I'm learning modern AI systems by rebuilding Transformer inference from first principles.
AI learning · Roadmap · Selected projects · Toolbox
My current focus is not just using models, but understanding what happens between tokens and generated text. I turn each topic into small implementations, tests, notebooks, concept notes, and benchmarks—and keep the whole learning trail public.
|
From Tokenizer to KV Cache, I'm building a Llama-style decoder-only inference engine step by step—starting from the math, making every tensor shape explicit, and verifying each layer before moving on. This is a 40-unit, test-driven learning path toward Hugging Face numerical parity and reproducible Prefill/Decode performance analysis. Start learning → · 40-unit roadmap · Source · Tests
|
Built: Next gate: The live checkpoint is maintained in the learning repository. |
- Match final logits with the Hugging Face reference implementation
- Produce identical generation results with and without KV Cache
- Publish reproducible Prefill/Decode benchmarks and performance analysis
| Stage | Focus | Evidence I want to produce |
|---|---|---|
| 01 · Foundations — built | Tokenizer, embeddings, LM Head, autoregressive generation | Shape contracts, notebooks, unit tests |
| 02 · Model internals — now | Attention, RMSNorm, SwiGLU, RoPE, decoder blocks | Layer-by-layer numerical alignment |
| 03 · Efficient inference — next | Sampling, padding, Prefill, Decode, KV Cache | Cache/no-cache correctness parity |
| 04 · ML systems — next | GQA, benchmarks, profiler, memory analysis | A reproducible performance report |
My learning loop is deliberately engineering-heavy: derive the operation, inspect the shapes, implement the smallest correct version, lock it down with tests, compare against a trusted reference, and only then optimize it.
Alongside the inference engine, these projects show how I turn learning into reusable material and ship systems-focused products:
|
A self-contained AI learning path with 40 units, 28 day guides, concept notes, executable notebooks, test-driven implementations, and explicit graduation gates.
|
A native macOS WidgetKit dashboard for system health, weather, and privacy-conscious Codex, Cursor, Claude Code, and Grok usage snapshots.
|
|
A profile-guided Rust splitter and TypeScript runtime for loading large WebAssembly modules in hot and cold stages.
|
Export Godot 4 projects to WeChat, Douyin, and TikTok Mini Games through an editor-native, CI-validated WebAssembly toolchain.
|
Transformer inference · ML systems · runtime engineering · developer tooling
Learning in public, one verified tensor shape at a time.