- NVIDIA's Maria Victoria Gromysch said her team uses statistical process control on line data because otherwise its AI could filter out all the outliers.
- Her team expects its agent to cut review time, but she did not say how the control sits next to the model's own scripts.
Maria Victoria Gromysch, a test engineer at NVIDIA, pointed to a quiet risk in factory AI: a model that makes the numbers look clean. She spoke in a GitHub Universe 2026 session published on the GitHub channel, about an agent that reviews chip manufacturing data. Her team uses statistical process control, a traditional quality-control method, on its manufacturing-line data.
What was said
Gromysch began with the goal. In her words, "Yield is the percentage of devices that pass manufacturing validation tests that are ready to ship to the customer."
The agent gathers data from Splunk, diagnostic releases, and bugs filed by customers and vendors. The volume is large, so the data is pre-processed first. Outliers are filtered and summaries are generated. The language model writes scripts that summarize and evaluate the data, then feeds the result to a handover agent. That agent runs at 9 a.m. and 5 p.m., and files issues for the right engineer.
She said the team expects to cut review time by 60%. She also said finding a root cause used to take several weeks, up to six.
An audience member asked whether the agents are simply turned loose on raw data. Gromysch said the team put in guardrails and uses statistical process control to evaluate the line data. The reason, she said: "Otherwise it could just filter out all the outliers" and the yield picture would look perfect. She did not detail how this control sits relative to the model's scripts.
Why it matters
Our reading: an outlier is often the fault you are hunting. A model asked to tidy data can remove exactly those points, and the report still looks professional.
A general note on the method, not from the session: statistical process control tracks measurements against limits calculated from the process's own statistics. It flags points that fall outside them. That differs from a model deciding for itself what looks odd.
So for leaders buying or building AI review tools, speed is not the only question. Ask what the tool discards, who set that rule, and whether a check independent of the model tests it. Gromysch's session did not settle those points for her own system.
The other side
Gromysch did not say who sets the control limits, or whether the control runs before, after or alongside the model's scripts. She gave no measure of how often it catches a wrongly removed outlier. The system is young. She said it was released a month or two earlier and is at the first stage of evaluation.
The 60% was stated as an expected saving, not a measured one. Asked about improving the system with reinforcement learning, she said that is something the team will work on.
She also gave a practical reason to lean on the agent. The team had considered traditional machine learning, such as classification and anomaly detection. But it needed summaries and action items, because the team is overwhelmed with issues and has limited bandwidth. That tension was not resolved in the session.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- GitHub: Building domain-specific AI agents for hardware reliability at NVIDIA (2026-11-20)





