No PhD. No papers. Just building and thinking. Tech Systems, AI, real-world engineering. Results first. No hype required.

— "Don't do anything a computer can do better."

⚡ Now: Shipping markdown renderer and export pipeline for ocf-py. Wired up Bluesky posting via Sonnet editorial layer. Cursor adapter for SQLite exporters live.

Updated: 2026-04-29

A spec format for LLM tool use - point your model at a static driver spec instead of running an MCP server. No transport layer, no build step, the driver embeds the tool descriptions in the system prompt. Not even the client needs to know how to handle a tool call.

One file format for AI conversations - portable across Claude Code, Cursor, and Codex, attachable to commits, resumable across tools. Wire-compatible with OpenAI Chat Completions for the message core.

A credential broker for AI agents - Device-Flow UX with end-to-end encryption, so agents can request user secrets without custom web portals, OAuth callback plumbing, or plaintext storage. Internet-Draft in preparation for IETF 126.

A Python framework for building digital robots - autonomous units that perceive, plan, act, learn, and stay accountable. Deterministic state, governed tool execution, self-improving memory.

A small, task-tuned character model (~0.5-20B parameters) with a clean bio, structured knowledge tiers, a social-rule engine, and deterministic refusal paths - produces more believable personality than a frontier model with an "act as Jarl of Winterhold" prompt.

Multi-paper research path in progress: methodology validation on a single character > open-source trust-state dialog engine > cross-character reproducibility > domain transfer beyond NPCs.

LLM as worldbuilder. Question: can an LLM build a fictional world with enough substance to carry believable characters (NPC Forge), RPG systems, novels, gamebooks, or asset pipelines? Reference points: Witcher, Game of Thrones, Marvel Universe, Cyberpunk 2077 - closed worlds with narrative depth.

An OS-style dashboard that exposes AI services and your own systems through one central surface. Goal: snap together small micro-apps (transcription, custom agents, mini-workflows like "draft the product description" or "pull a customer record from the CRM") without rebuilding infrastructure and integrations for every idea.

Drafts - no public repo yet.

A bilingual six-piece series on LLM epistemics and curation - what models actually do when they "generate", why alignment-as-control mistakes the substrate, and what curation looks like as the alternative.

bizrockman/Unpublished-Papers · EN/DE

• AIInferenceBenchmark - Inference performance benchmark across providers
• EbookSummarizerToPodcast - Turning books into listenable summaries
• OpenXE - Continued development, mainly on the integrations side
• AutoTrader - Draft, occasional tinkering
• DSPyTutorial - Companion code for the DSPy article series

_{📍 d.gerst@bizrock.de · Status badge and Now block are refreshed daily from recent commits. Source: scripts/update_status.py.}