The work I consider most significant — the highlights, without reading everything.
2022
Led a team impacting 90+ people
Led and guided a volunteer team on the Folhas Que Salvam educational environmental project, developing company policies and internal control procedures that secured the project's legal validity and financial transparency, and defining achievement metrics that raised the consistency and quality of weekly deliverables — work that impacted more than 90 people.
2026-09
CAIXA credit simulation ~2 min by hand (the correspondent's estimate) → 15 s in live tests on real buyers' data; every form field filled from the documents, checked by 8 pre-submission rules
Gestão Minha Nova.Casa is an internal tool for a CAIXA Aqui correspondent, the partner that takes home buyers through CAIXA's mortgage process. The team drops a buyer's documents into the app; an LLM extracts the data, the app builds the buyer's record, checks it against a catalogue of 8 pre-submission rules (CPF check digit, postcode against address, validity of the proof of income, extracted income against declared income, highly variable income, maximum age, maximum term, income commitment), and, at one click, an agent drives Chrome on the attendant's own machine to run CAIXA's mortgage simulation and saves the result, which CAIXA's site does not keep — so nobody retypes anything. CAIXA changes its site's layout often but not its fields, as the correspondent confirmed, so the agent fills the form by reading the whole screen rather than relying on fixed selectors; when it cannot find a field or CAIXA rejects one, it stops, the attendant completes that step by hand, and the agent resumes. A simulation the correspondent estimates at about 2 minutes by hand took 15 seconds in live tests with the correspondent on real buyers' data, and when the documents are sent complete every form field is filled from them, never from the conversation with the buyer. Daniel chose the extraction model by measuring four of them on 16 real customer documents (82 fields): the one picked read 89% of fields correctly with 2 silent errors (a wrong value at confidence 0.9 or above), at about US$ 0.05 per buyer, over a model with the same accuracy that never answers "unreadable" and made more than three times the silent errors. Since no model read every photo right — all four got the same birth date wrong with high confidence — he designed the extraction to be verified, not just read: every field ends either confirmed by a check that cannot be wrong (CPF and CNH check digits, payslip arithmetic) or seen by a person with the document beside it, and all 13 wrong CPF and CNH readings in the test failed their check digit. Daniel ran the discovery with the correspondent and built the whole system — the Electron desktop app and its browser agent, the NestJS server with its AI gateway, the partner portal and the sales site — everything except the brand name and the visual identity, which came from a landing page designed by the CodeSquare team. It is in acceptance testing with the correspondent, in a staging environment, ahead of release. Built with TypeScript, Electron, Playwright, NestJS, Drizzle ORM, PostgreSQL, better-auth, Zod, React, TanStack Router, TanStack Start, OpenAI API, Docker, Caddy, CSS Modules, Zustand, TanStack Query.
2026-09
Engineer's full analysis 3-4 days → under 1 day in a side-by-side test at the Uberlândia city hall (two engineers, same application, one with the system)
Square AI is CodeSquare's own AI pre-analysis system for building-permit applications: it checks an application's documents, zoning indices and architectural plans against municipal law, and returns a checklist of compliant / non-compliant / not-applicable / inconclusive items with evidence and a suggested opinion — the AI does not approve or reject; the engineer decides. Daniel built everything except the legislation data collection and the RAG index: the API, the queue executor, PDF and DWG extraction, the deterministic rule checks, the phased LLM analysis constrained by JSON Schema, and a review phase that filters false alarms. DWG plans, a closed format, are converted with LibreDWG into one the system can read, so rules such as a balcony's maximum depth and area or a bathroom's minimum area are checked against the plan itself. The plan is read twice — by an LLM and by the LibreDWG reader, matched through the room labels the designers write — and a room the two do not agree on, or drawn outside the convention, is marked "cannot verify" and left to the engineer, who can explain how to read it; that explanation feeds a list of keywords per room type that grows as the system is used. In a test at the Uberlândia city hall, two engineers analysed the same application, one with the system and one without: pointing the engineer to what to verify cut a full analysis from 3-4 days to under a day. The system has been submitted to a public tender and awaits budget for implementation. Built with Python, PostgreSQL (pgvector), MinIO/S3, PyMuPDF, LibreDWG, LLM with JSON Schema, Docker.
2026-09
Tailored résumé: ~2 hours to 3 minutes
A career Corpus that treats professional history as versioned source: Affiliations and Accomplishments stored as schema-validated markdown, read through a typed loader and rendered per audience. Daniel authored the three agent skills that drive it — capture interviews him to record an Accomplishment without ever inventing a number, generate turns a job posting into a structured Selection that a fixed template renders to PDF through Tectonic, and recon produces a verified dossier on a target company plus a direct approach to its decision-maker. The model never writes LaTeX and never asserts a fact absent from the Corpus: a malformed Selection fails to render rather than quietly producing a different document, and an Accomplishment with no metric is kept as an explicit draft instead of an embellished claim. Producing a résumé tailored to a posting went from about two hours of research and writing to three minutes. The public site is one Render Target of the same Corpus. Built with Next.js, React, TypeScript, Tailwind CSS, Framer Motion, Tectonic (LaTeX), Node.js. Live: https://www.teamdbsolutions.com Repo: https://github.com/daniel-bernardino747/personal-website
2026-10
8 of 8 chaos scenarios end with the oracle's ledger hash
Designed, directed and architected a system that keeps a per-user ledger of bank transactions in sync with Plaid and stays correct when things fail halfway, with every decision recorded in 13 ADRs. All 8 chaos scenarios, run as real processes — SIGKILL mid-pagination and mid-commit, every webhook delivered 10 times, two concurrent workers, 30% of Plaid calls answering 500 or timing out — end with the same ledger hash as the provider's ground truth. Built with TypeScript, Next.js, Cloudflare Workers, Node.js, PostgreSQL, Prisma, pnpm, Turborepo, Vitest. Repo: https://github.com/daniel-bernardino747/plaid-sync-ledger
2026-10
23 of 24 searches ranked the right note first across 14 real notes, written before the notes were saved; plan to production in 2 days
Designed and directed scholia, a self-hosted second brain for AI agents: a remote MCP server where Claude saves distilled notes from conversations (decisions, data with sources, positions, open questions) and searches them before answering, with hybrid search that fuses pgvector similarity and Postgres full-text through reciprocal rank fusion. AI agents implemented it under Daniel's direction, from first commit to a production deploy on Railway with GitHub OAuth in 2 days. Tested on his real knowledge base (14 notes in Portuguese, 13 of them distilled from a week of political conversations), with 24 questions written before any note was saved, 12 by meaning and 12 by exact term or acronym, by the same agent that wrote the notes: the right note came first in 23 and in the top 3 in all 24. The similarity floor was calibrated on real scores and left off, because related and unrelated notes overlapped. Built with Python, FastMCP, PostgreSQL (pgvector, full-text search), Voyage AI and OpenAI embeddings, GitHub OAuth, Docker, Railway. Repo: https://github.com/daniel-bernardino747/scholia-mcp
2025
Main kanban board load 7s → 1.5s (team testers, browser Network tab); all 5 client teams that threatened to leave stayed
Homeet is CodeSquare's own work-management platform — kanban, task manager, timesheet and per-task time tracking, in effect a Jira of its own — used by the software house's client teams. It already existed when Daniel joined in October 2025, and its main kanban board took 7 seconds to load; five client teams had told the co-founders to fix it or they would leave. He found four causes and fixed each. Components called oversized endpoints to read one or two fields — one request per card for the assignee's photo, every project loaded on the home screen — so he wrote narrow endpoints and cached their data in the front end. The SortableJS drag and drop broke under caching and real-time updates, so he replaced it with Atlassian's Pragmatic drag and drop. Every card move sent its own request, so he batched a user's moves into one request per 1.5-second window, the last write winning. Cards were ordered by an increasing integer that let two cards share a position with no tie-break, so each user saw a different order; he replaced it with alphanumeric keys placed between neighbours and recalculated when the gap between two cards runs out. The largest fix was moving business logic out of the browser: card position and the movement history ("so-and-so moved the card to X at this time") were computed inside the board's onDragEnd handler and repeated wherever a card could move, so no one knew every place a change had to touch and every change broke something; he moved them to the NestJS back end and left the front end to render. Main-board load fell from 7s to 1.5s, measured by the team's testers in the browser's Network tab, and browser memory by 30%, also measured by the testers; all five client teams stayed and were still using Homeet when Daniel left in October 2026. The gain was not the main board's alone: by Daniel's account, the system's slow paths — page loads, sending several card moves at once, real-time updates — improved in the same measure, about 79% (the board's 7s → 1.5s), and that system-wide improvement is why the client teams stayed. Built with Next.js, React, NestJS, Prisma, Supabase, Pragmatic drag and drop, TypeScript, Tailwind CSS, Zustand, TanStack Query.
2024
Delivery time 4 days → 1 day; team output +200%
Engineered a high-performance culture by implementing a weekly feedback loop and data-driven management adjustments, reducing delivery time for core tasks from 4 days to 1 day and lifting overall team output by 200%.