Blog
Insights on AI engineering, product development, and what we're building at TRON Technologies LLC written by the founders.
The Long Road From 'It Works on My Machine' to Something Real
Getting a model to work once in a notebook is the easy ten percent. The gap between that and something dependable is where the real work hides.
Read Post →Two Android Audio Decoders. Same File. Different Answer.
Android ships two production-quality MP3 decoders. They are not interchangeable. One re-reads the format header on every frame and keeps going. The other locks in the format from the track header and silently stops the moment a later frame disagrees. The kind of bug you only find when you have two consumers of the same audio.
Read Post →It's Not the Compute. It's the Bandwidth.
Everyone measures on-device AI by compute: TOPS, benchmark scores, peak throughput. The variable that actually decides whether a phone can run two models at once is memory bandwidth, and it is the one nobody puts on the spec sheet.
Read Post →Quantization: What Actually Happens When You Try to Run AI on a Phone
Most ML engineers glaze over the second you say "4-bit versus 8-bit." That's a problem, because quantization is the single most important tool you have for deploying models outside of a datacenter. Lessons from building Foxleaf, a mobile Stable Diffusion app, across four generations of Qualcomm NPUs.
Read Post →Your On-Device AI Works Great for 20 Minutes. Then It Melts.
Memory is the constraint everyone plans for. Thermals are the one that nearly kills the project. Notes from running Stable Diffusion and Whisper concurrently on Snapdragon hardware, the death spiral that took down 20-minute sessions, and why pinning transcription to efficiency cores was the only viable fix.
Read Post →A Field Guide to the LLM Models Behind Modern AI
The phrase "large language model" covers a wide range of systems built and trained for different jobs. A field guide to the major families in use today, base models, instruction-tuned chat models, reasoning models, code models, multimodal models, embedding models, and small on-device models, with notes on what each is for and how the training process shapes it.
Read Post →What Goes Into an AI Art Style
I have been working on image art styles for AI image generation. Each one is actually a carefully written prompt that has to fight the model's own training biases, era anchors, and the constant trade-off between style fidelity and content fidelity. Notes from the process so far.
Read Post →From Figma to Lottie: Building a Character Animation Without After Effects
A looping character animation for a mobile app, built entirely in the browser without After Effects or any desktop tooling. Vector artwork traced in Figma, animated in Lottie Creator with layered keyframe tracks, and exported as a single JSON file ready to embed.
Read Post →How I Use AI for Code Reviews and Security Audits
When you're a small team shipping real software, you don't have a dedicated security engineer. I use AI to review code before anything ships: structured adversarial analysis covering XSS, CSP misconfigurations, insecure storage, and OWASP Top 10. Here's exactly how that workflow runs.
Read Post →Blacksmithing: A Development Methodology for AI-Assisted Engineering
The design document is the product. Code is a compilation artifact - disposable, regenerable, worth no more than a .pyc file. Tom Moore on the methodology behind how TRON Technologies LLC builds with AI.
Read Post →The Wrapper Problem
Most AI products are wrappers: a clean UI on top of GPT-4 or Claude, with a subscription and some prompting in between. When your product is a wrapper, the AI isn't yours. The behavior, the cost, the privacy terms, all controlled by someone else.
Read Post →What "Building Your Own Engine" Actually Means
It doesn't mean reinventing the transformer architecture from scratch. It means designing the full AI pipeline for a specific problem, owning every architectural decision between the problem and the output, and building the inference logic that makes it work.
Read Post →The Ceiling Problem
When you wrap an existing API, your product's ceiling is set by that API. Every competitor using that same model has the same ceiling. The competition shifts entirely to non-AI factors. Custom engines let you decide where the ceiling is.
Read Post →It's Harder. That's the Point.
Building custom AI engines takes longer and requires more expertise than integrating an API. That's exactly why it's worth doing. The things that are easy to build are easy to copy. Difficulty is how the moat gets built.
Read Post →Your Data Leaves Every Time
Every time you use a cloud AI app, your data leaves. Audio files, documents, images, conversations, all of it travels to a server. Most people understand this in theory. Few think carefully about what it means when the data being processed is personal or commercially sensitive.
Read Post →How On-Device AI Changes the Equation
On-device AI runs the model directly on your hardware. The data never leaves your device. Apple Silicon, Qualcomm Snapdragon, and modern NPUs have made this viable. The privacy guarantee is hardware-level, not contractual.
Read Post →The Founding Conviction
TRON Technologies LLC was founded on a single conviction: the best AI products will be the ones where someone actually built the AI. The market in 2026 doesn't reflect that yet. Most AI products are wrappers. The companies that built their own engines will have a compounding advantage that's very hard to close from behind.
Read Post →