███╗ ███╗████████╗███████╗ ██████╗███╗ ██╗██╗ ██████╗
████╗ ████║╚══██╔══╝██╔════╝██╔════╝████╗ ██║██║██╔════╝
██╔████╔██║ ██║ █████╗ ██║ ██╔██╗ ██║██║██║
██║╚██╔╝██║ ██║ ██╔══╝ ██║ ██║╚██╗██║██║██║
██║ ╚═╝ ██║ ██║ ███████╗╚██████╗██║ ╚████║██║╚██████╗
╚═╝ ╚═╝ ╚═╝ ╚══════╝ ╚═════╝╚═╝ ╚═══╝╚═╝ ╚═════╝
Self-hosted AI out of Oshkosh, WI — wAIve.online · Fox Valley AI Foundation
Everything here runs on hardware I can physically kick. Two threads tie it together: agents that ship finished, validated software — not snippets — and making big models cheaper to run through tokenizer surgery, expert pruning, and token-efficient context formats. Results get published either way, including the negative ones.
Smaller models, fewer tokens, same behavior — with the measurements to prove it.
| Repo | What it does | ★ |
|---|---|---|
| tokopt | Post-hoc BPE tokenizer adaptation for deployed LLMs — script-aware pruning, continued BPE extension, embedding-only calibration. Validated end-to-end on Qwen3.5-4B; reproduces in ~3 hrs on a single RTX 3090. Ships a ~9,000-word PAPER.md and a step-by-step REPRODUCE.md. |  |
| research-test-Qwen3-Coder-Next-REAP-AWQ | REAP expert pruning + AWQ quantization of Qwen3-Coder-Next (149 GB BF16 MoE). 512 → 410 experts per layer via 4-dataset saliency calibration with super-expert preservation, then W4A16 @ group_size 32 for consumer GPUs. |  |
| ctx | CTX (Context Transfer Format) — an interchange format for LLM web consumption. Turns a 1.2 MB Wikipedia page into 150 KB of structure-preserving CTX: −87% bytes, ~90% fewer tokens. CLI, Python lib, and a FastAPI service with Redis caching + transparent proxy. |  |
Autonomy is easy. Finished is the hard part.
| Repo | What it does | ★ |
|---|---|---|
| cadillac | Autonomous coding agent: a sentence in, a validated app out. 12-stage pipeline (SPEC → … → CRITIC → RUNTIME → PACKAGE) with operational gates, a completeness CRITIC, runtime flow verification, and surgical-mode stuck-loop recovery. 616 tests passing · 146 apps built · 403 lessons learned. |  |
| cadillac-builds | The receipts: 26 runnable applications built unattended by Cadillac with zero human edits — published from a scan of ~99 build workspaces, failures acknowledged. Every project's README lists exactly which validation checks passed and which didn't. |  |
| workerAI | Agentic workflow system on vLLM + LangGraph — a ReAct-style autonomous agent that plans, reasons, and executes multi-step tasks with a library of specialized tools. |  |
Things the agents made (research artifacts, shipped as-is):
| Artifact | What it is | ★ |
|---|---|---|
| matrix-doom | Matrix-themed ASCII FPS raycaster in pygame — built autonomously by a modular LLM code generator. |  |
| silence-of-lolth | A 33,152-word, 14-chapter novel written end-to-end by an AI-authoring pipeline (AIRowling), voice-referenced on R.A. Salvatore's Dark Elf Trilogy. |  |
Tools for running a multi-node cluster from one chair.
| Repo | What it does | ★ |
|---|---|---|
| clusterspace | Tiled desktop workspace (Electron + React + TS) for terminals, embedded Chromium, and SSH auto-wrapped in tmux — sessions survive everything. Optional AI co-pilot gets first-class tools to read, type into, and drive any pane toward a goal. |  |
| model-chat-cli | Terminal command center for local AI servers. Auto-discovers Ollama, LM Studio, and vLLM on your network; chat with real TTFT/decode metrics, blind multi-model battles with auto-judged tournaments, load tests, and a 45-task agentic benchmark across 6 difficulty tiers. |  |
| cliide | The AI-native terminal IDE — file tree, editor, and an agent with tool execution in one TUI. Built AI-first rather than AI-bolted-on. On PyPI. |  |
| Repo | What it does | ★ |
|---|---|---|
| cardboard | Full-stack social platform for sports card collectors — share, buy, sell, trade. The day job and the code base, converging. |  |
| cblchat | Real-time enterprise chat with LDAP / Active Directory auth. |  |
| koding | CodeQuest — gamified, visual Python learning platform for kids 9+. |  |
| Cbl | Native iOS WebView wrapper for a WordPress storefront (Swift). |  |
Numbers pulled from the repos, on my own hardware:
| Result | Where |
|---|---|
| −4.9% bits/char and −4.2% tokens on held-out code, HumanEval within noise, greedy generation byte-identical | tokopt |
| −18% INT4 model size (4.4 GB → 3.6 GB) after tokenizer surgery + calibration | tokopt |
| 20% of MoE experts removed (512 → 410/layer) from a 149 GB coder model, then W4A16 | REAP-AWQ |
| −87% bytes / ~90% fewer tokens on real web pages | ctx |
| 146 applications built autonomously; 26 published unedited with per-check validation status | cadillac · builds |
- Reproduce or it didn't happen. Research repos ship a REPRODUCE.md you can actually follow, not a citation to a vibe.
- Negative results get published. tokopt's paper includes the method that didn't beat the baseline (hierarchical merge-tree init) right next to the ones that did.
- Validation status is public. cadillac-builds lists every failed check per project. No project is dressed up to look better than it is.
- Runs on hardware you can own. Everything above was built and validated on consumer GPUs — no cloud dependency, no API key required to reproduce.
wAIve.online — self-hosted AI platform · Fox Valley AI Foundation — open tooling for everyone else
Oshkosh, Wisconsin. Yes, really.