What I'm building toward
A living page about where my attention is and what I think the work is really about. Rewritten when the answer meaningfully changes, not on a schedule.
We are a few years into a shift where software is no longer a tool you pick up and put down — it's an agent you work alongside. Most of the current product surface area still treats AI as a chat window bolted onto an app. I think the real answer looks different: ambient, memory-rich systems that accumulate context with you over time and operate on your behalf in ways you can inspect and trust.
The unsolved problems aren't about model capability. They are about runtime — what the system does when the model is wrong, when context is incomplete, when sessions span weeks, when two agents need to agree on state. That's the layer I want to build at.
- SleepHub. Shipping the next generation of the bedside AI experience at Asleep. Core loop: understand what happened during the night, then turn that into useful daytime action.
- Knowledge OS. A markdown-first personal memory substrate. Files as the source of truth, agents as readers and writers, git as the history. This site is one translated output.
- Agent infrastructure. Session continuity, multi-agent coordination, runtime contracts, and guardrails encoded into the environment rather than left as prompt folklore.
- Validation surfaces. Exploring how AI systems move from internal prototypes to practical deployment, testbeds, and real users.
- Patient–family record sharing. An early exploration of a lightweight digital health layer that helps a patient and a patient-designated family member or care companion stay on the same page after a hospital visit — shared schedules, instructions, preparation items. Scoped to patient-led record sharing and communication outside the hospital, not medical judgment. Shaping the problem with real patient and family input first.
- Writing. Publishing what I learn in Writing. Short, concrete, from the workbench.
- 01 Demo-shaped doesn't scale.
A prompt that produces an impressive output once is not a system. What matters is behavior across sessions, edge cases, partial failures, and humans interrupting. That's the layer where correctness lives.
- 02 Memory should be explicit.
If you can't read, edit, or grep your agent's memory, you can't trust it. Files age well. Plain text survives tool churn. Opaque stores might win on convenience; they lose on auditability.
- 03 Runtime is the product.
Most AI product value is realized in the seams: tool errors, retries, state handoffs, partial context. The "product" is how the system behaves in those moments, not in the happy path.
- 04 Ambient over intrusive.
The best software disappears. AI should earn its presence by being useful quietly, not by asking for attention. SleepHub is a test of this.
- 05 Deployment matters more than posture.
The question is not whether a system sounds impressive, but whether it can survive contact with real operators, real users, and real constraints.
- 06 Build in the open when possible.
Public writing and public artifacts produce better thinking and better systems. Private drafts rot; public ones get corrected.
- 07 Small, durable, composable.
Prefer boring technology I can debug over novelty I can't. Prefer small scripts I can read over frameworks that abstract away the problem. The best code is no code; the best change is the smallest change.
Equally useful to name what's out of scope. I'm not building another agent framework. I'm not chasing benchmarks. I'm not trying to wrap models for generic productivity. The attention goes into the specific surfaces above, and the general lessons get written up in Writing.
This page is the anchor I rewrite when the answer changes. If you're reading it and the dates look stale, the Writing page is probably the more current signal. If any of the above resonates — or you think I'm wrong about something — reach out via X or GitHub.