Summary
Ahmad Awais, founder of CommandCode.ai and former VP of DevRel at RapidAPI, joins the Latent Space podcast to explain how he engineered a system called Taste that enables DeepSeek V3 to outperform Claude Opus 4.7 on real-world coding tasks. Taste is a meta-neuro-symbolic architecture that learns developer preferences by observing accepts, edits, and rejects across coding sessions, distilling them into per-repository skill files. Rather than relying on RAG over documentation, it encodes the engineer’s lived judgment — things like preferring pnpm for package installs but npm link for local CLI development — directly into the agent’s behavior layer. CommandCode.ai reports 1.2 billion agent runs per month on DeepSeek infrastructure.
Awais traces the lineage of CommandCode from a viral COVID-era terminal project and early GPT-3 access granted by Greg Brockman in July 2020, more than a year before GitHub Copilot launched. The platform was previously known as LangBase, a memory infrastructure layer, before the team concluded that coding agents are the universal interface and pivoted accordingly.
The conversation also covers how the same compositional repair-logic framework has been applied to eliminate ‘design slop’ — the indigo-purple gradient and generic layout patterns endemic to LLM-generated UI — by encoding a finite set of design rules as skill files. Awais argues this approach generalizes across commercial and open models alike, making it a broadly applicable technique for any team that wants to bend a frontier model toward a specific aesthetic or engineering taste.
📺 Source: Latent Space · Published June 06, 2026
🏷️ Format: Interview
