01 / Building

ZaimuTomo

I am building a personal accounting application in Elixir and Phoenix. The current work is mostly around invoice ingestion, document understanding, structured extraction, a separate semantic verification pass, and human review.

The useful question is not whether a model can extract an invoice. It is how the system exposes unsupported or ambiguous fields, preserves corrections, and improves without silently moving accounting decisions into the model.

02 / Exploring

Building a small software factory

I am experimenting with a workflow built around Hermes Cloud Agent, its Kanban board, and isolated Fly.io Sprites. The aim is to turn scoped work into testable, reviewable changes without losing human control of delivery.

I am interested in which parts need durable orchestration and which are better left as simple agent loops.

03 / Learning

Local inference and multimodal search

A recent proof of concept uses SigLIP 2, ONNX, PostgreSQL with pgvector, and a small local inference service to search an image collection with natural language. It is a useful testbed for model-serving boundaries, local hardware, and deliberately boring storage choices.

04 / Writing

Management problems with technical consequences

I want to write about capacity planning, succession, distributed teams, scope control, and staying technical without becoming the team’s bottleneck. I also want to document the systems work behind observability, testing, developer tooling, and applied AI.

This page changes when my focus changes, rather than on a fixed schedule.