João Raul Furtado Lucas
Forward deployed engineer · Porto Alegre, BR
I build back-end systems that have to stay correct under load. Market data at Nelogica, then payments and e-invoicing at Singular. Now at Telepatia, putting clinical AI inside the hospital systems it has to work with.
Impact
Case studies
Approval workflow for high-value transactions
Problem
On a platform serving ~15,000 users, every transaction followed the same path regardless of value, and nobody could reliably say who signed off on what.
Ownership
Decisions
- Threshold-based routing. Only transactions above a limit enter the flow.
- Limits per business unit, set by the client's management, no deploy needed.
- Full audit trail. Who decided, why, and who issued the invoice.
Outcome
R$230M+ a month under governance. Every high-value decision has a named owner and a recorded reason.
Reconciliation pipeline redesign
Problem
The pipeline exhausted the connection pool. Requests timed out and contract balances stopped updating.
Ownership
Decisions
- 12 Redis-backed workers, sized from work time, run budget and spare connections.
- Batching: ~500K → ~100K queries per run.
- Cache for rarely-changing reference data.
Outcome
Peak connections ~400 → 12, runtime 40 → 15 minutes.
End-of-day closing for an asset manager
Problem
The closing processed every record one at a time, on 5–6M records a day.
Ownership
Decisions
- Execution plans first, then targeted indexes.
- Batched writes instead of row-by-row. The biggest win.
Outcome
−80% execution time.
Things I made
-
A second model reviews my coding agent's plans before they land.
-
Pulls Jira, GitHub and Azure DevOps activity and writes the timesheet.
-
Farm management that works offline, because farms rarely have signal.
On AI: agents help me write and test code. The decisions, reviews and numbers here are mine.
Writing
-
A SourceMod plugin and a Python bridge let Julia 1, a small open model from Supersonic Labs, pick a bot's next move in Counter-Strike in about 28 ms per decision.
-
I ran 81 BenchmarkDotNet cases comparing System.Text.Json, MessagePack and protobuf-net. Binary was 14× faster on numbers and tied with JSON on text.
-
A timesheet tool that maps Jira tickets and Git commits onto working days with Gemini. Reliability came from deleting the MCP layer and a time-tracking integration.
-
Wiring the ralph-loop skill to the Superpowers skill set so Claude Code works through design, planning, implementation, verification and review, looping back on failures.
-
A weekly scheduled task scans my git history, Jira tickets and PRs, finds the work that never had a ticket, and writes it all to a Notion page.
Experience
- Telepatia AI · Forward deployed engineer
- Singular · Software engineer
- Nelogica · Software engineer