Quantization
Round 64 synthetic weights to fewer bits and watch the error grow.
Uniform symmetric quantization of a synthetic weight vector. An illustration, not a benchmark.
Palm Coast, FL··
ai engineer & game developer, Palm Coast, Florida
I ship tools, games and AI experiments, and I’m building Basketball Heroes, a multiplayer VR basketball game in Unity.
I’m an independent developer, studying AI engineering, machine learning and large language models. On GitHub: 221 stars on roblox-ai, 21 public repos and 254 contributions in the last year.
Also a .
Now
Developer & Manager, 2026 – Present
A multiplayer VR/XR basketball game built with Unity. I develop the gameplay mechanics and interactive systems, work with VR/XR, physics and multiplayer, design the interface and player experience, and coordinate development, beta testing and project management.

Present
Independent
Building websites, applications and developer tools, plus Discord bots, integrations and automation systems. I work on personal and open-source projects and keep experimenting with modern frameworks.
Present
Personal projects & experiments
Exploring large language models and AI-powered applications and developer tooling. I experiment with open-weight models and local inference, and I’m learning how tokenization, quantization, model architecture and training work. Taking AI classes to deepen my knowledge of AI engineering and machine learning.
Personal projects
Independent
Developing interactive game experiences and prototypes across Unity and Roblox Studio. I design polished interfaces and animations, build gameplay systems and work on making them feel good to use.
Build Roblox games with any AI agent: Claude, Codex, Gemini, OpenCode and Grok, wired into Studio. Free and open source. repo
12,000+ Roblox UI references for Claude Code, Codex and other AI coding assistants to learn from.
A skill that helps Claude Code and Codex build modern, good-looking Roblox game interfaces.
Show your live Claude Code chat, model and token usage on your Discord profile.
A fast Discord client that uses less RAM.
This site. Hand-written HTML, CSS and a little JavaScript, no framework.

01 / 18ClassicFrom Basketball Heroes. Pick one.
I’m learning how language models work by taking them apart. Two small experiments, running in your browser:
Round 64 synthetic weights to fewer bits and watch the error grow.
Uniform symmetric quantization of a synthetic weight vector. An illustration, not a benchmark.
A tiny byte-pair encoder, trained live in your browser.
Learned from a short built-in text, so common words merge into single tokens and unfamiliar ones fall apart. Real tokenizers train on vastly more data. A dot marks a leading space.
Building a tokenizer from first principles.
A neural network from scratch in NumPy: backpropagation and gradient descent.
A minimal tool-calling agent runtime with schema-validated tools and loop detection.
Deterministic LLM output evaluation with JSON contracts and regression gates.
Offline guardrails for LLM pipelines: PII redaction, secret detection, prompt-injection screening.
Zero-dependency span tracing for LLM apps with latency percentiles and token rollups.
Reproducible model-routing experiments with UCB exploration and budget caps.
Offline BM25 passage retrieval with citations, overlapping chunks and a Python CLI.
254 contributions in the last year
Have something to build? Message me on Discord () or X (@IsaiahSrc), or read the code on GitHub.
click anywhere to close