A good system is one that doesn't go down.

A choropleth of Brazil's 5,570 municipalities that draws the same indicator through all five classification methods at once. Where the breaks fall changes the conclusion more than any other decision on the map, and almost no published map reveals which choice was made. It also computes local Moran's I, with neighborhoods extracted from the topology and multiple-comparison correction.

Natural-breaks classification is the exact global optimum in O(k·n·log n): 5,570 municipalities in 2.7 ms against 354.8 ms for the quadratic version, which stays in the repo without serving the app — it exists to prove the fast one reaches the same cost. The /validacao page reruns that proof in the browser, over the data in production.

• Next.js
• TypeScript
• MapLibre GL
• TopoJSON
• Jenks
• Moran's I

Surf forecasting computed per break, not per beach. The public swell grid is 8 km wide and returns identical numbers for neighboring breaks, so the difference between them comes from a physical model: dispersion, shoaling, Snell refraction and breaking depth, plus tide predicted by summing harmonic constituents.

The score never stands alone: it comes broken into the six factors that produced it, each with its measured quantity, and every break parameter declares whether it was measured or estimated. The predicted tide matches the Brazilian Navy's table within 21 mm, and the model's invariants are checked over a thousand generated cases per run.

• Next.js
• TypeScript
• Wave physics
• Harmonic tide
• fast-check
• Zod

A bitemporal energy-billing engine: it ingests readings idempotently and out of order, prices them against a versioned ANEEL tariff (period, flag, moving holiday) and posts to a double-entry ledger that always sums to zero. It keeps history on two axes — valid_time × transaction_time — reproducing any invoice exactly as it was known on any date.

The money invariant doesn't live in a code branch: Postgres rejects an overlapping bitemporal assertion with an EXCLUDE gist, and the zero-sum ledger is proven by property over thousands of scrambled sequences — not claimed.

• Next.js
• TypeScript
• PostgreSQL
• Bitemporal
• Double-entry
• fast-check

An AMI mesh-network simulator with an observability console: meters routing through collectors up to a head-end, live mesh reconvergence, chaos injection and an SLA panel. At its core, a real DLMS/COSEM codec — HDLC, APDUs, OBIS codes — that encodes every reading to bytes and decodes it back.

The whole simulation runs in the browser, on purpose: no backend to time out, free to host on Vercel, and the mesh reconverges in front of you when you kill a collector. It's how I brought what I do in the power sector into something you can open and break.

• Next.js
• TypeScript
• Zustand
• DLMS/COSEM
• OpenTelemetry
• Vitest

A maintenance-shutdown optimizer modeled as RCPSP: thousands of orders, precedences, limited crews and a fixed window. The solver — serial SGS with simulated annealing — compiles to WebAssembly and runs in a Web Worker, on the client. It imports SAP PM exports (IW39/IW49), computes the critical path, date risk via Monte Carlo, and measures its gap against the PSPLIB dataset, which has proven optima.

I spent years planning maintenance shutdowns in SAP, deciding months ahead when each asset would stop. CritPath is that problem turned into a solver: it reads the same SAP export I used and optimizes the schedule — with the gap measured against proven optima, not claimed.

• Next.js
• TypeScript
• WebAssembly
• RCPSP
• Monte Carlo
• Web Workers

In-browser vibration diagnostics for rotating machinery: windowed FFT, Hilbert envelope, and bearing defect frequencies (BPFO, BPFI, BSF) computed from geometry. A rule engine names the fault — outer race, misalignment, looseness — with severity graded by ISO 20816.

I spent years in the refinery diagnosing bearings through vibration analysis. Auscult is that knowledge turned into code: all the signal processing runs in the browser, and the detector is validated against Case Western's bearing dataset — including the cases where it gets it wrong.

• Next.js
• TypeScript
• DSP
• FFT
• ISO 20816
• Vitest

An order-book matching engine — the core of an exchange. It matches buy and sell orders by price-time priority, supports limit, market, IOC, FOK, cancel and replace, and records every trade via event sourcing, with bit-for-bit deterministic replay from a single seed.

In a system that moves money, correctness is the feature. Every invariant — cash and shares conserved, the book never crossed, no one jumping the queue — is proven by property testing over thousands of random order flows, and checked live on screen. Proven, not promised.

• Next.js
• TypeScript
• Matching Engine
• Event Sourcing
• fast-check