Senior Staff Software Engineer at Bags. I build financial systems and the AI agents that run on them: twenty years from trading desks to LLM accounting, and an engineering practice run with coding agents that tripled the team's merged pull requests in a year.
Former founder and CTO. CFP® professional. Brazil, working with New York.
AI work
- Agent-run engineering at BagsDesigned the pipeline in March 2026 and built it in April: a control plane dispatches coding agents that open pull requests, review them, triage production errors and turn Slack requests into working previews. Scope gates run in CI; a human merges every change. With the same four engineers, merged pull requests went from 213 in Q3 2025 to 762 in Q3 2026. This year 1,012 of the team's 1,726 merged pull requests are mine, 644 written directly and 368 through the pipeline, where I write the task and merge the result. The other three engineers combined shipped 714. Year over year, my output is 10x.
- The reviewing agentReviews every pull request, runs a sandboxed test probe before it reports a finding, and approves only when deterministic gates pass. Fixes its own pull requests on a per-PR budget. 79 of its 80 commits are mine.
- From a bot fleet to one agent at BagsFrom May 2026, fifteen Slack agents served tax, sales, operations, funding and marketing on one shared substrate. In October 2026 I replaced the fleet with one agent, Baggins, on the open-source QM runtime: 17 skills hold the team playbooks, the controller and the customer-support inbox workers run on one host, and the fifteen bots and the old controller host are switched off. Eight pull requests in two weeks, all mine. Customer-facing writes stay propose-then-confirm.
- Production error intakeErrors from monitoring become one root-cause ticket for an agent instead of a burst of duplicates. Slack feature requests become login-free previews through the same pipeline, with the requester approving before execution.
- AI transaction classificationShipped April 2025. Every bank line gets an account from a model, then an accountant review queue. Audited backfills verify postconditions before anything touches the books. A golden dataset in CI scores each model change.
- CFO Chat and the evidence-first harnessAn AI CFO for small businesses that reasons in a bounded observe, inspect, answer loop over findings the server computed, never over raw books. Immutable business binding, fail-closed period handling, at most 8 tool calls, 6 model requests and 60 seconds per run, a monthly spend cap per plan. Built on it: a P&L line explainer that persists every hierarchy explanation, revenue and expense drill-downs by customer and vendor, balance-sheet variance over versioned snapshots, and variance narration that can only choose statements the server wrote. Specified and shipped September 2026.
- Accounting document reviewRule checks, concurrency and prompts in July 2025, scored against a golden dataset in CI. In 2026, a general-ledger pilot: findings grounded in ledger slices, cited by id, with a bounded "not evaluated" instead of a guess.
- Insight tools for internal agentsAggregate-only MCP tools over customer financials, audited per run, so agents can answer business questions without reading a single customer's data.
- SuperfalvaA software house run by agents where a Lean-proved model of each rule is the contract: Modeler, Builder and QA roles in fresh sessions with pinned models and mutant proofs. A human merges.
- FalvaCustom agents for Brazilian operations teams on WhatsApp plus ERP and CRM: a parts assistant for auto-parts sellers, a clinic intake secretary. Each pilot has one narrow promise, a real data path, a human handoff and a per-client runtime.
- MaestroGovernance and organizational memory for companies run by AI agents: is this action ready for a human decision, and what reviewed knowledge should the agent know. A plugin for an open-source agent control plane, 142 commits since May 2026.
- Multi-model review councilPull request review by two frontier models, with a third that verifies each finding before a human reads it.
- Shared agent memoryA wiki that agents compile from raw sessions and read before acting, served over MCP to five harnesses on one machine. 114 curated pages since June 2026.
Work
- Bags, New York (remote)Senior Staff Software Engineer. Core backend, QuickBooks integration, AI classification and analysis, the agent engineering practice.
- HublaSenior Engineering Manager, Senior Staff Engineer.
- ProviSenior Engineering Manager.
- IperumCTO.
- VetyFounder, CEO and CTO. Angel-funded veterinary platform, 350+ veterinarians.
- BTG PactualSenior Software Engineer, then Market Risk Analyst. Seven years on trading and market risk systems at one of Latin America's largest investment banks: trade lifecycle, pricing and risk metrics.
Writing and talks
Essays from incidents, with dates. All essays · RSS
- What breaks when an AI reads a bank statementEighteen months of a language model against real books: the invented account id, the false deletions, the PDF with no numbers, and why approval is not posting.
- When agents write most of the codeGoogle reports 75% of its new code is AI-generated and engineer-approved (April 2026). What that number hides when the code moves money: scope gates, verified findings, and why a human still merges.
- The engineer who became a financial plannerWhy I took the CFP® exam after years of writing trading code, and what it changed about the systems I build.
- Lessons from VetyFour years as a founder, 350 veterinarians, angel money, and what I would not do again.
Press
I give interviews and written quotes on fintech engineering, AI in accounting, and building software companies in Brazil. English or Portuguese. felipealvarengarc@gmail.com
Short bio. Felipe Alvarenga is a Senior Staff Software Engineer at Bags and a former founder and CTO. Over twenty years he has built trading platforms at BTG Pactual, founded the veterinary platform Vety, and led engineering at Provi, Hubla, and Iperum. He is a CFP® professional based in Brazil.