⋯ 138 unchanged lines
<div class="hero">
<h1>The best AI products aren't the smartest. They're the ones that change safely, every week.</h1>
<p class="sub">That is what we build: the evals, workflows, and infrastructure that let a team ship fast without shipping regressions. Four proofs from this quarter.</p>
<p class="sub">That is what we build: the models, workflows, and infrastructure that let a team ship fast without shipping regressions. Measured outcomes from this quarter's client work, below.</p>
</div>
<section class="sec">
<h2>Rebuilding without stopping the business</h2>
<div class="figures">
<div class="fig"><b>2 weeks</b><span>to cut over, Java → Python</span></div>
<div class="fig"><b>0</b><span>rollbacks after launch</span></div>
<div class="fig"><b>0</b><span>releases frozen during the migration</span></div>
</div>
<p class="lede">A consumer AI video app had outgrown its original engine and was slowing down with every release.</p>
<p>We rebuilt it and cut over in <em class="k">two weeks</em> with zero rollbacks, so the team kept shipping to users the whole time instead of freezing for a quarter. When an AI product outgrows its first codebase, that is the work we do: replace the foundation without stopping the business.</p>
<p class="note">This year the team grew to 20, and shipped to production every week of the quarter.</p>
</section>
<section class="sec">
<h2>More client work</h2>
<h2>Measured outcomes this quarter</h2>
<div class="work">
<div class="item">
<h3>Made prompt changes safe to ship</h3>
<p>For a B2B AI product, the content that most shaped answer quality was edited by hand with no safety net: you shipped an edit and hoped. We put evals in the release pipeline as a gate, so a change that would lower answer quality is caught before a customer sees it. <em class="k">193</em> automated checks now sit between an edit and production, and nothing ships unreviewed — so the team can improve the product every day without gambling on quality.</p>
<h3>+40% click-through, and still climbing in production</h3>
<p>For a lending and collections company, we built a model that sends each borrower's message at their most responsive moment instead of on a fixed schedule. Click-through rose <em class="k">40%</em> over a randomized baseline, and keeps outperforming it live.</p>
</div>
<div class="item">
<h3>Made the product feel 50–60% faster in the moment that sells it</h3>
<p>In a consulting-AI product, the live board is what a consultant shows the client, so its speed is the product's felt quality. We made that board render <em class="k">50–60%</em> faster and stream in as it builds, at cache-read pricing rather than full cost. The moment that matters most, live in front of a client, now feels instant.</p>
<h3>$1–2 cheaper and 30–75 seconds faster on every render</h3>
<p>For a consumer AI music-video product, we removed a paid lip-sync step by carrying the song straight into generation. Each render now costs less, finishes sooner, and has one fewer thing that can fail.</p>
</div>
<div class="item">
<h3>Turned a capability that needed screen-sharing into a link</h3>
<p>Letting a client watch that board update live once meant screen-sharing or emailing a file, and only worked on a single server. We made it stream over the web as live snapshots, across any number of machines, with no account and no install. Friction became a link anyone can open.</p>
<h3>80–90% more assets per creator, ~140 steps down to one</h3>
<p>For a creative marketplace, bulk intake and reusable assets replaced a twenty-form, 120-upload catalogue process with a single submission, and let creators reuse what they make instead of rebuilding it each time.</p>
</div>
</div>
</section>
<section class="sec">
<h2>Worth reading</h2>
<div class="rows">
<a class="link-row" href="https://usething.ai/unravel/prompt-caching-and-how-to-think-about-it">
<span class="lr-main"><span class="lr-title">Prompt caching, and how to think about it</span><span class="lr-desc">Where caching saves real money and latency with large models, and where it quietly does not.</span></span>
<span class="chev">→</span>
</a>
<a class="link-row" href="https://usething.ai/unravel/eval-metrics">
<span class="lr-main"><span class="lr-title">The metrics we judge AI output by</span><span class="lr-desc">How we measure whether a model's output is good, and keep our evals honest as the product changes.</span></span>
<span class="chev">→</span>
</a>
<a class="link-row" href="https://usething.ai/unravel/migration-work">
<span class="lr-main"><span class="lr-title">How we run a big migration without regret</span><span class="lr-desc">The questions we answer before a rewrite, so it ships in weeks instead of stalling for a year.</span></span>
<span class="chev">→</span>
</a>
<div class="item">
<h3>New AI models judged in days, not a rebuild cycle</h3>
<p>We built an evaluation studio across nine models and four vendors, so identity drift and regressions are caught on reproducible evidence before customers ever see them.</p>
</div>
<div class="item">
<h3>A production engine rebuilt in two weeks, with zero downtime</h3>
<p>Java to Python, verified against the real clients and cut over without freezing releases or rolling anything back.</p>
</div>
</div>
<p class="note" style="margin-top:26px;">This year the team grew to 20, and shipped to production every week of the quarter.</p>
</section>
<section class="sec">
⋯ 38 unchanged lines