Affaan Mustafa

Affaan Mustafa co-leads compute trading at Jump Trading alongside Alejandro Soumah Fernandez and created ECC, an open-source toolkit for giving coding agents reusable skills, memory, planning and security workflows. He previously co-founded Itô Markets, a compute-trading startup acquired by Jump. His work connects three interests that developed together: building computers, trading markets and using software agents to run more of a business. His career account traces that path from custom PCs to Jump’s dedicated compute desk.

From custom computers to trading agents

Mustafa began building PCs at 11 and trading CS:GO skins, then moved into custom computers, electronics commerce and cryptocurrency. While running his own business, he earned two bachelor’s degrees at the University of California, San Diego, in Mathematics–Computer Science and Business Economics. He later entered graduate study in Applied and Computational Mathematics at the University of Washington, leaving at 21 to build full-time. These early experiences joined technical work with the practical demands of buying, selling and operating a company. (Education and early career)

His company, DCUBE, developed from ecommerce and custom computers into language-model-enabled customer relationship management and administrative automation. Running a business while studying gave him a concrete reason to experiment with AI: he wanted software to handle consulting, secretarial and management work he could not delegate. The venture did not succeed commercially. The ambition survived: expanding how much one person could build and operate with the right tools. (Background)

From November 2024 through January 2025, Mustafa built autonomous trading software on Solana and streamed its development publicly. He also contributed to elizaOS, the open-source autonomous-agent framework, and co-founded AYO Protocol, an AI-powered yield platform combining agent-driven workflows with on-chain execution. From September 2025 through February 2026, he served as head of product and founding engineer at PMX/Polycule, helping build its core prediction-market trading infrastructure. Across these projects, agents moved beyond answering questions into gathering market data and supporting financial execution. (Project history)

He also applied agents to customer research. In 2025, Mustafa and teammate David Rodriguez won first prize at a Forum Ventures × Anthropic hackathon. Their entry combined synthetic customer-discovery testing with an automated email feedback system, giving early-stage founders ways to investigate an idea before building it. Rodriguez’s account calls the entry PMF Probe; Mustafa’s portfolio presents his customer-research product and hackathon win under the name Zenith.

Making a coding workflow reusable

ECC grew out of the system Mustafa used to build his own products. It packages specialized agent instructions, reusable skills, automated hooks and ways to carry project knowledge between sessions. Rather than requiring developers to adopt a new coding environment, it supplies working practices for tools they already use, including Claude Code, Codex, Cursor, OpenCode and Kimi Code. Mustafa’s career account reports that the project reached 100,000 GitHub stars in about two months and 200,000 in less than five months.

The toolkit includes several ways to make an agent’s work easier to inspect and continue. Plan Canvas lets developers annotate a plan in the browser, send feedback to an agent and review its revision. Unified memory saves handoffs and makes project context searchable across coding tools. A separate GitHub App uses repository history to prepare guidance that a team can review and keep with its code. These features address different points where a coding session can lose direction: an unclear plan, missing context or lessons that never make it into the next task. (ECC tools and workflows)

Mustafa’s longform guide to ECC explains the working habits behind the toolkit:

• Preserve what the next session needs. Record approaches that worked, attempts that failed and unfinished work, then review that record before starting a fresh session. Recurring debugging lessons can become reusable skills, so the developer does not have to explain the same problem repeatedly.
• Choose tools with context costs in mind. When a service already has a capable command-line interface, a small skill or command can sometimes replace a permanently loaded Model Context Protocol integration. His example wraps GitHub’s pull-request command instead of keeping a larger tool catalog available throughout the session. The recommendation is selective: use the smaller interface when it supplies the functionality the task actually needs.
• Parallelize separable work. Add agent sessions because there are distinct tasks to do, rather than aiming for an arbitrary terminal count. Define responsibilities clearly, minimize overlapping changes and use separate Git worktrees when parallel agents need to edit related code.

Mustafa also created AgentShield, a security scanner for agent configurations, tool permissions, hooks and MCP connections. Initially built at the February 2026 Cerebral Valley × Anthropic Claude Code hackathon, it checks for exposed credentials, unrestricted command permissions and injection risks. Developers can run it through a command-line tool, a GitHub Action or a GitHub App integration. Its findings help them inspect what an agent can reach and decide what to change; a scan is a review aid, not a guarantee that an agent or repository is safe.

From GPU procurement to compute trading

With Soumah Fernandez, Mustafa co-founded Itô Markets, where he served as chief executive. The company developed an agent-assisted over-the-counter desk for sourcing, contracting and hedging GPU capacity. A buyer could specify a workload, GPU type, quantity, region and term; the desk would obtain comparable supplier quotes, generate a capacity contract from the selected quote and bring the resulting position into its trading workflow.

The financial problem extended beyond finding available GPUs. Buyers wanted predictable costs without committing too much capital before demand materialized. Suppliers needed dependable future revenue to support infrastructure financing. Itô’s proposed model put a desk between those needs: sell the client a fixed GPU-hour rate, source capacity and arrange hedges against price increases. If the desk’s supply costs rose, the hedge was intended to offset that movement. The agent-assisted procurement workflow and the management of financial exposure were parts of the same service. (Itô’s launch description)

Following Jump’s acquisition of Itô, Mustafa and Soumah Fernandez became co-leads of its dedicated compute desk. The move brought Mustafa’s earlier hardware business and agent-building work into the same market: buying, financing and supplying the infrastructure on which AI software runs. (Current role and acquisition)