After 7 Years in Technical Support, His Side-Project Local AI Engine Hit 1.5k Stars
Ash Hart spent seven years in technical support, moved into product, then returned to engineering. In private, he wrote TensorFold, archived it three times, and when he released it, he only dared expect 5 to 10 users. A few days later, GitHub stars passed 1.5k, the community rewrote it in Zig, and got Macs and NVIDIA machines to share memory.
Ash Hart's résumé doesn't look much like the usual AI founder template. In a tweet, Pratik Gandhi laid out a path: grew up in the UK, spent seven years in technical support, joined a friend's food-delivery tech company as a PM for two years, then returned to his first company as a technical lead, where he remains today. More than 15 years of experience in software, DevOps, AWS, networking, and security. Eventually he became fourth-line support—the person they call when nothing else works.
That background fits local AI perfectly. TensorFold is a local inference engine that lets models run on your own machine instead of sending every request to the cloud. It supports Metal and CUDA, corresponding to Apple's and NVIDIA's hardware stacks, respectively. Developers can get directly at the layers involving drivers, memory, networking, and OS versions, while ordinary users don't have to hand over every sentence they type.
He first spent six months writing the engine in private. He archived it three times. Then he released TensorFold anyway, expecting only 5 to 10 users. The result blew up within the first few days: 1.5k+ GitHub stars, some leading local AI developers began contributing, and the community rewrote it in Zig together. Then came MCDMA, letting Macs and NVIDIA machines share memory; two Mac Studios connected via Thunderbolt 5; and a model that keeps learning during use was released. Later, a DGX Spark was donated to him so he could continue working on TensorFold. Within four weeks, his followers grew from 300 to 9k+. He also called on the UK to build its own open models.
Pratik Gandhi said he was honored to collaborate and bring TensorFold into a YukonResearch challenge. The challenge is run by Darkbloom, Yukon, and EigenLabs together. Ash wrote in his tweet that CUDA is racing toward 200%. The performance chart gives weighted benchmark gains: CUDA on DGX Spark +146.6%, MLX on Apple M5 Max Metal +53.9%. The bottom of the chart marks a reference benchmark at 2026-10-31T10:46:00.000Z.
In the replies, someone was preparing to update the CUDA driver and rerun the benchmark, someone said their patch was submitted a step too late, and someone posted a screenshot showing they were close to the leaderboard. Ash replied that the community cooks too fast. One commenter joked that going from support tickets to TensorFold was a pretty good plot twist. Another said TensorFold had helped them a lot. Someone else said outright that the story inspired them to build and contribute.
The barrier to local AI is often not model weights, but drivers, memory, networking, and OS versions. Someone who spent seven years in technical support knows exactly how those places break. The AI industry likes to tell stories about flashes of insight in the lab. Ash's path looks more like another version: do enough boring troubleshooting first, then wait for a direction you can actually act on. At the end of his tweet, Pratik Gandhi wrote that the story is far from over. TensorFold's commit history is still moving, so that statement isn't hollow.
发布时间: 2026-11-01 15:42