Four projects across shipped software, agent systems, and active research builds. Everything here is something I have actually built, tested, or released.
Open-source, local-first AI dictation for Windows. Press a shortcut, speak, and Voxdara transcribes locally before inserting the result into the application you are using. Built around local speech models, privacy-first inference, and a desktop workflow designed to stay out of the way.
WindowsLocal AI
Speech-to-textOpen source
A fictional university Biology course planner where a person and an AI agent act on the same live application state through WebMCP tools. The agent reads the plan and can replace the selected courses, but it cannot touch the scheduling preferences the human owns. Change the page yourself and the agent's next read sees that newer state, not a stale snapshot.
get_course_planread current state
set_course_planbounded course replacement
A local-first tool for AI coding agents, aimed at project context, evidence, scope, freshness, invariants, and human-controlled authorization. Foundation components are implemented in the repository. It is still under development, has no published release, and there is nothing to demo yet.
foundationimplemented
releasenone yet
Small hands-on builds I use to test what a new model can actually produce, past the benchmark score. Each one ends as a page you can open, inspect and use.
Click image to preview webpage
A research-driven webpage documenting my hands-on evaluation of Astra across capability, reliability and practical use.
Click image to preview webpage
An interactive solar system used to test whether Astra could produce a detailed visual web experience from a constrained prompt.
Click image to preview webpage
A hands-on evaluation of DeepSeek across coding, reasoning, agentic tasks and real-world use.
Areas I actually spend time in. Building, breaking, and reading about. Some of it is work, some of it is study. None of it is a rating.
Adapting an open-weight base model to one narrow job: choosing and cleaning the data, tuning it, then checking afterwards whether the change actually held.
Deciding whether a change actually helped. I am more interested in a test that fails honestly than a number that looks good.
What an agent can see, what it is allowed to touch, and how much of the loop stays under a person's control.
The system around the model matters as much as the model itself: what context it receives, which tools it can use, how actions are executed, and where to look when something goes wrong.
Weights you can inspect, run and modify. What a small model can be pushed to do reliably, and where a larger general-purpose model is simply the better tool.
Not a methodology. Just the rules I keep coming back to when a system starts getting away from me.
Evidence first.
A claim about a project traces to something you can open, run, or read.
Human approval.
A person stays in control of anything consequential. The agent proposes, the human decides.
Real experiments.
Build the thing and watch it fail. A demo that only works once is not a result.
Verify, then say done.
Rendered, clicked, and checked before it gets called finished. Not inferred from the source.
Understandable systems.
If I cannot explain how it works at three in the morning, it is too clever to keep.
Unanswered on purpose. The ones I keep coming back to whenever a result does not hold up.
I started out building software and wiring AI systems together to see what held up. The interesting layer turned out to be the one underneath: how an agent gets context, which tools it can use, how much authority it should have, and where to look when something goes wrong.
That pulled me into evaluation, because “looks good in chat” is not a result. Then I moved deeper into the models themselves: open weights, post-training, and specialization. Eventually those threads came together in WEIGHTWRIGHT.
WEIGHTWRIGHT is an AI research and engineering company focused on specialized models for narrow, verifiable tasks.
The work centers on model post-training, evaluation, failure-driven iteration and reliable deployment. Narrow tasks where a result can be checked rather than argued about.
WEIGHTWRIGHT Research
The research arm: experiments, evaluations, technical reports, datasets and research models.
Status
Being built now. This site stays personal: it is where I write about the work, not a company page.
Website
I record what I am building: the parts that work and the parts that do not.
YOUTUBE / @heytenzyWEIGHTWRIGHT first: specialized models for narrow, verifiable tasks. Around it the same threads as always: coding agents, evaluation, and post-training on open weights. Voxdara has a public Windows alpha, WebMCP is live, and DSH Forge and Evidline continue.
Post-training and evaluation, mostly: how small a specialist model can get before it stops being reliable, and what the harness should control instead of the model. The open ones are listed under Questions I'm exploring.
On GitHub. The projects above link straight to their repositories, all under github.com/TenzyZ
For research and engineering conversation, yes. Email is open and I read everything. My working hours go to WEIGHTWRIGHT, so I am not taking on client projects right now.
It lets a web page expose tools that an AI agent can call, so the agent and the person work against the same live application state instead of the agent guessing from a screenshot. The course planner is my working example of that.
Contact
Research, engineering, collaboration, or something interesting.
If any of the questions above is also yours, or you have something worth discussing, my inbox is open.
