Alistair Pullen

Alistair Pullen is a co-founder of Cosine, where he leads the ML team, research direction, synthetic-data curation and model fine-tuning. His work on Genie, its software engineering agent, centers on a distinction between generating code and learning the process of solving engineering problems. He was identified as Cosine’s CEO in his August 2024 interview and in his authored article dated April 16, 2026.[3][1][4]ⁿ¹

In his August 22, 2024 interview, Pullen called a pull request “a super lossy thing”: its final diffs preserve changes but lose much of the investigation and reasoning that produced them. His team’s training pipeline sought to reconstruct how and why engineers solved problems, rather than teach a model only to imitate completed code changes.[1]

The arrival of a 128,000-token context model prompted Pullen to ask OpenAI for fine-tuning access. He wanted enough information in context to trace generated output back to material the model had actually seen. The team prepared its data pipeline while waiting, then gained experimental GPT-4 Turbo fine-tuning access. Pullen described the first trained version as “a disaster,” exposing dataset biases and problems with producing Git diffs that applied correctly.[1]

One early experiment made the consequences of data selection concrete. Pullen recalled that a minimally cleaned dataset produced a model adept at updating READMEs and writing comments, but also prone to argumentative review responses. The team filtered those examples and worked to align the training data with the engineering behavior it wanted.[1]

To teach iterative repair, the team used two kinds of failure. It deliberately introduced runtime errors by changing abstract syntax trees, creating out-of-bounds accesses or referring to nonexistent variables. It also let a previously trained model attempt a task, compared its incorrect result with a known correct codebase state, and used the resulting corrective diff in the next training batch. Pullen described this as a way to teach the model to recover from mistakes rather than depend on getting the answer right in one attempt.[1]

Pullen described the retrieval problem as “searching for what code does rather than what code is.” An earlier approach translated an English query into a hypothetical code snippet, then embedded that snippet to search for similar code. That became part of an engine combining semantic search, keyword search and language-server tools. Genie subsequently took a different approach: the team trained it through self-play to choose how to investigate a codebase, rather than route its requests through the earlier engine.[1]

Genie’s learned retrieval process inspected directories and candidate files, then followed definitions and references using running language servers. It retained search and graph tools, but learned when to use them. Pullen reported that emphasizing language-server navigation raised internal retrieval performance from roughly 54% into the 60s. He treated missing information as a major opportunity for improvement: an agent must find the relevant code before it can reproduce and solve an issue.[1]

The team trained Genie to emit patches for token efficiency. Pullen also analyzed token probabilities and issue-solving performance across its context window. He reported that, for the system and SWE-bench tasks he examined, solving probability fell to about 0.5 at 60,000 tokens. That observation informed short-term optimizations to keep work within a more successful range; it was a measurement of that system, rather than a general limit on usable context.[1]

Pullen worked with OpenAI’s fine-tuning team, including John Allard, on questions involving larger LoRA adapters, learning-rate schedules and limits encountered as training-data volume increased. Other implementation work remained with his colleagues: he explicitly credited the rest of the Cosine team with the CI integration. The design he described invoked a repository’s existing CI and returned its outputs to the model, rather than asking the model to set up and run the repository itself.[1]

In August 2024, Pullen explained the company’s decision to withhold Genie’s reasoning trajectories while leaving its patches public. He believed the trajectories mirrored its training approach closely enough that competitors could use them to distill a competing model. In his account, that choice prevented recognition on the official SWE-bench leaderboard when trajectories were requested. He nevertheless reported resolving 219 of the 500 SWE-bench Verified tasks, or 43.8%; this was a historical result reported by Pullen.[1]

Pullen’s August 2024 view of delegation was that “the model should do the coding and you should be in charge of guiding the model.” He also wanted users to be able to edit plans and correct patches while the agent continued working. In his article “Cosine: The Coding Agent for Engineers with Taste”, dated April 16, 2026, he defined developer taste as judgment about trustworthiness and maintainability. He argued that agents should expose their work for inspection and intervention, respect existing architecture and conventions, and make long-term code quality a priority.[1][4]

The article explained the team’s choice of one underlying agent system across terminal, desktop and cloud interfaces: developers should be able to change where they work without adopting a different operating model. Pullen also advocated choosing models to fit a task’s performance, cost and governance constraints. As an example of oversight during parallel work, he described agents producing independent diffs in isolated worktrees, with review required before merging.[4]

Before Genie, Pullen worked in mobile development and startup engineering. In his 2024 recollection, he and his future co-founder Sam built a small consultancy during university and continued taking projects afterward. Pullen then joined Fancy, a fast grocery delivery startup, and Sam later joined him. He recalled that the two built its mobile apps, backends, stock-management system and driver-routing algorithms. Looking back, he said they would have bought more components off the shelf. After GoPuff acquired Fancy, they completed an integration period and left around 2022 to pursue their own projects.[1]

Their mobile-development background supplied the first AI experiment: Pullen and Sam chained prompts through small context windows to generate an app from an idea, progressively specifying its stack and components before asking Codex to write the code. Pullen recalled occasional working results from an otherwise fragile system. They showed it to Yang, whom Pullen had met through Fancy and who became their co-founder and COO.[1]

Pullen recalled that a Y Combinator interview rejected the app-generation idea but accepted a proposed direction helping developers work on existing codebases, with two weeks to produce an MVP. The team initially built code-retrieval tooling while model capability and context limits constrained its broader ambition. Trying to keep code local led to a semantic-search implementation inside VS Code whose HNSW index consumed substantial memory. This earlier work supplies context for Genie’s later retrieval decisions: the team had already spent months confronting the difficulty of locating useful information in large codebases.[1][3]