// BLOG
Latest notes.
Notes from inside the build. No thought-leadership, no recycled takes, just what I figured out the hard way so the next person doesn't have to.
Automating Social Media Posts With AI, Without the Slop
I'm building Latent, an AI post generator. Here's how to automate social posts end to end: exact platform sizes, human review via Kanban, and zero AI slop.
LGTM: A Claude Code Orchestration Framework I Built
LGTM runs parallel Claude Code agents in isolated git worktrees, one PR per issue, human merges everything. The architecture and the design decisions behind it.
Small Language Models for Vertical Agents
Small models fine-tuned for a narrow task outperform generalist frontier models on that task, cost a fraction to run, and unlock deployment options the big ones cannot touch.
AI Evals in Production: The Work Nobody Sees Until It Breaks
A model that scored well on benchmarks will quietly rot in production. Evals are the unglamorous work that turns an AI demo into a system you can operate.
LLM Cost Control: Strategies That Actually Move the Bill
Most LLM cost advice stops at "use a smaller model." Real bills get reduced by caching, routing, context discipline, and knowing which 20% of calls drive 80% of spend.
Agentic RAG: When Retrieval Becomes a Reasoning Step
Traditional RAG retrieves once and hopes. Agentic RAG treats retrieval as a loop the model controls, which changes everything about how you build the pipeline.
MCP Servers in Production: What the Hype Misses
Model Context Protocol finally gives agents a clean contract for tool use, but the production story depends on auth, versioning, and rate limits most teams skip.