01
Models and data
Governing question
How much of a model's capability is decided by data and evaluation before architecture receives the credit?
Current position
Model behavior starts below the interface: curation, language coverage, contamination, evaluation, and implementation choices shape what a system can become.
Related writing
No published essay yet; the public work below holds the question open.
Public artifacts
- Opens in a new tab
Project EKA — model and data infrastructure
Multilingual curation infrastructure and reproducible experimentation for a sparse-MoE program, with the program's scale kept distinct from an individual contribution.
- Opens in a new tab
Qwen3 MoE — understanding by rebuilding
A from-scratch PyTorch implementation for understanding sparse routing, grouped-query attention, RoPE, normalization, and KV caching at code level.
02
Agents, memory, and authority
Governing question
What should persist when an agent crosses sessions, tools, and models—and what authority should never move with it?
Current position
Useful continuity needs correctable context, bounded permission, inspectable action, and a person who can intervene.
Related writing
Public artifacts
Waldo
A user-owned personal agent exploring how intent, evidence, correction, and unfinished work can remain coherent across agents and time.
Atlan — production agent systems
Production work across context, tools, authentication, integrations, deployment, and reliability made the gap between a finished run and a completed outcome impossible to ignore.
- Opens in a new tab
Systems Around Models
A practical fieldbook where I share what I’m learning about harnesses, memory, authority, recovery, evidence, and completion.
03
Outcome truth and evaluation
Governing question
When an agent says done, what became true—and what still belongs to the person?
Current position
A completed run, an artifact, evidence, an accepted outcome, and an open loop are different kinds of truth.
Related writing
Public artifacts
Atlan — production agent systems
Production work across context, tools, authentication, integrations, deployment, and reliability made the gap between a finished run and a completed outcome impossible to ignore.
Waldo
A user-owned personal agent exploring how intent, evidence, correction, and unfinished work can remain coherent across agents and time.
04
Agents in the world
Governing question
How should agent state and control change when action crosses into physical environments?
Current position
Physical systems keep uncertainty, timing, material cost, safety, and irreversible consequences in view.