Shahir M A2026 / portfolio

Shahir
M A.

I build production apps and experiment with local LLMs and inference on Apple Silicon.

Founder, The Smart Language
20K+ commits in 3 years
Currently building Haku, Haku Terminal & SuperTutora

01 — Apps in production

Apps in production.

02 — On-device experiments

A “why not” experiment.

Rung-models · BitNet b1.58 · ternary from scratch

Rung 3.

Curiosity-driven: can a 1.58-bit base learn real grammar, and can a small LoRA on top turn that into something useful?

335 M ternary base trained from random init on 7 B tokens, then a 6.2 M-param LoRA on top that turns natural language into AppIntent JSON. The base never learned JSON or app catalogues — the adapter taught it those.

on M3 Air ANE
243 ms
on disk (4-bit)
168 MB
intent acc.
~98 %
1.
50 M · 0.9 B tok · 40 MB · PPL 133, ~77 % grammar — recipe trains.
2.
99 M · 2.8 B tok · 38 MB · PPL 44, 94 % base grammar → 96.8 % with a 2.6 M LoRA.
3.
335 M · 7 B tok · 168 MB · NL → AppIntent JSON in 243 ms on M3 Air ANE, 100 % schema-valid.

~$50 of A100/A10G time through Rung 2, plus one $26 retrain for Rung 3.

03 — Other work

Building blocks.

The substrate underneath the apps and experiments — runtimes, inference engines, small models, voice infrastructure. Mostly private on GitHub.

Inference runtime · Rust + Metal

silicon-llm

Custom hybrid inference runtime for LLMs on Apple Silicon. Qwen3.5 prefill on the Neural Engine, decode on the Metal GPU, KV cache shared via IOSurface zero-copy.

1.7B @ LUT-8 on 8 GB M1
RustMetalCoreML / ANEIOSurface
LLM runtime · Swift + Metal 4

mlforge. in progress

A full Apple-Silicon-native LLM runtime written from scratch in Swift against Metal 4 and the private ANE path. Custom quantisation including TurboQuant-3.

byte-identical to ref
Swift 6Metal 4CoreMLGGUF
Speech-to-text · Rust + Metal

silicon-stt

Qwen3-ASR speech recognition running natively on Apple Silicon via Metal GPU. Quantised GGUF, encoder-decoder, INT4 encoder variant.

1.88 % WER (1.7B) · 2.90 % WER · 823 MB (0.6B)
RustMetalGGUF
On-device LM · CoreML

spectra.

A 1.5B on-device language model with per-app LoRA packs — Swift library plus a haku-run CLI. Each app gets its own adapter rather than its own model.

1.5B base · per-app LoRAs · ANE
SwiftCoreMLANELoRA
WhatsApp call bridge · Go + Rust

WhatsApp call bridge

A bridge that connects the WhatsApp Calls API to real-time voice AI over WebRTC — pluggable STT, TTS, and LLM. Built on Pion (Go) and a Rust voice engine.

Pion · WebRTC · OpenAI Realtime
GoRustWebRTCWhatsApp Calls API
From-scratch model · MLX

rune-lm

A 20.5M-parameter decoder trained from scratch in MLX. Natural-language commands → AppleScript. Routes out-of-scope intents upward via PASS_TO_CLOUD.

20.5 M · 78 MB · sub-second on M1
MLXPythonBPE 8kRoPE

Also worth a look: turn-taking (5.26 M EoT detector, < 1 ms CPU), silicon-tts (dual-engine Qwen3-TTS + Kokoro), silicon-embed-v2 (Metal 4 ML tensor passes), haku-voice-tiny (~10 MB ternary CoreML on ANE), teachezee-voice-rs (production Rust WebRTC voice service).

04 — Teaching & talks

Programs & talks.

05 — Hackathon wins

Hackathon wins.

06 — Open source

Open source.

qwen3-asr-llamacpp — Qwen3-ASR for llama.cpp; patch, GGUF models, benchmarks.

claude-accountability-partner — exercise & hydration reminder where Claude Code watches the screen.

07 — Elsewhere

Where to find me.