SignalPilot, a US-based, Bangladeshi diaspora-led startup, became the first team to break through a performance threshold that OpenAI, Anthropic, Google DeepMind, and JetBrains had been unable to cross for nearly a year — the 50% mark on the hardest benchmark in data engineering AI. 

The open-source AI agent layer hit 51.56% accuracy on the benchmark on April 21, crossing the 50% ceiling for the first time in the test's eleven-month history. JetBrains, which had previously held the top position, responded within ten days — a turnaround that is itself a signal of how seriously the result was taken. SignalPilot proved the wall could be broken. It was the first team in the world to do it.

The benchmark tests AI agents against broken enterprise dbt pipelines — a scenario that closely mirrors what data teams actually encounter in production. Its difficulty stems from a core problem that haunts AI agents deployed on real databases: they guess at schemas, misread relationships, and in some cases attempt to execute destructive operations on live infrastructure. "Please don't drop tables" in a system prompt, as Tarik Adnan Moon, SignalPilot's founder and CEO, has put it, "is a wish, not a security control."

SignalPilot is built as a Jupyter-native agentic harness for data teams — connecting directly to a data warehouse, dbt lineage, query history, Slack threads, Notion, and Jira to give the AI the kind of institutional knowledge that general-purpose tools like ChatGPT or IDE copilots cannot access. Rather than single-shot code generation, it runs long-running agent loops that plan, execute, and iterate with an analyst in the approval chain. It retains memory across sessions — tracking past hypotheses, validated assumptions, and known data quirks — and can be taught a team's custom business logic, coding standards, and domain-specific analysis patterns.

The governance layer sits underneath all of it. Where most AI tools rely on prompt instructions to prevent dangerous operations, SignalPilot enforces rules at the architecture level — what the company describes as governing agents with physics instead of prompts. Destructive database operations are blocked at the parser level before they can reach the database. A 7-Check Verifier Subagent validates row counts, fan-out metrics, and schemas before any pull request opens. The product runs locally, retains zero data, and operates in read-only mode by default.

The underlying meta-harness is called AutoFyn — a system that runs Claude in sandboxed loops and iteratively tunes its own agent architecture until it converges. Tarik has described the approach as inspired by the Banach fixed-point theorem from mathematics: repeatedly apply a transformation until the system reaches a stable state. AutoFyn also autonomously discovered 26 zero-day vulnerabilities in major open-source projects during development. The product is 100% open-source and deployable locally in 60 seconds. Over 100 developers had used it within its first week.

SignalPilot is led by Bangladeshi founders operating out of the United States — part of a growing cohort of diaspora builders competing at the frontier of global tech. For eleven months, no team at any major AI lab crossed 50% on this benchmark. A five-person Bangladesh-led startup was the first to do it.

Tarik Adnan Moon trained as a mathematician at Harvard and worked as a quantitative analyst at Goldman Sachs before turning to AI agent infrastructure. Fahim Aziz, co-founder, is a two-time Y Combinator founder who previously built Nala (YC W19) and Backpack (YC S14). The remaining three — Adib Hasan, Daniel Schaffield, and Luiz Fernando — handle engineering. The team operates from the United States and is targeting a market where roughly 80,000 data teams globally depend on dbt, the open-source data transformation tool that SignalPilot is built around.

‍