9 MIN READ ON OVERLIFT.AI
An agentic system, small enough to inspect
Follow a shipment from question to evidence, approval, and replay inside the browser.
Read the summarySEAN FINDLEY / SELECTED WRITING
Ideas from
the work.
What I’m learning about building useful, accountable AI systems.
These notes connect my engineering practice to OverLift, the product I’m building. Browse the summaries here; read the complete articles and try the related experiences at OverLift.ai.
Explore the projectsStart here
Foundations
21 MIN READ ON OVERLIFT.AI
Deterministic authority: deciding what AI may change
Separate a useful proposal from permission to change accepted state.
Read the summary19 MIN READ ON OVERLIFT.AI
Inside the OverLift WASM Kernel
How focused WebAssembly modules check permissions, execute changes, and preserve receipts.
Read the summary25 MIN READ ON OVERLIFT.AI
The browser as an AI runtime
Put useful intelligence near the user while keeping shared authority explicit.
Read the summary24 MIN READ ON OVERLIFT.AI
Semantic search is not truth
Retrieve by meaning, then check source authority, freshness, and support.
Read the summary20 MIN READ ON OVERLIFT.AI
Q-Lens: compare possibilities before committing
Explore candidate paths with classical computation and make uncertainty visible.
Read the summary29 MIN READ ON OVERLIFT.AI
WebGPU: faster paths must earn authority
Qualify acceleration against a reference result and recover when the device fails.
Read the summary22 MIN READ ON OVERLIFT.AI
AI security begins with authority
Contain code, limit capabilities, and verify the exact effects a system permits.
Read the summary30 MIN READ ON OVERLIFT.AI
Measure the system, not the demo
Evaluate complete tasks, unsupported answers, authority failures, and recovery.
Read the summaryEngineering
5 MIN READ ON OVERLIFT.AI
MCP, A2A, and the authority between them
Tool connections and agent collaboration still need application-level permission.
Read the summary6 MIN READ ON OVERLIFT.AI
One architecture, different consequences
Keep the core boundaries while qualifying evidence, language, and risk for each domain.
Read the summary9 MIN READ ON OVERLIFT.AI
Multi-agent systems need a world authority
Coordinate specialist proposals against one accepted world state.
Read the summary12 MIN READ ON OVERLIFT.AI
Local-first AI is a product decision
Design for private state, real device limits, offline work, and recovery.
Read the summary8 MIN READ ON OVERLIFT.AI
Browser intelligence, server authority
The hosting split behind the original .NET and Azure application, and what it teaches.
Read the summary9 MIN READ ON OVERLIFT.AI
Real-time systems test the architecture
Deadlines, input, sound, and world state expose weak execution boundaries.
Read the summary11 MIN READ ON OVERLIFT.AI
From prototype to proof
Turn a product hypothesis into testable contracts and evidence for release.
Read the summaryPeople & evidence
16 MIN READ ON OVERLIFT.AI
Better writing must preserve meaning
Protect facts, numbers, uncertainty, and authorship while improving the prose.
Read the summary10 MIN READ ON OVERLIFT.AI
What was knowable then?
Keep later evidence from silently rewriting the record of an earlier decision.
Read the summary7 MIN READ ON OVERLIFT.AI
When AI acts around people
Keep consent current, uncertainty visible, and interruption meaningful.
Read the summary