Founder of Tribunus, an independent software company in San Francisco. Engineer. Sole maintainer of Tessera and Tessera Studio. Before that: software engineer, freelance photographer and designer.
The work is on GitHub. Each item says what it is and where it stands.
A C++ fork of llama.cpp, tuned so models run fast and use less memory: per-tensor calibration (T640), speculative decoding (DFlash / DSpark), and Apple Neural Engine prefill on M-series Macs. Outputs match the reference model, run for run.
The desktop app that puts Tessera to work: an agent loop, manifest-based skills (SKILL.md), and files that stay on your computer. macOS first; Linux in progress.
The Python tooling behind the calibration: per-tensor search, AWQ and outlier handling, and per-step spec-decoding records (llama.tessera.spec.v1).
An experimental AI systems project in Rust — open research, not part of Tessera's shipping path.
These apply to the code, the company, and this site.
I research and design before writing code. When the plan and the evidence disagree, the evidence wins.
Code lands only when the actual build passes — not a script that claims it does. Progress only moves forward.
Every run keeps a log — what loaded, what ran, what happened. If we can't check it, it doesn't count.
No v1/v2/v3 copies of the same code. Refactor in place — a version is a release tag, not a label.
External accounts, keys, and APIs are risks we design out — not defaults we inherit.
I use AI tools heavily, but I review everything that ships and can explain all of it. The tool drafts; I own what goes out.
I founded Tribunus in October 2024 in San Francisco with one conviction: AI should run on your own computer — fast, private, and useful — and work with you, not against you. That's what Tribunus is built around, and Tessera Studio is its first product.
Before Tribunus I worked as a software engineer in the Bay Area, and before that as a freelance photographer and designer. That background shows up in the product — the care taken with design, documentation, and the feel of every screen in Tessera Studio.
I am currently building one product — Tessera Studio — as the sole maintainer of both the Swift app and the C++ engine underneath it. I am the only full-time person, and I intend to keep it that way until the product earns a team.
"I am one founder, doing one product, and I am not going to ship it badly." — Julian Torres, founder of Tribunus
Julian Torres founds Tribunus in San Francisco as an independent, founder-led software company. The idea: AI should run locally, privately, and work with you — not against you.
The C++ side is named Tessera. Per-tensor calibration (T640), DFlash / DSpark drafters, and Apple Neural Engine prefill land.
The desktop product is renamed Tessera Studio. The Swift app and the C++ fork ship from a single repo — one company, one product.
Tessera Studio is in developer preview on macOS with Linux building in CI. The company is pre-seed and self-funded. One founder, one product, no hires planned before a pre-seed close.
For press inquiries, podcast appearances, conference talks, or partnerships related to Tessera Studio or the calibration work — reach out directly.