 LUT quantization for inference; custom fused CUDA kernels for training. Documented in an IEEE-format write-up.
▸ field metric +169–226 Elo over matched MLP
AI / Platform Engineer · Remote, USA
AI Intern · Remote, USA
Cloud Engineering Intern · Cairo, Egypt (Hybrid)
KAN vs traditional architectures across a chess engine, a computer-vision pipeline, and behavioral anomaly detection — implemented as a Rust workspace on the Burn framework.
Merged a fix adding an async Tokio wakeup channel that resumes the parent agent loop the moment parallel sub-agents finish — a real concurrency fix in an agent runtime.
RepoCharacter 03 — now playing
Full-stack Engineer
Full-stack engineer who ships products end to end across Rust and TypeScript — from database and auth to UI and deploy. Experience with AI-powered web platforms, OAuth and billing integrations, queue-based backends and production observability. Comfortable owning a feature from schema to screen.
▸ objective Crowdsourced document-investigation platform — shipped and used in the wild.
Turns a 3.5M+ document release into a votable, verifiable feed with an interactive document viewer, layered annotations and threaded discussions. Heavy speed/UX engineering: lazy-loaded rendering, Supabase Realtime, large-corpus indexing under production load.
▸ field metric 500k+ requests · ~100k in one day
▸ objective AI calendar-intelligence platform — full-stack in a single Rust binary.
Leptos (SSR + WASM hydrate) + Axum + SQLx + PostgreSQL. AI assistant (Rig + Claude) with intent detection and a user-intelligence memory; Google Calendar, OAuth (PKCE) + JWT, Stripe, SSE streaming, per-IP/per-user rate limiting, circuit breaker, OpenTelemetry → Langfuse.
▸ objective Lecture-to-knowledge-graph converter via a multi-agent AI pipeline — live.
FastAPI + Next.js app that turns lectures, notes and papers into interactive knowledge graphs. Bring-your-own key or free models via OpenRouter, Graphviz rendering, credit-based usage, Google OAuth (PKCE), Celery async workers, PostgreSQL, Docker.
▸ objective Quran similarity visualizer — a prefix trie over all 6,236 ayahs, rendered with D3.
Builds a prefix trie over the entire Quran and visualizes it with D3 to reveal Mutashabihat (similar verses) for memorizers. Non-trivial algorithmic and Arabic/RTL text-normalization work (pause marks, Uthmani script).
▸ objective Remote control for Claude Code — drive sessions from your phone, end-to-end encrypted.
Rust workspace (daemon + relay + shared) on tokio/axum with X25519 + AES-256-GCM encryption; the relay only forwards opaque ciphertext. Flutter mobile app, QR pairing, self-hostable, distributed via curl, AUR and Homebrew.
AI Engineer (Full-stack) · Remote, USA
AI Intern · Remote, USA
AI Software Engineering Intern · Remote
Cloud Engineering Intern · Cairo, Egypt (Hybrid)
Full Stack Intern · Cairo, Egypt (On-site)
Built and shipped Attio CRM and Linear integrations end to end — MCP tools, GraphQL/REST clients, OAuth 2.0 setup, tests and docs — plus a credential-decryption bug fix. 2 PRs merged.
RepoCharacter 04 — now playing
Applied ML / Research Engineer
Applied ML engineer with published research on Kolmogorov-Arnold Networks (two IEEE-format papers) and hands-on experience fine-tuning open models with disciplined, reproducible evaluation. Comfortable owning the full research loop: hypothesis, data pipeline, custom kernels/quantization, benchmarking (SPRT, paired eval, ablations) and validation in real systems.
▸ objective KAN-based chess evaluation research — trainable activations inside a real engine.
Integrates Kolmogorov-Arnold Networks (trainable B-spline activations) into NNUE evaluation, validated with SPRT gauntlets. Custom B-spline IR op + int8 LUT quantization for inference; custom fused CUDA kernels for training. Documented in an IEEE-format write-up.
▸ field metric +169–226 Elo over matched MLP
▸ objective Fine-tuned Gemma IBS health assistant with a disciplined, safety-gated eval.
QLoRA fine-tune of Gemma for the Kaggle Gemma hackathon. Data pipeline mines clinical guidelines, synthesizes Q&A under grounded prompts, and gates every sample through a medication whitelist + hallucination blocklist + LLM judge. 3-seed paired eval vs baseline.
▸ field metric +0.073 vs baseline · 4/4 red-flag handling
▸ objective Fine-tuned Gemma for modern Rust frameworks that mainstream code LLMs handle poorly.
Open pipeline building instruction datasets and fine-tuning Gemma for Leptos, Axum and Rig. ~4,200 train rows collected via tree-sitter + scraping; multi-judge eval (9 judges) tracking regressions per dataset version.
▸ field metric #1 overall on RustWebBench-15 (a trust-me-bro benchmark)
▸ objective Multi-domain KAN research framework in Rust (graduation thesis).
Public Rust Cargo workspace where Burn is the main DL framework across a computer-vision pipeline (Hook Lens) and anomaly detection (Rough Guard). Generic-over-Backend code with switchable CUDA / wgpu / ndarray backends and custom #[derive(Module)] models.
▸ field metric KAN +1.81% over CNN+MLP baseline
AI / Evaluation Engineer · Remote, USA
AI Intern · Remote, USA
AI Software Engineering Intern · Remote
Trainable B-spline activations for chess NNUE evaluation, with custom CUDA kernels, int8 quantized LUT inference, and SPRT playing-strength validation.
KAN vs traditional architectures across a chess engine, a computer-vision pipeline, and behavioral anomaly detection — implemented as a Rust workspace on the Burn framework.
Switch character. Same engineer.
New Cairo, Egypt · open to remote · relocation possible for the right role
Arabic (native) · English (fluent)
Ain-Shams University, Cairo · Oct 2021 – Jul 2025
CGPA 3.48 / 4.0
Neural Networks & Deep Learning, Machine Learning, NLP, Image Processing, Operating Systems, Algorithms.