I build the parts that have to be correct, not just look right.

polygon-offset-exact

Offsetting a polygon so that every point of the result is actually at the distance you asked for. The intuitive method — push each vertex along its angle bisector — puts the corner at d / sin(θ/2) instead of d; on a star offset by 40 the worst point lands 121.8 units out. It still looks like an offset, which is the problem.

The library ships maxDeviation as part of its public API, because a picture cannot tell you an offset is correct. It also ships the naive method, so the error is measurable rather than asserted. 57 tests, the core of them one property asserted across five shapes: every returned point is at the requested distance, to 1e-6.

Published on npm: npm install polygon-offset-exact

Parametric product configurator

A 3D configurator with no model file. Most load a finished mesh and swap its materials, which demos well and says nothing about whether the thing can be made. Here the parameters are dimensions in millimetres: changing the gusset depth doesn't select a variant, it rebuilds the geometry — and moves the gusset on the 2D cutting layout by the same millimetres.

The cutting line is a polygon offset of the sewing line, and the page measures every point of it back to the sewing line: 0.00 mm. Three.js vendored rather than pulled from a CDN, so it has no external runtime dependency.

Nesting playground

Nesting looks like bin packing until the material starts making rules: the roll is a fixed width and unbounded length, pieces must run along the grain so most get two rotations rather than four, and everything needs a gap. Each constraint is a control here rather than a paragraph, so its cost is visible in metres — dropping the grain lock on the default four-garment marker takes it from 2.54 m to 2.44 m.

Bottom-left fill with a skyline accelerator on a raster grid, run across four orderings because greedy placement is order-dependent and none wins on every mix. The grid resolution is a control too, so you can watch efficiency fail to improve monotonically as it gets finer — which is the greedy showing through, and the reason production nesters run search on top.

Geometry for design tools

Four pieces of 2D geometry every CAD-like editor needs — orientation, point-in-polygon, distance, tolerance — each written around its failure mode rather than its formula. Every figure is draggable and every number is computed live. The formulas take thirty seconds to look up; knowing which assumption breaks and what it costs downstream is the part that isn't written down anywhere.

looks-glassy

Refractive glass components for React and Angular, built on physics-based SVG displacement maps rather than a backdrop blur pretending to be glass. A framework-agnostic core generates the displacement map and builds the SVG filter; thin React and Angular packages wrap it.

Refractive index, glass thickness, bezel width, surface profile, specular angle and saturation are all parameters, so the effect is tuned like an optical material instead of dialled in by eye — with graceful fallbacks where the filter chain isn't supported.

Prompt Ops

I build the automation systems at Prompt Ops, a Brazilian studio that wires AI into the tools a business already runs — CRM, customer support, internal operations — rather than adding one more app beside them.

The work is integration and tuning: connecting existing systems, shaping prompts and workflows for accuracy and cost, and putting in the guardrails that let an automation be left running unattended instead of watched.

LLM outreach pipeline

A six-stage pipeline where the interesting part is everything around the model call. Daily and per-record spend ceilings are checked before the request rather than after. A circuit breaker opens on provider failure and the workers re-queue with a delay instead of burning their retry budget — and it only counts API errors, so a bug in my own code raises through instead of hiding behind an open circuit.

Sources are pluggable behind one interface, deduplication leans on a unique index rather than check-then-insert so concurrent workers can't both win, and nothing reaches a recipient without passing a human approval gate. Each stage sits behind its own feature flag, switchable without a deploy.

replayable-ingest

An ingestion pipeline built around the four things that break data platforms without ever raising an exception: the same export shipped twice, one row carrying "N/A" in an integer column, last week's re-export arriving after this week, and the same key appearing twice in one file. A plain upsert gets all four wrong quietly — a double count finance finds later, a cast to zero nobody can tell from a real zero, a table walking backwards, and a guess presented as a fact.

Batches are identified by the hash of their bytes rather than their name, deduplication rests on a unique index rather than on the loader being correct, a record older than what we hold never overwrites it, and a batch either lands whole or goes to quarantine with every failing line, field and reason named. Each guarantee has a test named after it, so a red test says which promise stopped holding — including one that loads the same files in reverse order and asserts the table converges on the same fingerprint.