Measurement first.
Research here pays twice: once as a paper, once as a capability the products and the enterprise work can use. One paper is published; the rest are under review or available as preprints.

Published
−41.8%
Latency Optimization in Production Voice AI Pipelines: A Systems Approach
A systems treatment of latency in voice pipelines running across more than 40,000 locations: where the time goes, and a set of changes that cut median end-to-end latency by 41.8 percent (3,277 ms to 1,909 ms).
Rodrigues, C. and Vas, O. (2026). In Artificial Neural Networks and Machine Learning, ICANN 2026, Part V, Lecture Notes in Computer Science, Springer, pp. 450-461.
doi:10.1007/978-3-032-38404-1_36 ↗Under review and preprints
Further manuscripts are under review at journals and conferences or available as preprints, across these lines of work. Titles, venues and preprint links are kept on the research index and the ORCID record so that status is always current.
Evaluating language models by severity
Probes that score how badly a model fails, not just whether.
Hallucination in generative models
How and when outputs drift from the source.
Developer identity and agent memory
Benchmarks for what an AI assistant should know about the engineer it works with.
Visual attention
Predicting where the eye lands first on a creative, the research line behind FirstGlance.
Trust and personality in language-model interaction
A review of what changes when a model has a character.
- 2026-08
First place, Data Visualization, Adaption AutoScientist Challenge
A dataset of real published chart distortions with a continuous severity measure, a reproducible fine-tuning pipeline driven through the platform's API, and a public instrument for reading the results. Won on data originality and depth of use, not on the leaderboard number.
- 2026
Vesuvius Challenge, tooling
An audit tool and atlas for checking ink labels on the Herculaneum scroll scans, contributed to the open effort. Not a prize entry; the kind of measurement work the papers above come from.
- 2026
Kaggle and hackathon entries
Medical imaging, agricultural ML, agent and trading-agent sprints. Some won, most did not; each left a reusable pipeline or a product, such as Nameproof. Results are stated exactly where they are cited.
Contributions go upstream to the projects WYTOR builds on. Merged pull requests include work in Medusa, SAP UI5, Ably and CALL-E, with more open and waiting on maintainers.
The full log, with links to each pull request, is kept on the founder's site.
Contribution log ↗- Nameproof speaks MCP, so any agent can run a name-collision check.
- labelscope the Vesuvius label-audit tool, is public.
- Skillfile will open when it has proven itself on WYTOR's own repositories.