Work

Migrated from an earlier TypeScript/Express + Agent-SDK design to a self-hosted Python stack — a parity-verified, independently peer-reviewed cutover (8.1/10). An in-house stateless step loop on the OpenAI Python SDK talks to vLLM over its OpenAI-compatible API, with chain-of-thought reasoning on by default; tools run in-process (MCP dropped). A durable, crash-safe job spine — a bespoke scheduler, worker, and idempotent delivery ledger with leases and retries that survive crash and redeploy — runs the proactive/scheduled substrate, with a three-tier egress safety ladder (auto / gated+logged / human-approved) bounded by a fail-closed permission ceiling. Retrieval is single-vector semantic search over a local bge-small embeddings sidecar. A full self-hosted observability stack (Prometheus, Grafana-as-code, alerting, uptime checks, GPU exporter) ships alongside.

A continuous Python transcriber polls stations every 10 seconds, feeding audio chunks to faster-whisper with VAD filtering and hallucination detection. MeCab post-processes segments to add kanji→furigana readings. An SSE streaming API delivers captions in real-time to the frontend, which highlights active words every 250ms with click-to-seek navigation. A caption gate pauses audio at the transcription horizon so playback never outruns captions. HLS.js handles stream playback with a three-tier recovery strategy, and a Zustand store persists player state across page navigations.

Runs a configurable Monte Carlo of log-normal returns per scenario — 500 paths by default, up to 2,000 — with percentile tracking (10th/25th/50th/75th/90th). The phased engine carries portfolio state across phase boundaries so a 'work 5yr → part-time 3yr → quit 12yr' plan produces a single continuous simulation. Break-even uses binary search to find the minimum work years that keep probability of ruin below 5%.

Country data lives in structured YAML research files and gets generated as TypeScript via a code-gen script — no manual edits to the runtime module. State is managed with useReducer, and a shared FX rate service with live currency conversion connects the budget output directly to the planner's scenario inputs.

Each prefecture and ward is an FSRS card (ts-fsrs) scheduled by spaced repetition, with all progress persisted in localStorage. Correct = Good, incorrect = Again; response time is tracked for analytics but kept out of the recall signal. The interactive SVG map drives per-region click targets and a mastery heatmap derived from FSRS stability plus recent accuracy.

Built on Vite + React 19 with PixiJS 8 for the space-themed game, VexFlow 5 for real-time music notation rendering, and smplr for piano sample playback. Performance data is collected per-session with a versioned localStorage schema.

ESP-IDF firmware on an ESP32-S3. Two INMP441 MEMS I2S microphones in stereo configuration. The firmware reads the I2S samples and streams raw PCM over WiFi as UDP to a Python asyncio UDP server (Opus encoding is the planned next step), which ingests the audio and hands it off to the MrHuang backend. Work-in-progress: the audio path is still stubbed.

Seeed Studio XIAO RP2040 microcontroller soldered to mouse switch contacts, powered by an internal LiPo. Firmware randomises intervals using a PRNG seeded from analog floor noise. A small project for learning USB HID descriptors and embedded power management.