Cosine

Cosine develops AI models and coding agents for software engineering, including work in defence, national security and regulated industries. Its Genie agent used a fine-tuned GPT-4o model trained on examples of engineers at work, with an emphasis on the process of finding and fixing problems rather than generating finished code alone. In June 2026, Cosine announced an industry coalition to co-design Lumen Sovereign, a model targeted for late-2026 delivery and deployment within customers’ own infrastructure.[7][1][6]

In the August 2024 interview, co-founder and CEO Alistair Pullen explained that a pull request preserves final changes while losing much of the investigation and reasoning that produced them. Cosine’s training pipeline sought to reconstruct that process. Its retrospective history also describes building a custom keylogger to capture its own engineers navigating and editing code. OpenAI’s account, reproduced in the interview, identified Genie’s launch-era model as a fine-tuned GPT-4o; OpenAI’s announcement described training it to produce formats such as patches that could be committed to codebases.[1][5][2]

Cosine also trained for failed attempts and subsequent corrections. Pullen described introducing synthetic runtime errors by changing abstract syntax trees, creating invalid indexes or referring to nonexistent variables. In successive training batches, an earlier model attempted a task; when it failed, Cosine compared the resulting codebase with a known correct state and trained on the corrective diff. This supplied examples of iterative repair rather than only one-shot answers.[1]

Genie’s workflow comprised finding files, planning, writing code and running tests. Pullen said it was trained to write diffs for token efficiency. The intended shipping product would call existing GitHub CI pipelines and read their outputs, leaving repository setup outside the model’s responsibilities. He also described an internal variant called “the John,” created by running Cosine’s data pipeline over its own repository and further fine-tuning Genie on that codebase.[1]

In August 2024, Pullen reported 219 resolved instances out of 500, or 43.8%, on SWE-bench Verified, calculated by matching the benchmark’s instance IDs against an earlier run. He attributed Genie’s absence from the official SWE-bench leaderboard at that time to Cosine’s decision to withhold model trajectories: the intermediate outputs could reveal its proprietary training approach and help others distil it. Cosine left the patches public. The result was self-reported, and the disclosure decision traded leaderboard inclusion for protection of the training method.[1]

Genie followed an earlier period of app-generation experiments and codebase-search development. That earlier work supplies context for both the company’s focus on existing repositories and the retrieval approach it later replaced.[1]

Cosine was founded in 2022 and joined Y Combinator’s Winter 2023 batch. An August 2024 report identified the company as Buildt AI Inc. and named three co-founders: Alistair Pullen, Yang Li and Sam Stenner, then serving respectively as chief executive, chief operating officer and chief information officer. Pullen recalled that he and Sam had built mobile apps, backends, stock-management software and driver-routing systems at grocery-delivery startup Fancy. After its acquisition by GoPuff and an integration period, they left around 2022 to pursue their own projects.[3][8][1]

Pullen and Sam initially chained prompts around small context windows to generate mobile apps from an idea, progressively specifying the stack and components before asking Codex to write code. Pullen recalled that the prototype occasionally produced a running backend and database, and that they showed it to Yang, whom he had met through Fancy. At their YC interview, he said, criticism of the app-generation idea prompted a move toward automating development in existing codebases. Their first launched product became a codebase-retrieval tool.[1]

The retrieval product aimed to run locally and minimise sending code to the cloud, but Pullen recalled memory problems when indexing large repositories inside VS Code. To connect English queries with code, Cosine trained a model to generate a hypothetical answering code snippet, then embedded that snippet for similarity search. Its broader retrieval engine combined semantic search, keywords and language-server tools. Genie did not use that earlier engine: Pullen said it learned navigation through self-play, inspecting directories and candidate files, then following definitions and references through running language servers. Search tools remained available within that learned process.[1]

Genie’s development passed through GPT-3.5 experiments with a 16k context window, GPT-4 fine-tuning with an 8k window, and experimental access to GPT-4 Turbo fine-tuning. Pullen said Cosine had prepared its data pipeline before receiving Turbo access; early training exposed dataset biases and difficulties producing applicable Git diffs. These stages preceded the GPT-4o launch-era model. He described working with OpenAI’s fine-tuning team, including John Allard, on adapter sizing and learning-rate schedules while training at billions-of-tokens scale.[1]

Cosine’s retrospective history attributes its move into air-gapped deployment to demand from an unnamed major investment bank. The company says it fine-tuned an open-source model to operate entirely on premises, then built Kubernetes infrastructure and specialised post-training pipelines for organisations cut off from the public cloud. This extended its work beyond hosted coding agents into customer-controlled environments.[5]

On April 16, 2026, the UK government named Cosine among six startups receiving access to the AI Research Resource supercomputer network through its Sovereign AI initiative. This was compute support; the separately announced first equity investment went to Callosum. The initiative’s £500 million budget was a national programme amount, not financing awarded to Cosine.[7]

On June 8, 2026, Pullen announced a coalition to co-design Lumen Sovereign. Named participants included BAE Systems, Babcock International Group, BT, Era4, HSBC, Leonardo, Lloyds Banking Group, LSEG, NatWest Group, PwC, Telefónica Tech UK&I, Thales UK and The Alan Turing Institute. Under a memorandum of understanding, the coalition would help define use cases, security protocols and governance standards. Its participation represented collaboration on requirements rather than announced purchases or investments.[6]

Cosine’s announced plan was to expand a known-good sparse mixture-of-experts architecture with a 256k context window, using continual pre-training and expert expansion within a 500,000 GH200 GPU-hour budget on Isambard-AI. Approximately one-third of training compute would address complex software engineering; the remainder would cover broader uses such as cybersecurity adversarial testing, KYC and AML investigations, and clinical-trial coordination. Delivery was targeted for late 2026, with eventual deployment inside customer infrastructure, including fully air-gapped environments. These were architecture, training and deployment plans.[6]

On August 12, 2024, SiliconANGLE reported that Cosine said it had closed a $2.5 million seed round, led by Uphonest and SOMA Capital, with participation from Lakestar, Focal and others.[8]

Cosine’s newsroom summary of outside coverage, dated August 11, 2026, described the London company as having about 30 employees and $15 million raised. Those figures describe the approximate headcount and cumulative financing reported in that dated summary.[10]