23 July 2026
Over the last two days, Trelis Research has released two open-weights models: Trelis/Tara, a frontier model for Hindi and Hinglish transcription, and Trelis/Tiron, a multi-speaker transcription model performing ahead of commercial APIs on long meetings.
Next, I am taking these stepping stones much further, with a new company focused solely on conversational intelligence that is seamless, trusted, and judged by what we accomplish with its help.
I believe voice will grow enormously in the coming years as our primary means of interacting with technology.
I am beginning the search for a founding team across four specialties:
- Evals & Post-training
- Infrastructure (training + inference)
- Data
- Architecture & Mid-training
Think Karpathy in 2015, working with the tools of 2026+. The search is global. The work will be intensive, focused, and in-person. The location will be decided based on getting the founding team together in one place. The team will be well compensated. The company will handle relocation and visas for team members and their families.
I’m looking for nominations for each of these roles, which you can submit by writing to ronan at trelis dot com with:
- Who you’re nominating, and for which specialty
- Why they are absolutely exceptional, pointing to a specific project
- If we don’t end up hiring this nominee, what the most likely reason would be
More info on these four positions, ordered by priority
Evals & Post-training
You will:
- Own our evaluation benchmarks: the company’s ground truth for what our models should be capable of doing through voice, informed by customer feedback.
- Lead post-training and release of our models, beginning with v0.
- Build the systems that iterate on data mixtures and post-training runs.
Infrastructure (training + inference)
You will:
- Build self-improving and self-healing systems that orchestrate training and inference. They must be extremely reliable, extremely cost efficient, and able to scale robustly as we grow.
- Own compute capacity: decide and manage our investments in networking and compute, laying out our six-month and six-year strategic plans.
Data
You will:
- Decide and manage our investments in data, with great emphasis on quality control and the cost versus value of any data source. You must develop a clear framework for prioritising data we collect ourselves versus procure, taking into account the obligations and responsibilities of each source. You must think about how we get to the best data for our needs on a six-month and six-year timeline.
- Build the shared pipelines that evaluate, filter, prepare and move data, delivering quality-audited datasets to the training teams against their specs.
Architecture & Mid-training
You will:
- Lead the development of our base models, beginning with v1.
- Build the experimentation system that trains and evaluates new architectures to climb our benchmarks, a system that improves its own experimental design and ability to efficiently allocate compute over time.
Ronan