With $3 Million, Sqwish Labs Is Training AI to Choose, Not Chat

The Cambridge startup is betting that real-time, specialized decision engines will be the infrastructure layer inside AI products.

About Sqwish Labs

Published

The most expensive part of an AI product isn't the model. It's the electricity and time spent waiting for it to think out loud. Sqwish Labs, a Cambridge startup with eight people, is building models that don't think at all. They just choose, in about the time it takes to blink.

The bet on structured decisions

Founded in 2024 by Ushnish Sengupta and Federica Freddi, Sqwish Labs trains what it calls 'small decision models' [Sqwish Labs, 2026]. Its flagship product, the kettei family, is designed to make a specific, structured choice from a set of predefined options,like routing a customer support query, flagging a contract clause, or detecting a prompt injection,and return a confidence score. The key is that it does this without generating a single token of explanatory text. According to the company, the median end-to-end latency for a kettei decision is 160 milliseconds, and the core inference cost is $0.02 per million tokens [Sqwish Labs, 2026]. For comparison, sending a query to a frontier model and waiting for a reasoned response can take seconds and cost dollars. Sqwish's bet is that for countless production workflows, you don't need a verbose explanation. You just need a fast, cheap, and accurate answer.

The OneHot wedge

Sqwish isn't just selling a set of pre-trained models. Its wedge is a platform called OneHot that lets customers create and customize these decision engines using their own examples and, crucially, their ongoing production data [PERPLEXITY SONAR PRO BRIEF, 2026]. A developer can upload a dataset, and the platform reportedly returns a bespoke model in minutes. More importantly, the model continues to learn. As the AI product runs, new outcomes and user feedback can be fed back into the training loop, allowing the decision model to improve on the specific task it's been assigned [Sqwish Labs, 2026]. This turns a static component into a learning system, a feature the company calls 'outcome optimization infrastructure' [Sqwish Labs, 2026].

The founding team brings a focused technical pedigree to the problem. Sengupta, the CEO, holds a PhD from Cambridge and worked on generative AI and foundation models at MediaTek Research [PERPLEXITY SONAR PRO BRIEF, 2026]. Freddi, the CTO, also a Cambridge alum, has a background in model efficiency techniques like pruning and meta-learning from roles at ARM and MediaTek [PERPLEXITY SONAR PRO BRIEF, 2026]. Their research-heavy approach is reflected in the product's architecture, which offers four model sizes,from 'fat' for the hardest calls down to 'dot',letting customers use the smallest model that gets the job right [Sqwish Labs, 2026].

Benchmarks versus giants

The company's November 2026 pre-seed round of $3 million, led by Cambridge Enterprise Ventures, was announced alongside a set of bold performance claims [Cambridge Enterprise / Cambridge Founders, November 2026]. In a published benchmark, Sqwish compared its kettei models against industry leaders on specific tasks:

Task Sqwish kettei Model Competitor Model Result for Sqwish
Legal Contract Review 94.5% accuracy, $0.03, 220 ms Claude Opus 5.5: 93.6% accuracy, $6, 2.7 s More accurate, 200x cheaper, 12x faster [Sqwish Labs, 2026]
Support Escalation 82.5% accuracy, $0.008, 163 ms GPT-6 Luna: 72.7% accuracy, $0.032, 2.8 s More accurate, 4x cheaper, 17x faster [Sqwish Labs, 2026]
Prompt Safety 61% safer, same price Llama Prompt Guard 2: 5% safer 12x safer [Sqwish Labs, 2026]

These figures, while sourced from the company, frame the ambition clearly: to outperform general-purpose giants on narrow decision tasks where speed and cost are paramount.

The incumbent to beat

The obvious counter-bet is that developers will simply continue to call a single, powerful API for everything, accepting the latency and cost as the price of simplicity. The infrastructure of prompt engineering, function calling, and output parsing around these large models is becoming deeply entrenched. For Sqwish to succeed, developers must be convinced that the operational savings and performance gains are worth integrating a separate, specialized decision layer. The risk is that this adds complexity to a stack that many are trying to simplify.

Furthermore, the company's public traction is currently measured in benchmarks, not named enterprise deployments [PERPLEXITY SONAR PRO BRIEF, 2026]. The $3 million will be used to expand the OneHot platform for enterprise customers and hire commercial and developer relations staff [Cambridge Enterprise / Cambridge Founders, November 2026]. Proving that the impressive lab numbers translate into scaled, reliable production use is the next hurdle.

Do the economics work? Take the support escalation example. If a product handles one million decisions a month, using GPT-6 Luna would cost about $32. A kettei model doing the same job would cost $8. Over a year, that's nearly $300 in direct cost savings per million monthly decisions, before you factor in the latency benefits for user experience. The incumbent to beat isn't just a model provider; it's the inertia of a good-enough, familiar workflow. Sqwish Labs must prove that its specialized, learning decision engines are not just faster and cheaper, but reliably better enough to change how AI products are built.

Sources

  1. [Sqwish Labs, 2026] Sqwish Labs: Building System One for AI | https://sqwish.ai/
  2. [PERPLEXITY SONAR PRO BRIEF, 2026] PERPLEXITY SONAR PRO BRIEF
  3. [Cambridge Enterprise / Cambridge Founders, November 2026] Eight-person Sqwish Labs raises $3M as its decision models outrank those from OpenAI, Microsoft and TypeSafe | https://founders.cam.ac.uk/eight-person-sqwish-labs-raises-3m-as-its-decision-models-outrank-those-from-openai-microsoft-and-typesafe/
  4. [TokenPost, November 2026] Sqwish Raises $3 Million to Build Its Tunable Decision Model | https://www.tokenpost.com/news/business/59736

Read on Startuply.vc