Leonardo Cardoso's AI portfolio. Local-first agents, schedulers, routers, and pipelines, plus consultancy and project work. 2026-06-18T00:00:00+02:00 https://a7t.ai/feed.xml Leonardo Cardoso 2026-06-18T00:00:00+02:00 2026-06-18T00:00:00+02:00 https://a7t.ai/articles/moe-speedup-measured/

A 50-prompt head to head between a dense 27B (Q8) and a 35B Mixture-of-Experts model on a 48GB Mac Studio. The MoE is 4.5 times faster at near-parity quality, and the one trade-off it showed at Q5 disappears at Q6, which is what I now run.

Leonardo Cardoso 2026-06-09T00:00:00+02:00 2026-06-09T00:00:00+02:00 https://a7t.ai/articles/context-layer-bottleneck/

An engineer's look at why vector RAG hits a ceiling on relational enterprise data, what graph-structured context unlocks, and where MCP fits in.

Leonardo Cardoso 2026-04-09T00:00:00+02:00 2026-04-09T00:00:00+02:00 https://a7t.ai/articles/almost-sherlocked-by-anthropic/

Anthropic solved the infrastructure problem the same day I shipped the workflow one. The value is moving up the stack and the two pieces compose.

Leonardo Cardoso 2026-04-08T00:00:00+02:00 2026-04-08T00:00:00+02:00 https://a7t.ai/articles/orchestrating-agents-with-github-actions/

Why GitHub Actions beats the official action for long-running agents, how six labels do the work of a dependency graph, and the numbers from running this at scale.

Leonardo Cardoso 2026-04-01T00:00:00+02:00 2026-04-01T00:00:00+02:00 https://a7t.ai/articles/how-i-built-an-agent-factory/

Principles for running an autonomous coding pipeline: keep it boring, separate research from execution, make 'done' provable, and treat your rules file as the OS.

Leonardo Cardoso 2026-03-23T00:00:00+01:00 2026-03-23T00:00:00+01:00 https://a7t.ai/articles/cognitive-debt/

Why clean code from autonomous agents can still leave you in the dark, and what to actually do about it.

Leonardo Cardoso 2026-03-20T00:00:00+01:00 2026-03-20T00:00:00+01:00 https://a7t.ai/articles/how-autonomous-coding-agents-are-changing-solo-dev-productivity/

Why products that needed three engineers can now be built by one. And the single use case that changed my mind: agent-driven security review.

Leonardo Cardoso 2026-03-19T00:00:00+01:00 2026-03-19T00:00:00+01:00 https://a7t.ai/articles/let-ai-agents-handle-refactoring/

If you only use AI for new features, you're missing its biggest superpower. Three iOS migrations that an agent handled while I reviewed the output.

Leonardo Cardoso