Background

Most bugs at least have the decency to crash. They throw, they log, they leave a trail. The code at UpscaleAI doesn't get that courtesy.

UpscaleAI builds the switch silicon and networking fabric for large-scale AI clusters. Its flagship, SkyHammer, is an open-standard switch chip that wires GPUs and XPUs together at line rate, so a whole cluster trains and infers as one machine. Founded in 2025 in Santa Clara by Barun Kar and Rajiv Khemani, both Palo Alto Networks and Juniper Networks alumni, and backed by NVIDIA, Tiger Global, Temasek, and Salesforce Ventures, it hit a $2 billion valuation in under 18 months.

Their codebase models that silicon down to the bit: packet forwarding pipelines, ACL and TCAM matching, route lookups, ECMP tables, hardware counters. When a mask is off by one bit or a hash bucket is miscounted, nothing errors out. The chip just quietly forwards the wrong packet, at millions of packets a second. That's the bar.

The problem

This is where human review quietly gives up. A reviewer can catch a naming nit or a missing test at a glance. But asking someone to trace whether a mask hit both sides of a TCAM comparison, or whether a counter index survived a refactor three commits ago, is asking for an hour of unbroken concentration on a Tuesday afternoon with six other PRs waiting. Under a deadline, that read doesn't happen. The change looks clean, the tests pass, and the bug ships wearing a green check.

The switch

UpscaleAI brought Entelligence in on a trial and never turned it off, running continuous code review on their core hardware-model repo long past the trial window.

What was caught

A counter quietly writing to the wrong slot. With the counter-enable flag set, the index that should have pointed at the configured base offset was never assigned. Every counter would have reported the wrong number, and nothing would have complained. It kept creeping back across pushes until it was fully nailed down.

An ACL match thrown away for stale data. A register mixup silently dropped a real ACL match and served stale data in its place. The kind of bug that sails through every test except the one that matters in production.

A guaranteed crash, hiding one branch away. A variable was deleted from one code path while a later branch still reached for it. Any input that took that branch would hard-crash on contact. Caught and fixed.

A build racing against itself. A multi-target Makefile rule let parallel jobs regenerate the same file at once, each clobbering the others' output. Flagged as a reliability risk and resolved before it could produce a corrupt build.

Not one of these looks dangerous in a diff. They only surface if something follows an index or a mask all the way through, every time, without getting bored.

Results

On this codebase, roughly every other comment Entelligence left turned into a code change. Sit with that number. On most repos a hit rate like that would look too good to be true. Here it just means half of what the reviewer flagged was a real correctness problem, not a style preference, because at this depth there is barely any style left to argue about. And it held that pace cleanly for months past the trial, with no meaningful downtime.

Bottom line

When a wrong array index is a hardware bug and an unmasked comparison ships in silicon, review stops being hygiene and becomes the last line of defense. Entelligence found the exact class of bug that human attention misses, and caught it while it was still just a diff.

Try Entelligence on your own repos →