01
Give machines context.
Agents need more than prompts. They need durable memory, visible provenance, recovery paths, and the good sense to ask before they act.
- memory
- provenance
- judgment
Alvis Tang · AI founder · London
Context for AI.Purpose for humans.
I build systems where agents remember what matters, tools explain themselves, and people recover the purpose automation too often forgets.
Math PhD by chance persistent enthusiast by choice
The Analytical Engine has no pretensions whatever to originate anything.Ada Lovelace, 1843
— The premise · 01 —
Execution is getting cheap. Direction is becoming sacred.
The world already has enough output. I am interested in AI that understands the situation around the task: memory, provenance, boundaries, uncertainty, and taste.
Because once machines become good at execution, the question is no longer “can we do more?” It is “what is still worth doing—and who gets to choose?”
One mission / two halves
02
Return humans to purpose.
When execution becomes abundant, authorship becomes precious. I build for work that protects agency, meaning, and the part only a person can choose.
- agency
- meaning
- authorship
In the arena / 02
One company. Two connected questions.
01 · Primary company · Context-aware AI infrastructure
Contexta
A system for giving AI agents durable working memory, provenance, and context that can be inspected, updated, and trusted.
↗02 · Exploring · AI-native workflow and virtual employees
Amino.ai
An exploration of virtual employees: agents that can own bounded work, report progress, ask for judgment, and improve through feedback.
↗03 · Researching · Purpose, careers, and future-of-work systems
Work for Future
A research and product thread about meaningful work, career agency, and the tools people need to shape better futures.
↗Shipped proof
Presetter v9
A configuration system for TypeScript projects and monorepos that keeps shared engineering decisions composed, reviewable, and current. The public release, migration guide, and repository show how I turn a thesis about memory into infrastructure other engineers can inspect.
The throughline / 03
Math PhD by chance. Builder by choice.
Probability taught me to respect uncertainty, hidden structure, and answers that arrive a little too cleanly.
Building taught me that rigor only counts when a real person can hold it, use it, and trust it. That is the work I keep choosing.
Thinking in public
Ideas are part of the work.
Notes on agents, context, tools, and the human judgment that makes any of them worth building.
Configuration Should Have a Memory: Presetter v9
Presetter began as a way to stop copying config files. Version 9 turns that small irritation into a clearer system for preserving engineering decisions across projects, packages, and time.
PresetterTypeScriptOpen sourceDeveloper tools
On the Context Harness
A note on why the apparatus around the model — and not the model itself — has quietly become the substance of intelligent work.
Context harnessAgentsPhilosophy of tools
A Harness in Eighty Lines
The smallest loop that calls a model, dispatches its tools, keeps a memory, and decides when to stop — and what each of those choices is actually saying.
PythonAgentsEngineering
The work stays public
The machine may execute. The human must still mean it.
AI founder · open-source builder · Math PhD, Imperial College London · persistent enthusiast