Who decides, how fast we ship

Today Silo is a one-person company. I set product direction, write the code, and use AI as leverage inside the build loop rather than as a substitute for judgment.

Team

I am a practicing Brazilian lawyer, admitted under OAB/PR 96.036, with work across corporate and litigation matters. I am also the person writing the software. That combination is not a branding line. It is the reason the product keeps gravitating toward real workflow pain instead of generic AI demo logic.

Today the team is me. There is no management layer being assembled ahead of validation, and there is no artificial org chart to make the company look larger than it is. The practitioner-as-builder bet is that many of the early errors in legal AI are errors of problem definition, not raw execution capacity. I would rather stay small while the core workflow is still hardening, then add people in response to a proved bottleneck.

Operating model

The operating model is short-loop and judgment-heavy by design. I try to resolve architecture and product questions as close as possible to a working artifact: a narrower interface, a real case trace, a graph query, a failing path, or a live reviewer surface. That keeps the system falsifiable. If something cannot be made legible in the product, it is usually not clear enough to keep.

Agentic coding is part of that loop, but not the owner of it. I use it to expand implementation bandwidth: drafting alternatives, reviewing code, pressure-testing decisions, and moving faster across integration work. The critical decisions still stay with me: what counts as a correct legal structure, what standard an output has to meet, which risks are unacceptable, and whether something is ready to ship. That compression of implementation into the same loop as product and legal judgment is the reason a one-person team can move at this pace.

Cadence

• Since the first repository in this stack was created on January 23, 2026, 1,522 commits have landed on the current default branches across 11 repositories.
• Across the public-facing web surface, 309 production deployments have landed since those projects were created on Vercel.
• The shape of the work is not one application with a thin wrapper. It is a multi-repository system spanning ingestion, extraction, retrieval, agent orchestration, evaluation, and the public review surface.

What agentic development changed

This would not have been a one-person company without agentic coding. Since January 23, 2026, the stack has grown to 11 repositories, 1,522 commits on default branches, roughly 250,000 lines of source and tests, and 309 production deployments across the public web surface. On a conventional path, I would estimate this scope at roughly 12,000 to 15,000 engineering hours: something like a 6-8 person team over 18-24 months, with a fully loaded cost in roughly the $2M-$6M range.

What agentic coding made possible was implementation throughput: scaffolding, refactors, test writing, code review, integration work, and repeated iteration across a wide surface area. What it did not do was the critical work: product direction, legal judgment, ontology and taxonomy design, system architecture, evaluation standards, factual validation, security decisions, and the final call on what ships.

Those human steps are the project. If the legal structure is wrong, the graph is useless. If the evaluation standard is weak, the outputs are not trustworthy. If the security and provenance model are wrong, the system is not usable in serious legal work. Agentic coding accelerated construction; it did not replace the judgment that makes Silo defensible.

What this document does not cover

• Detailed hiring plans
• Compensation structure
• Equity and ownership details

Those are available on request where they matter to the discussion.