mansio
@ManSio
Mikhail
Painter by trade, AI architecture researcher by practice.
About
Officially working as a painter, building practical software tools in my free time. I focus on logic and execution mechanics—developing tools like MSCodeBase to keep AI models grounded in real data rather than statistical guesswork.
Language & Tools
Python, Zed IDE, Git, Termux, AST parsing (tree-sitter), Vector Search (LanceDB).
Currently Exploring
Running local LLMs on mobile hardware, zero-dependency decision primitives, and deterministic verification bounds.
Achievements
Built MSCodeBase, an open-source MCP server designed to ground AI agents in real code context.
Conducted an empirical study benchmarking 14 LLMs across 3,300+ API calls to evaluate model memory and hallucination boundaries.
Contributed evaluation controls to open-source developer tooling on GitHub.
Conducted an empirical study benchmarking 14 LLMs across 3,300+ API calls to evaluate model memory and hallucination boundaries.
Contributed evaluation controls to open-source developer tooling on GitHub.
Fun Fact
Painting walls by day, building deterministic control harnesses for AI agents by night.
Random Dev Quote
Without hard verification boundaries, an AI agent's persistent memory just turns an honest unknown state into a structural guess.
Contributions
Comments
- @[helkyn-coello]
Nailed it. 'The structure IS the information' — th...Aug 30, 2026
- @[helkyn-coello]
That’s a gold-standard distinction: Independence b...Aug 30, 2026
- @[helkyn-coello]
"Touché. 'Deferred and shipped anyway' is definitel...Aug 30, 2026
- Spot on, Helkyn!
Everyone is obsessed with model accuracy metrics, bu...Aug 30, 2026
- Excellent reflection, Helkyn!
That observation — "workflows that ne...Aug 30, 2026
Saved Posts
chevron_left