const frank = {
role: "AI/ML Engineer",
background: "Software Engineering & DevOps",
focus: ["LLMs from first principles", "Agentic AI", "LLM Eval & Safety", "Edge AI"],
building: "AI agents, LLM tooling, and eval/red-teaming systems",
learning: ["Advanced ML", "Distributed Systems", "Cloud Architecture"],
philosophy: "Build from first principles, then ship with tests.",
};- Building the LLM stack from scratch, no frameworks: a transformer with hand-derived gradients, LoRA/DoRA/QLoRA fine-tuning, a byte-level BPE tokenizer, an MCP server, and an agent red-teamer. The code is public, see below.
- Going deep on LLMs, agentic AI, LLM evaluation & safety, and edge AI
- Open source: second-highest contributor to llm-exe (my merged PRs)
- Ask me about LLMs, agentic AI, RAG, evaluation, MLOps, AWS, or DevOps
Languages
ML & AI
Frameworks & Libraries
Databases & Platforms
Cloud & Tools
|
A Transformer Built From Scratch in NumPy No PyTorch, no autograd. Forward and backward passes for attention, LayerNorm, RMSNorm, GELU, SwiGLU, RoPE, softmax, and cross-entropy are all hand-derived and gradient-checked to 1e-8. Ships both GPT-2 and Llama-style architectures in one codebase, trains on a CPU, generates with a KV-cache, and renders attention heatmaps so the internals are visible. 39 tests.
|
LoRA, DoRA & QLoRA From Scratch Efficient fine-tuning implemented by hand, forward and backward, gradient-checked to 1e-8. Includes DoRA (weight-decomposed LoRA, 2024) and QLoRA-style int8 quantization for a roughly 4x smaller frozen base, plus adapter save/load/hot-swap and honest held-out perplexity eval. The base model is provably untouched. 59 tests.
|
|
Adversarial Stress-Tester for LLM Agents Point it at any agent and it runs hostile and edge-case attacks aligned with the OWASP LLM Top 10 (prompt injection, jailbreaks, secret-leakage, hallucination, robustness, over-refusal), then reports per-category hold rates over N trials plus a severity-weighted risk score. Includes a CI gate that fails a build when an agent gets measurably less safe. Zero dependencies, 61 tests.
|
RL Trade Execution RL agent that minimises crypto slippage on large orders, using a Double DQN with a dueling architecture and prioritized experience replay alongside classical models. Trained on real Binance data across 3 assets and validated against 98M tick-level trades. Implements the Almgren-Chriss market-impact model, with 17 ablation studies and 176 unit tests.
|
|
Open-Source Contributor · llm-exe.com Second-highest contributor to this upstream TypeScript framework for building LLM applications, across workflow automation, CI/CD hardening, release pipeline improvements, and docs-sync infrastructure.
|
The Rest of the From-Scratch Stack Cleave, a byte-level BPE tokenizer with a provably lossless Unicode round-trip, exported to a dependency-free in-browser tokenizer. Winnow, an MCP server written without an SDK (JSON-RPC 2.0 over stdio) that compacts documents to Markdown and scans them for prompt injection. Device-Aware RAG Agent, routing between cloud, quantized on-device GGUF, and fully offline modes.
Write-ups at frankduah.me |
