The laboratory
From the control plane down to specialized applications. Each page states what the project does, the architect's call that shaped it, and what went wrong along the way. The measured post-mortems are the useful part.
Control plane
RAM-aware MLX gateway (80 GB budget, LRU eviction, pinned model), shared memory bus, ReAct intent router, usage journal and a 10-section nightly eval. The "nervous system" that unifies the whole ecosystem.
Tabular foundation model
SOTA tabular ML, 100% local: calibrated classification + SHAP explainability. 694 books recategorized in the live database, offline medical demo.
Observe → propose loop
Five bricks shipped: usage journal, nightly habit mining, proposal cards, auto-distillation, feedback loop. The assistant notices, proposes, and learns from what you decline.
Library distillation
Distills the library into sourced digests, memories and an automatic export to KlodyAI.
Persistent lore engine
Entities, relations and narratives of a fictional universe, in a database: hybrid text + vector search fused by RRF, exposed to the agent as six tools. Standalone interactive HTML map, plus a lore → image bridge.
Text-to-SQL & semantic layer
Natural-language warehouse querying, 100% local: declarative semantic layer, constrained SQL generation, DuckDB execution. klodmetrics ask CLI.
Coding agent · ReAct
ReAct loop, 44 native tools + 18 MCP servers, write/read approval, Tauri desktop dashboard. 2,061 tests.
macOS memory guardrail
Swift daemon watching real memory pressure and suspending heavy consumers before suffocation. Two supervised agents: the guard and the menu bar.
100% local RAG
Answers sourced only from the books (zero hallucination). 24,799 books, 1.68 million chunks, a 21 GB database, cross-encoder reranking, cross-lingual.
Musical structure analysis
Scores a symbolic work across 29 weighted axes (form, harmony, melody, rhythm, texture, coherence) → one number. MIDI parser, drag-and-drop web UI, comparator, one-click push into Reaper.
Music generation
Lyrics → song → cloned voice → master: ACE-Step 1.5 (MLX), Demucs separation, RVC conversion, lookahead limiter. Asks the gateway to free RAM before running.
Incremental sample indexer
Semantic search across an audio sample bank: CLAP embeddings, LanceDB vector index, a watch agent reconciling the disk every six hours.
TTS & voice cloning
TTS orchestration, 4-D model router, semantic memory (recalls the right track, zero LB import).
Video world model
Action-conditioned 5B video world model, wired as a post-processing stage of the music pipeline: a clip is rendered from the final mix.
Generative groovebox
Embedded generative groovebox (Teensy + Pi + M5 Max). Hardware design in progress.
Fan control · anti-throttle
Swift daemon: proactive fan curve against thermal throttling under MLX load (M5 Max). IOHID + SMC IOKit sensors, watchdog and auto-restore of the system curve at idle.
macOS AI app · pure-Swift
Pure-Swift local AI app (MLX LoRA) — consumes the gateway optionally, never via a Python bridge.
Vocal channel strip
Audio plugin for zouk, afro-love and Caribbean R&B vocals: eight processing modules in a chain, calibrated by ear against a reference mic-preamp path. Built and validated as VST3, AU and standalone.
Most repositories are private: they index personal data. Whatever generalizes ships to GitHub under the MIT licence — auditable line by line, not merely described.
Two of them run as public demos, and not on my server: the computation executes in the visitor's browser through WebAssembly. Nothing is uploaded. The principle this site argues for, applied to its own shop window.
Which local model do you kill to make room for the next one? LRU eviction under a shared memory budget — the gateway's core, extracted. Zero dependencies.
PythonCalibration gate for tabular classifiers — model-agnostic, offline, CI-ready.
PythonMusical structure analysis across 29 axes. Zero dependencies, MIDI parser included.
PythonLocal coding agent: ReAct loop, native and MCP tools, write approval.
PythonTauri + React desktop dashboard driving the agent locally over WebSocket.
TypeScriptDrop a MIDI file, get it scored on 29 axes — the analysis runs in your browser, the file goes nowhere.
SpaceCan your classifier's probabilities be trusted? ECE, reliability curve and gate verdict, with scikit-learn running in WebAssembly.
SpaceThat is precisely the problem I solve. A 30-minute call is enough to scope an audit.