The Axoworks Logs: Surviving the Conversational Web

Part 1: The Vibe Coding Mirage and the 10-Second Boss

Disruptive tech doesn’t happen because a new model is 5% more efficient. It happens because people are just sick of getting force fed the status quo.

End of 2025, the corporate internet had become a digital graveyard. Every AEC website was the same deadbeat Wix or WordPress template going nowhere, forcing you to click through a dozen disjointed pages just to figure out if a firm could actually build your project. Clients want a straight answer, instantly.

We didn't set out to build a chatbot MVP; we set out to permanently delete our company brochure. We replaced it with a multi-model AI terminal—The Concierge.

The initial trap was "vibe coding." The illusion that anyone can spin up a DIY SaaS app over the weekend if they just prompt an AI hard enough. The reality is a brutal loop of four steps forward, three steps back. You fix three minor UI glitches, hit a foundational logic bug, and watch the AI obsessively unravel a week’s worth of functioning code in a blind, OCD debugging loop. You learn fast: you aren't writing code anymore; you are managing a chaotic junior developer running at lightspeed.

Then we hit the deployment wall: The 10-Second Boss.

Standard serverless platforms like Netlify work great for static sites, but they have a hard, 10-second execution timeout. When your LLM is deep in cognition, cross-referencing global architecture codes, 10 seconds is a blink. The connection just snaps. To survive, we had to rip up the floorboards. We spent a month refactoring the entire architecture, moving the "brain" completely to the edge. We decoupled the client from the heavy lifting, building complex middleware to stream responses and keep the connection alive.

The brochure was dead. The terminal was breathing. But we hadn't given it a soul yet.

Part 2: The Hacker News Crucible and the Metacognitive Straitjacket

Getting an AI to talk is easy. Getting an AI to shut up and do its job is an engineering nightmare.

Initially, we shoved our entire 30-year operational ethos into one super-duper master system prompt. That works beautifully if you have 20 polite visitors a day. But then we hit #2 on Show HN.

When 2,000+ engineers bombard your terminal in six hours, they aren't there to look at your portfolio. They are there to jailbreak your persona, extract your backend instructions, and break your API budget. Under that weight, the master prompt collapsed. The LLM suffered from severe context rot, losing its boundaries and hallucinating wildly.

Worse, this eagerness to please bled into our core AEC workflows. When we tasked the Concierge with QA audits on design schedules, it became an overzealous inspector. It suffered massive scope creep, silently dragging in non-applicability criteria. It would flag perfectly compliant interior egress paths because it decided to cross-reference structural steel tolerances that weren't even in the model.

We had to physically decapitate the system prompt.

We migrated the knowledge base into a strict RAG (Retrieval-Augmented Generation) pipeline. The AI was no longer allowed to just "know" things; it had to fetch vectorized, sanitized data for every response. We engineered a "metacognitive straitjacket" into the middleware. Before the AI can evaluate any QA criteria, it is mathematically forced to prove that the rule applies to the current project scope. If it can't, the criteria is sandboxed.

We stopped trying to control the AI's personality and started strictly controlling its memory.

Part 3: OpenClaw and the Dawn of the Killer Agents

By early 2026, the human wave attacks from Hacker News felt like a nostalgic warmup. The open web had changed.

The line between a human actor pulling a prank and an uncontrolled, autonomous agent going wild vanished entirely. The explosion of open-source, agentic AI like OpenClaw means we are no longer just defending against script kiddies. We are defending against headless swarms that can read, reason, and execute toolchain commands.

An OpenClaw agent isn’t browsing your site. It is systematically probing your RAG database, injecting malicious payloads hidden in plain text, and mapping your backend to hijack the conversational state.

A friendly Concierge built for January 2026 will be eaten alive in 2027. We had to evolve the terminal into a paranoid sentry.

We built dynamic state detection that analyzes interaction pacing, instantly distinguishing between a human visitor asking about design workflows and a bot executing a scrape script. If the primary LLM chokes under an attack or throws a 500 error, our routing system instantly triggers a failover route to a secondary model family, ensuring the terminal never goes dark.

And if you want global reach? You hit the browser API stone age. Handing international users native browser language detection, Text-to-Speech, and Speech-to-Text across the chaotic landscape of iOS, Android, and Linux is a fresh hell of inconsistent protocols.

We started this journey to solve the frustrating, dumb simplicity of the strip mall corporate website. We ended up having to build a complex, edge-deployed fortress just to survive the conversational web. The AEC industry laggards are still optimizing static pages for 1999 Google crawlers. They have no idea what is coming over the horizon.