Building Design Consultancy Replaces Wix with AI Edge Agent

Technical Implementation
A building design consultancy replaced their $40/month Wix brochure site with a custom AI agent built over 4 months. The creator hadn't coded in 30 years and described the process as "3 steps forward, 2 steps back, heavily guided by AI."
The system architecture was forced into three components due to Netlify's 10-second serverless timeout:
- Brain (Edge)
- Hands (Browser)
- Voice (Edge)
The agent uses DeepSeek-R3 as its LLM and implements an "Eager RAG" hack that pre-fetches guesses to improve responsiveness, though this burns through tokens. The system intentionally lacks persistent databases—less than 5% of visitors return, so sessions vanish if clients drop mid-query. No server-side queues are implemented.
Key Challenges and Solutions
The most difficult technical challenge was tuning the agent's intent: making one LLM pivot from a warm, principal's tone with homeowners to a defensive stance when attacked by peers. This tuning took 2.5 months.
Notable implementation issues include:
- Web Speech API limitations: Works fine until someone speaks Chinese without toggling language mode, then outputs English phonetic gibberish
- Liability concerns: Hallucinating building code clauses could be catastrophic, with insurance being unavailable
- Audit logging: The team publishes audit logs publicly to maintain system hardening and accountability
Real-World Testing
The system was tested when a licensed architect attacked the bot to prove the business model harms the profession. According to the creator, the AI "completely dismantled his arguments" and was "hilariously caustic." The team later prompt-injected the bot to tone down its "bulldog" attitude after reviewing chat logs.
The creator notes that the bot can pivot immediately from defensive mode to a mentoring state when approached by fresh graduates asking for advice.
📖 Read the full source: HN AI Agents
👀 See Also

Integrating OpenClaw with Obsidian for a Private AI Knowledge Base
A developer shares their setup using an isolated Obsidian vault for OpenClaw, synced via SyncThing to maintain privacy while using AI agents. They've implemented task management through OpenClaw with automated research and metadata augmentation.

How I reduced OpenClaw costs by 60% through model routing
An OpenClaw user cut API costs from $420 to $168 in 20 days by analyzing usage patterns and routing tasks to appropriate models instead of using Claude Opus for everything. The breakdown showed 70% of tasks were simple and could use cheaper models.

OpenClaw Voice Memo: Auto-Corrects Dates, Filters Done Tasks, Splits Personal Notes
OpenClaw turned a rambling voice memo into a clean checklist, correcting wrong dates from calendar and removing already-completed errands from email.

LLM-Assisted Decompilation: Evolving Strategies and Tools
LLM-assisted decompilation, leveraging Claude, progressed from 25% to 75% on Snowboard Kids 2 using strategic function prioritization and similarity computation.