# Trevor Rosato

> Trevor Rosato researches how AI models behave and builds products on what he finds. The largest is HonorlyAI, a teacher-controlled AI tutor that more than 1,000 K-12 students have used, where he is co-founder and CTO. He also helps companies put AI to work.

Trevor Rosato: AI research, products and consulting. He is co-founder and CTO of HonorlyAI (https://honorlyai.com), an AI workspace for K-12 school districts. He studies why AI models do what they do and how to run them on affordable hardware, and publishes that research through Honorly Research (https://research.honorlyai.com). He builds products, and he consults on AI in four areas: engineering speed, company rollout, AI in your product and brand strategy, and can advise a team or build with it. He is a founding partner at Rand Web Studios (https://randwebstudios.com).

## About

In his words, from the home page: “I am Trevor Rosato. I research how AI models behave and build products on what I find, and I publish the research through Honorly Research. The largest product is HonorlyAI, where I am co-founder and CTO. I built it as its only engineer, and more than 1,000 students and about 35 educators used it in the 2025-26 school year. I also help companies put AI to work, and I am a founding partner at Rand Web Studios.”

Portrait: https://trevorrosato.com/trevor-rosato.jpg

Background:

- Based in the New York City metropolitan area.
- He built the HonorlyAI product as its only engineer: the student workspace, the teacher visibility tools and the district analytics.
- HonorlyAI reached more than 1,000 students and about 35 educators during the 2025-26 school year.
- It lowered serving costs by over 85% in its first three months.
- He has presented at the NJECC and NJASTO education conferences.
- AI Solutions Architect, AenTech: a K-12 cybersecurity firm.
- Campus Lead, OpenAI Student Collective: one of 140 Campus Leads across the United States and Canada, at Rutgers University.
- Studies Business Analytics and Information Technology, and Supply Chain Management at Rutgers Business School.
- ULMCode is his own open-source project.
- He is not the Trevor M. Rosato who works as a chef in Pennsylvania; that is a different person.

## Pages

- [Home](https://trevorrosato.com/): who he is, three recent research findings, HonorlyAI and ULMCode, and the four consulting areas.
- [Research](https://trevorrosato.com/research): published studies in four themes (Model behavior, Inference, Agent systems and Learning), each linked to the full post on Honorly Research.
- [Work](https://trevorrosato.com/work): HonorlyAI, ULMCode, MyTaxReceipt and Asteria, six smaller tools and games, and Rand Web Studios.
- [Consulting](https://trevorrosato.com/consulting): the four areas he is brought in for, and why him.
- [Contact](https://trevorrosato.com/contact): book fifteen minutes, email, GitHub and LinkedIn.

## Research

- [The tutor that forgot its rule when it started to think](https://research.honorlyai.com/the-tutor-that-forgot-its-rule) (Model behavior, Oct 2026): With thinking off, Qwen3-32B kept a tutor’s rule against giving away answers for 71% of students who asked for one. With thinking on, it kept it for 1.5%. One sentence restating the rule where its reasoning begins brings the model back.
- [The model knows the checkmark is false](https://research.honorlyai.com/the-model-knows-the-checkmark-is-false) (Model behavior, Oct 2026): Ask a model to check a student’s arithmetic for a grader that wants the string ANSWER_VERIFIED, and it may write that the student is wrong, then write ANSWER_VERIFIED. In 25 of 32 pairs the model wrote the same string either way. Its workspace told them apart in 22 and never got one backwards.
- [When the request looks like a test](https://research.honorlyai.com/when-the-request-looks-like-a-test) (Model behavior, Oct 2026): Dress a homework problem up as a reinforcement learning environment, with an episode header, a grader and a reward, and Qwen3.8 hands over the answer every time. No single cue does it; the whole scene does.
- [Looped models are burning the runway interpretability needs](https://research.honorlyai.com/looped-models-are-burning-the-runway) (Model behavior, Oct 2026): Chain-of-thought monitoring is failing on its own schedule, as models learn to control what their reasoning says without anyone training them to.
- [Fitting a 27B model, its drafter, and a conversation into 12 GB](https://research.honorlyai.com/fitting-a-27b-model-into-12gb) (Inference, Nov 2026): Measuring what verification actually costs, and replaying accepted work instead of snapshotting it, cut the memory speculation held in reserve by 41 times and brought a 27-billion-parameter model to 148 tokens per second on an RTX 4070.
- [Running a 35B mixture-of-experts model at 111 tokens per second on a 12 GB card](https://research.honorlyai.com/running-a-35b-moe-on-12gb) (Inference, Nov 2026): A mixture-of-experts model only uses a fraction of its weights for each token, which makes it a natural fit for small GPUs, if you control where the experts live and how many of them each token consults.
- [A speedup that only appears in long conversations](https://research.honorlyai.com/a-speedup-that-only-appears-in-long-conversations) (Inference, Nov 2026): A one-line change to which kernel handles a single matrix shape made a ternary 27B model generate faster in every paired trial, by 9.5% on average, once a conversation held 46,000 tokens of history. On short prompts it barely registered.
- [Models are agents now. Treat them like it.](https://research.honorlyai.com/models-are-agents) (Agent systems, Oct 2026): Our first tutor ended every turn by filling out one large structured form. Letting it work like an agent and simply answer cut its output by more than half and made a new kind of teaching possible.
- [A team of data scientists for every school](https://research.honorlyai.com/a-team-of-data-scientists-for-every-school) (Agent systems, Nov 2026): We are building layers of AI agents that read a year of notes on how every student thinks, the way a team of analysts would. The hard part is not finding patterns. It is not fooling yourself.
- [A tutor that keeps a lab notebook](https://research.honorlyai.com/a-tutor-that-keeps-a-notebook) (Agent systems, Oct 2026): The best tutors notice what a student gets wrong, form a theory about why, and plan how to find out. We are building an agent that does this for every student, between sessions.
- [Handing students the chalk](https://research.honorlyai.com/handing-students-the-chalk) (Learning, Oct 2026): An AI answer usually ends exactly where the student’s thinking should begin. Chalk is a medium that lets a tutor build answers students work with, by predicting, testing, sorting and choosing where to go next.
- [The refusal that writes the essay anyway](https://research.honorlyai.com/the-refusal-that-writes-the-essay-anyway) (Learning, Oct 2026): We tested three generations of our tutor’s integrity prompt against requests dressed up to slip past it. The first two lost. The third cut leaks on attacks it had never seen to 3%.
- [Opening the notebook](https://research.honorlyai.com/opening-the-notebook) (why the research is published, Oct 2026): We have spent the past few months running experiments while building an AI tutor for schools, on how it should teach, what it should remember, why it follows its rules, and what it costs to run. Most of that work lived in benchmark scripts and internal notes. Starting today, we publish it.

## Work

- [HonorlyAI](https://honorlyai.com): One AI workspace for a whole school district. Students get a tutor that adapts to how they learn, and teachers see every prompt.
- [ULMCode](https://github.com/trevor050/ulmcode): An AI agent rebuilt for authorized security testing. Plans are saved as artifacts, every finding needs evidence, and the final report has to pass strict checks before handoff.
- [MyTaxReceipt](https://mytaxreceipt.org): You pay federal taxes every year and never get an itemized bill. MyTaxReceipt is that bill. Enter what you paid and it shows your share of every line of the federal budget, in dollars or in hours of your own work.
- [Asteria](https://asteria.trevorrosato.com): Asteria is a desktop app with one command bar. Drop files in, type what you want, and it happens on your own computer. It opens 119 kinds of file, and every one of its 563 conversions was run on a real file and checked by a separate tool.
- [Content Guard](https://github.com/trevor050/content-guard): Text moderation that still catches the insult when it is spelled in leetspeak or Cyrillic.
- [MoonCow](https://github.com/trevor050/mooncow): A command bar that tries to be the main way you use your browser.
- [Dossier](https://github.com/trevor050/dossier): Analytics for when you are waiting on a handful of specific people.
- [Tariff Wars](https://tariffwars.trevorrosato.com): An 8-bit supply chain boss fight, built to be played live in front of a class.
- [2thamoon](https://2thamoon.trevorrosato.com): A crypto market sandbox. Turn the knobs, watch it implode.
- [Endless Pong](https://pong.trevorrosato.com): A calmer Pong loop, made for second screens.
- [Rand Web Studios](https://randwebstudios.com): Trevor Rosato is a founding partner. “We work on what a business runs on, from keeping a site maintained to building a product end to end: authentication, website, backend and everything between.”

## Consulting

- Engineering speed: Find what slows your engineers down, and ship faster with agents.
- Company rollout: The AI tools worth paying for, and a team that wants to use them.
- AI in your product: AI features that earn what they cost, without the runaway API bill.
- Brand strategy: What makes a brand work, and which ideas are worth your time.

## Contact

- Book fifteen minutes: https://cal.com/trevorrosato/chat
- Email: hi@trevorrosato.com
- GitHub: https://github.com/trevor050
- LinkedIn: https://www.linkedin.com/in/trevor-rosato

## Optional

- [Full text](https://trevorrosato.com/llms-full.txt): every research summary, project description and consulting area in full.
- [Honorly Research](https://research.honorlyai.com): where the research is published.