Engineering blog
Platform updates, engineering deep-dives, and honest numbers.
Starting today, Nyquest has one line item: half of what compression saved you on the call. No subscription, no markup on the provider's rate, no fee on a call we did not improve. Here is the arithmetic, a worked example, and the case where we earn exactly nothing.
2026-08-12Stemming and lemmatization are standard practice in search. We measured them as prompt compression: 45,366 inflected words said a lemma is cheaper 39% of the time. Twelve real prompts said zero — on six of seven tokenizers, exactly zero, every rewrite one token in and one token out. Here is why both numbers are right, and what we changed about how we evaluate a rule.
2026-08-10A 50-byte English sentence costs 8 tokens. A 17-byte MAC address costs 16. We doubled the tokenizer vocabulary from 100,277 entries to 200,019 and the MAC address, the IPv6 address, the UUID and the interface name did not get one token shorter — the URL got longer. That is not a vocabulary problem, it is a merge-selection problem, and it is what our new tokenizer lab exists to test.
2026-08-05Our compression engine rewrote "should not" to "shouldn't". On Claude's current tokenizer that costs three tokens, and Claude is a third of the prompt tokens we send. We measured 150,895 words against seven tokenizers, audited all 203 literal rules, found 38 that lose somewhere — including one we had already published and were still shipping — and took Claude savings from 3.0% to 9.8%.
2026-08-02The obvious prompt optimisation is to swap long words for short symbols. We measured 32 of them against three tokenizers: "and" to "&" saves exactly zero, "with" to "w/" costs you one, and the same three records cost 25 tokens as JSON and 3 as YAML. Plus the converter of ours that had been dead for three weeks — and why a character-based metric could never have told us.
2026-08-01Providers retire models. If your integration pinned one, every request fails — and because a caller-side 4xx looks like your bug, nobody investigates. One of our own keys did exactly that for two months. Nyquest now routes retired models to their successor, always reports the swap on a response header, and watches for the integrations still asking for something that no longer exists.
2026-07-30Live web grounding is a feature until it overrules the thing you actually trust. Retrieved search results were being injected ahead of the developer's own instructions — and telling the model to prefer themselves. Nyquest's grounding now sits below your instructions instead of above them, callers can switch it off or force it on, and the headers reporting it finally reach the stateless callers who needed them most.
2026-07-21Nyquest's memory now behaves like a belief state, not a diary: corrections replace what they correct, single-valued facts update mechanically, a nightly pass merges duplicates and retires stale notes, and every fact carries a confidence score your conversations train. All of it visible and controllable in Settings — plus the wiring bug we found and fixed along the way.
2026-07-20Nyquest is twelve complete design languages, not one look. We rebuilt them onto a single shared foundation — faster, consistent, real on mobile — then used it to make Prism: Nyquest rendered in 2036.
2026-07-18Every chat input on Nyquest now shows a live token count and real cost estimate, and can optimize your prompt before you send it - with an integrity validator that refuses to lose your URLs, IPs, numbers, or 'do not' instructions, and repairs the result automatically when a rewrite drops one. The validator caught five of our own bugs before any user did.
2026-07-16The best security news is no news. A quick note on a routine hardening pass across the platform — tighter access, a smaller network surface, defense-in-depth at the API, and automated daily monitoring.
2026-07-15On a phone, the chat composer rendered off-screen and the conversation drawer was unreachable. The postmortem: one in-flow banner above eleven 100vh themes, one CSS sizing rule wearing three costumes, and a hamburger a stylesheet loading order had silently deleted.
2026-07-12Condensing conversation history used to add about a second to every turn of a long chat - on exactly the conversations where compression saves the most. Measured in production, it is now ~15ms on warmed turns, roughly 70x faster, with token savings unchanged.
2026-07-11Seven engine releases in seven days, all driven by measurement: grounded traffic covered, a savings curve that compounds with conversation length (15% at 7 messages to 75%), the loop's first self-driven fix, and provider-side caching on top.
2026-07-10We built a shadow-eval loop that judges our own compression against uncompressed baselines. It found where compression hurts answers on day one - code and numbers - and the fix makes them untouchable.
2026-07-09Nyquest won Best Tech at the FAU x The AI Collective Demo Night. The three features that carried the four minutes: the Splicer (multi-model consensus), the compression engine, and the AutoRouter.
2026-07-01Per-domain quality scorecards, recency-weighted grading, and bandit-math exploration — the routing brain now learns from every judged answer.