Most of what I build sits on one idea: the LLM is a constrained, contract-bound function and the engine is the runtime. Control flow, file access, retries and recovery belong to the engine — so a run can be reproduced, audited, and read back months later to find out why it did what it did.

• SkillFlow · pip install skillflow-py — a config-agnostic pipeline graph executor. Pipelines are YAML DAGs walked by the engine; loops, gates, retries and human approval checkpoints are the engine's job, not the model's. A step sees only the context it declares and writes only through tools the engine generates for its declared outputs — the tool to reach outside its contract does not exist in its schema.
• AItelier — the host application over that engine: a software-delivery pipeline, a State DAG for long-lived goals with revisioned acceptance contracts, and the whole surface exposed over MCP — so another agent can delegate bulk work to a deterministic pipeline and only decide at the checkpoints.

• An Index, Not a Store · preprint — when a frozen model writes facts into a LoRA adapter online, one per turn, day after day, the adapter becomes a recognition index, not a fact store. Recall fails within a handful of writes; recognition is the last readout to die. Five acts, each backed by the raw per-fact timelines in the repository.
• ActiveMemoryIndex — the system that follows from that result: memories stored twice, as verbatim timestamped turns and as atomic first-person facts, with retrieval that asks the log the question the user themselves would ask. Built for the Agent Memory Challenge 2026 (Academic Methods track).

• Robin — one OpenAI-compatible endpoint over every LLM plan you hold, so their windows add up instead of one burning out while the rest expire idle. Rotates per conversation, not per request, so provider prefix caches survive; a spent window is parked until the provider's own reset.
• 场记 / Continuity — local image, speech, music and SFX generation for an agent, plus a memory of what it made: the same character stays the same character across every call, and a degenerate generation is refused rather than returned. Models load per request and are released when idle — 0.21 GiB resident, measured.
• 随身翻译官 — a self-hosted real-time voice gateway (push-to-talk → ASR → LLM → streamed reply) on heterogeneous multi-GPU hardware, in three modes: translator, language tutor, meeting recorder.