Blog

What we learned shipping.

Engineering writing from the people who did the work: production AI, frame budgets, live mainframe migrations, offline-first architecture and how we hire. Click a post to read it.

The most valuable thing an AI agent does is refuse

AI · 18 June 2026 · 8 min read · Priya Raman, Principal AI Engineer

An agent that answers every question is not confident. It is uncalibrated. The distinction matters enormously once real money or real patients are downstream of the output, and it is almost never what teams measure.

In the Helios Orbit deployment, the behaviour that convinced a compliance team wasn't the retrieval quality or the reasoning depth. It was the refusal path: when span-level confidence dropped below a threshold, the system escalated to a human instead of producing a fluent, plausible, wrong number.

The uncomfortable truth is that refusal is expensive to build. You need calibrated confidence, provenance for every claim, and a definition of ‘unsupported’ that survives an adversarial reading. None of it demos well. All of it is what keeps the system in production eighteen months later.

Our rule of thumb: if you cannot describe, in a sentence, the conditions under which your agent will decline to act — you do not have an agent. You have a very expensive autocomplete with a tool-calling API.

Sixteen milliseconds: how we cut a frame budget in half

Gaming · 30 May 2026 · 11 min read · Dana Okonjo, Technical Director, Games

When we inherited Ashfall it ran at 40 milliseconds a frame on a PlayStation 5 devkit. The launch window required 16. Eleven months, three platforms, and a prototype that everybody already loved — which is its own kind of constraint, because the fun was in the parts that were slowest.

The first thing we did was stop calling it optimisation. Optimisation is what you do at the end, when there's no time left. We made the frame budget an acceptance criterion on every ticket: if your feature costs 0.4ms, you say so, and somebody has to give 0.4ms back.

The wins were unglamorous. GPU instancing on the crowd system returned 6ms. Nanite discipline — actually auditing which meshes deserved it — returned another 4. The animation graph, which nobody suspected, was costing 3ms on a thread we thought was idle.

None of that is a clever trick. It's a culture where the cost of a feature is visible at the moment somebody proposes it, rather than discovered in a panic six weeks before certification.

We migrated a 22-year-old mainframe with the lights on

Technology · 12 May 2026 · 9 min read · Marcus Feld, Principal Architect

Vantatek's system of record was COBOL, undocumented, and processing four hundred thousand orders a day. Two previous rewrite attempts had been abandoned mid-flight — which is the normal outcome, because a rewrite asks an organisation to hold its breath for two years.

We didn't rewrite. We put an event gateway in front of the mainframe and moved one bounded context at a time behind it. Every context dual-wrote to both systems until the new path had proven itself against real production traffic — not staging traffic, which is a fiction.

The rollback flag stayed live for the entire fourteen months. We used it four times. That number is the whole argument: a migration where rollback is theoretical is a migration where nobody is willing to move fast.

Zero downtime, zero lost orders, 3.4× throughput. The unsexy conclusion is that big-bang rewrites don't fail on technology. They fail because they ask for a year of faith, and organisations don't have a year of faith to give.

Offline is not an edge case

Apps · 22 April 2026 · 7 min read · Elena Cortez, Product Lead, Mobile

The clinicians using Orbital work hours from a cell tower. The previous tooling treated that as an exception: it showed a spinner, then an error, then — worst of all — silently dropped the record on the floor.

In a patient chart, silent data loss is the most dangerous failure mode there is. It is worse than a crash, because a crash tells you something happened.

So we inverted the default. The local store is the source of truth. Sync is an optimisation that happens when the network shows up. A CRDT handles the merge, deterministically, without asking a clinician to resolve a conflict between two versions of a vital sign — which is not a question a human should ever be asked at the point of care.

Eleven months, six thousand clinicians, not one record lost. The architecture is more complex than a REST client. The user experience is dramatically simpler: the app just works, and nobody thinks about sync at all. That trade is almost always worth making.

We stopped interviewing for algorithms and started interviewing for taste

Company · 28 March 2026 · 6 min read · Ilya Novak, Co-founder

We used to run a whiteboard round. It measured, as far as we can tell, how recently a candidate had practised whiteboard rounds — a quantity uncorrelated with anything we care about.

Now we ask for something they built and a critique of it. The build tells us what they can do. The critique tells us what they can see, and what they can see is the ceiling on what they will eventually build.

The best signal in the entire process turned out to be a single moment: the candidate who, unprompted, points at the part of their own work they're not happy with and explains precisely why. That person will raise the standard of everyone around them.

Our offer-acceptance rate went up. Our first-year attrition went to nearly zero. And the interview stopped being an ordeal and started being the most interesting conversation in the candidate's week — which, it turns out, is also good recruiting.

In XR, a dropped frame isn't a bug report — it's motion sickness

Technology · 4 March 2026 · 8 min read · Sofia Marchetti, XR Director

On a flat screen, a frame drop is a stutter. In a headset strapped to somebody's face, it is a vestibular event. The user does not file a ticket. They take the headset off, and a meaningful fraction of them never put it back on.

This makes 90fps a floor rather than a goal, and a floor changes how you design. It means the frame budget is fixed before the art direction is decided. It means you choose the interaction model — teleport, hand-tracked, seated — partly on comfort grounds, not purely on immersion grounds.

For Northwind Forge we simulated the robotic cell against real controller firmware, which meant the physics had to be accurate and fast. The temptation was to trade frames for fidelity. We didn't, and the rollout to eleven plants is the reason: technicians used it daily, because it never made them feel ill.

Every immersive project has this fork in it. The teams that take comfort seriously ship tools people use. The teams that chase fidelity ship demos people admire once.