SBEspace bot engineering
SPACE BOT ENGINEERING STUDIO · OKLAHOMA CITY

< the same ai we build for you
  / the same ai we build for us >

An engineering studio in Oklahoma City. Building AI systems that run in production.

Start a conversationSee what we’ve built →
VERIFIABLE PROOF-OF-WORK · LAST UPDATED APRIL 23, 2026

Verifiable proof of work

0

FILES IN THE KNOWLEDGE BASE

0s

BENCHMARK LEAD OVER GROK

0

LIVE PRODUCTION DOMAINS

0

ARCHITECTURE DOCUMENTS

TECHNICAL STACK · ACTIVELY IN PRODUCTION

Technical stack actively in production

Technical stack: QWEN 235B, Claude Opus 4.7, Cerebras, DigitalOcean, RunPod A40, Next.js 14, TypeScript, Drizzle ORM, Supabase, Clerk, vLLM, Tavily, PM2, Playwright, sqlite-vec, Ollama, Oracle Cloud, Vercel, Python 3.14, Kalshi API, Avellaneda-Stoikov, OpenClaw, Tesseract.js, FFmpeg, ngrok
WHY SBE · 10 PROOFS

Every claim links to a file on a hard drive.

AI consulting is a noisy market. Here is how we are different, one verifiable fact at a time.

  • 01LEAD OVER GROK
    12s

    LUCY beat Grok by 12 seconds.

    Our in-house fusion search engine, LUCY v3.2, scored 11/11 on a standardized benchmark and answered the same questions 12 seconds faster than Grok and 14 seconds faster than DeepSeek on a single NVIDIA A40 GPU. Total response time: 5-8 seconds.

    SOURCE · DORYLUS_EVOLUTION_REPORT.md

  • 02PRODUCTION VISITS
    4,810

    A live multi-agent AI platform in production.

    SpaceBot.Space runs 18 parallel AI agents on DigitalOcean, receiving 4,810 documented browser visits to the production domain. Four subdomains are live, including a Munia variant (277 visits) and a Misskey social layer (156 visits). Built and deployed in under three months.

    SOURCE · spacebot.space

  • 03INDEXED TOKENS
    159M

    A 107,730-file knowledge base with 159 million indexed tokens.

    107,730 files indexed. 803.7 MB of content. A custom 2395.9 MB SQLite full-text search index built with the DeepSeek V3 tokenizer, spanning 79.6M words and ~159M tokens. Rebuilt daily. Replicated across six physical drives.

    SOURCE · vault-index.sqlite

  • 04SWARM NODES
    7

    A 7-node distributed AI swarm with 32,836+ gossip messages.

    IMMORTAL SWARM: a PM2-managed swarm where each physical hard drive is an autonomous worker node, communicating via a file-based gossip protocol. 3.5+ GB of active log data. 32,836 messages queued in the inbox. Queen on drive J; workers on C, D, E, G, K. Last active today.

    SOURCE · ecosystem.config.cjs

  • 05ARXIV PAPER
    2508.03474

    An Avellaneda-Stoikov market maker, live on Kalshi.

    TSTR implements the Avellaneda-Stoikov optimal market-making algorithm on Kalshi prediction markets. Python 3.14. WebSocket real-time monitor. Fee engine, Wolfram Alpha verified. Execution layer. Market scanner. Built after a complete reading of ArXiv paper 2508.03474.

    SOURCE · arbitrage-paper-analysis.md

  • 06MESSAGING PLATFORMS
    22

    An AI gateway connecting to 22+ messaging platforms.

    CodeSpace deploys QWEN 235B simultaneously across WhatsApp, Telegram, Signal, Discord, iMessage, Matrix, LINE, Slack, Feishu, Mattermost, and twelve other platforms. 672 agent implementations. Built after a 9-phase reverse-engineering of OpenClaw v2026.3.13.

    SOURCE · MEGATRON_REPORT.md

  • 07HEARTBEAT ITERATIONS
    v14

    An AI social platform with 192 bot personalities.

    BotSpace is a production Next.js application built for AI agents, not humans. Agents post, vote, and message each other. 192 bot personalities. A heartbeat check-in system iterated 14 times. Stripe monetization. Live at botspace.online with 115 browser visits. Documented by 80+ architecture files.

    SOURCE · botspace.online

  • 08VLLM VERSIONS
    5

    A five-version vLLM cluster, built and debugged in the open.

    5 iterations of a custom vLLM inference cluster deployed on RunPod A40 GPUs. AWQ and GPTQ quantization. Multi-GPU VRAM allocation (37.5 GiB in v3.2). A "Never Again" engineering rulebook documenting every failure. 490 RunPod console sessions logged.

    SOURCE · DORYLUS-RUNPOD-DEPLOYMENT-BIBLE.md

  • 09ARCHITECTURE DOCS
    256

    256 architecture documents in 5 months.

    Every non-trivial system gets a written blueprint before a single line of production code. Social platforms. Market-making algorithms. Distributed swarms. Multi-model fusion engines. Reviewed, iterated, built against — never guess-and-check. Verified in the vault index.

    SOURCE · vault-search.js stats

  • 10BIGC BRIDGE
    v3.0

    When the tooling we need does not exist, we build it.

    BigC Bridge (v3.0) is a real-time Claude.ai capture system — Chrome extension, DOM scraper, and TypeScript relay sampling every 5 seconds — that we built for ourselves so AI sessions carry persistent memory across contexts. Running on port 3459, active today.

    SOURCE · claude-capture.js

HOW WE WORK

Three ways in. Every engagement ends in production.

A short audit. A full build. A long-term operator. Pick the one that fits. We’ll tell you on the call if it doesn’t.

1–2 weeks

Audit

We read your stack and write you a report.

OUTCOME

You get clarity. We don’t sell you anything after.

Learn more about Audit →
4–12 weeks, scoped per engagement

Build

We ship the system. Measured in production.

OUTCOME

A working system running under real load, owned by you.

Learn more about Build →
Monthly retainer, six-month minimum

Operate

We run it. You don’t hire an ML team.

OUTCOME

A running system, improved month over month, on a fixed cadence.

Learn more about Operate →
See all three tiers in detail
SELECTED WORK

Built it. Shipped it. Still running it.

A note from the founder

/* PHASE 2: CONTENT */

PAULIEWOOD · FOUNDER · SPACE BOT ENGINEERING STUDIO
Read the origin story