About
I'm a Member of Technical Staff on the Reasoning team at xAI, where I lead search and factuality post-training for the Grok model family - powering real-time capabilities of Grok. In addition, I lead the fast/instant mode training for model in grok consumer app.
Previously I spent five years at Apple as a ML Engineering Manager in search and language modeling, and before that I was an Applied Scientist at Uber building fraud detection systems and a data scientist at LeanTaaS building scheduling systems.
What I Do
xAI
- Grok 4.20 / Grok 4.3:
- Led Grok 4.20 / Grok 4.3 instant/minimal-reasoning mode post-training.
- Led search & factuality post-training for multi-agents.
- Grok 4.20 / Grok 4.3 beta-1 ranked #1 in Search Arena upon launch.
- Grok 4.20 / Grok 4.3 reduced 80%+ factual errors from Grok 4 Fast when both equipped with browsing and search tools.
- Grok 4.1 & Fast:
- Led search & factuality post-training, reducing hallucinations by 70% in instant mode.
- Grok 4.1 Fast non-reasoning with agentic search powers Grok in Tesla, helping users seek information and navigate in seconds.
- Grok 4 Fast:
- Led large-scale search RL and SFT.
- Grok 4 Fast (Menlo) ranked #1 in Search Arena upon launch and achieved Pareto-frontier intelligence.
- Grok 4:
- Core contributor to Grok 4.
- Contributed the first set of synthetic data that makes agentic search work.
Apple
- Trained sub-15ms efficient language models that correct spelling, auto-complete, and query understanding for 1 billion users across various apps.
Uber
- Built real-time ML systems that cut fraud losses by 60+% at a third of the action rate.
LeanTaaS
- Built the optimization algorithm for scheduling and wait-time optimization serving the top 20 hospitals in the United States, cutting wait times by 30% during peak hours.
Education
M.A. Statistics, UC Berkeley · 2015–2016
B.S. Mathematics & Statistics, Purdue University · 2012–2015
Contact
tyzhang at berkeley dot edu
