Lemonade Interviewer
TypeScriptA local-first mock interview agent: speech in, a local LLM in the middle, speech out — no cloud required.
Projects
Work I can stand behind — local AI, on-device inference, agents, and research — grouped so you can scan it. Each card opens a project page. Public repos stay one click away. More source lives on GitHub.
A local-first mock interview agent: speech in, a local LLM in the middle, speech out — no cloud required.
The local AI server that serves chat, coding, speech, and image models from your own GPU or NPU.
A specification approach to prompt optimization — and the measurements that showed a 29.8% token cut.
AMD's open-source framework for building AI agents that run entirely locally on Ryzen AI hardware.
Maxwell's 1873 Treatise as executable Python — all 866 articles, in the units he actually used.
The local AI server that serves chat, coding, speech, and image models from your own GPU or NPU.
Close-to-metal Python for AMD Ryzen AI NPUs — the layer under friendlier local-inference stacks.
A WIP toolkit sketch for XDNA NPUs: ONNX Runtime, IRON, Quark quantization, and hardware monitoring.
A WIP Swift and Python toolkit for converting and running LLMs on the Apple Neural Engine.
A C++ inference engine on MLX — local LLMs on Apple silicon, AMD GPUs, and CPU, without a Python runtime.
Graph-level bisection so large ONNX models can run on machines that cannot hold the whole graph.
A local-first mock interview agent: speech in, a local LLM in the middle, speech out — no cloud required.
AMD's open-source framework for building AI agents that run entirely locally on Ryzen AI hardware.
A long-form note on semantic markup and template logic for prompting — the document I used when I contributed to BMAD.
An open-source method for turning an idea into working software with structured AI agents — I contributed expansion-pack templates.
A public teaser of agents and tools generated by an AI agent-ecosystem creator.
A specification approach to prompt optimization — and the measurements that showed a 29.8% token cut.
Maxwell's 1873 Treatise as executable Python — all 866 articles, in the units he actually used.
The unglamorous middle of the Maxwell project: scans in, structured articles out.
The edition layer: 1892 indexes, OCR extracts, and architecture maps — not a typeset LaTeX book.
A swappable attention interface, and PhaseAttention: alternating access and suppressed streams.
See the full list at github.com/antmikinka.