AI engineering studio · Santa Catarina, Brazil · UTC-3
Production software, shipped through governed agentic workflows.
Coding agents do the volume. Guardrails, quality gates and evals keep it honest. A senior architect owns the result and signs the merge.
- commits shipped in 2026
- 3,000+
- merged MRs in ten weeks, one platform
- 125
- query latency, 1.29M rows
- 35 s → 2 s
- app stores publishing our product
- 2
What we do
Agentic software delivery
Features, modules and whole products built with coding agents under written process: an issue before every commit, quality gates before every push, adversarial review before every merge, CI guards for schema drift and contract freshness. Velocity with an audit trail.
LLM features with evals
Tool-calling assistants, retrieval and classification shipped with eval sets, guardrail tests and multi-provider fallbacks across Anthropic, OpenAI and open models. Measured before it is called done.
Data and platform engineering
Query engines, lakehouses and cloud cost, measured from outside: replayed production traffic, engine benchmarks, compute-amplification audits and FinOps findings with numbers attached.
How we work
Measure first
We replay real traffic and benchmark before recommending anything. Decisions carry p50s, row counts and costs, not opinions.
Guardrails, not gates
Agents run inside hooks that block unsafe actions and ask before anything irreversible. Humans approve merges; agents never close issues.
Contracts as the source of truth
Versioned Protobuf contracts generate types for every language in the stack, with breaking-change gates and freshness checks in CI.
Written handoffs
Every session ends with a report: what changed, what was tested, what is blocked. You can read the diary at any time, in English or Portuguese.
Selected work
Delivered by our principal engineer. Client names withheld; numbers from the repositories.
Fiber-optics ISP · CRM module
A CRM built from zero inside a live operations platform
28-table schema, 16 migrations, 105 FastAPI endpoints behind 103 permissions; kanban, lead workspace and proposal editor in React 19. Test coverage from nothing to 248 backend tests and 105 end-to-end cases. Pull-request delivery introduced to a team that merged straight to main. 446 hours, part-time.
Trade marketing · field platform
Backend, contracts and an offline-first field app in ten weeks
FastAPI and PostgreSQL backend, Protobuf contracts as the single source of types for TypeScript, Python and Dart, and a Flutter app re-platformed to offline-first. 125 merged MRs across five repositories; tests from 84 to 725 and from 145 to 736; 21 tester builds in 24 days.
BI platform · data engineering
Latency and cost, measured from outside and fixed at the root
A 1.29-million-row paged query from 35 seconds to 2 by pushing limits down to Trino; a 46-million-row out-of-memory aggregation redesigned to 12 seconds per page; Trino benchmarked against StarRocks on replayed dashboards; FinOps findings worth 30 to 60 percent of monthly cloud spend.
Products · in the stores and on the web
Our own products, run the same way
Chordo Pro, a chord-sheet app on the Apple App Store and the Microsoft Store with 449 tests and a CI release pipeline. GeoDome, a geodesic-dome calculator whose geometry passed a 41-finding adversarial review. Alpha Tester, a feedback platform with web and Flutter SDKs.
About
PDB Digital is the studio of Marco Santana, a senior software architect with more than 25 years across Microsoft Brasil, H2O.ai and two decades of SaaS and platform work. He has worked in English with North American and European teams since 2018.
The studio is small on purpose: one accountable architect, a fleet of coding agents under written process, and specialists brought in per engagement. Based in Santa Catarina, Brazil, three hours ahead of US Eastern and one behind Lisbon.
Résumé and profile
LinkedIn
GitHub