Intelligence that stays on-device

Open-source chatbots that run entirely in your browser. Transformer embeddings, BM25 retrieval, dense-sparse ensemble ranking, and a 25-feature MLP policy network — zero servers, zero LLMs, zero telemetry.

Game Theory Chat Golden Age Inquiry Data Science Chat ML Tools Source
New ebook
Get Fringe Learning: Resource-Efficient RL for Edge ML for $1

Build your own custom AI systems with practical reinforcement learning methods designed for resource-constrained edge machine learning.

Get the ebook
How It Works
01 Embed
Sentence Embedding
sentence-transformer (MiniLM-L6-v2) downloads once and is pre-cached by the Service Worker. Progressive loading: heuristic/BOW fallback gives instant first-turn answers while the model warms up; full dense embeddings hot-swap automatically. Query memoization + top-k keep repeats fast.
02 Signal Layer
Multi-Signal Extraction
BM25 sparse retrieval, entity extraction, ensemble dense-sparse ranking, and temperature-calibrated intent scores are fused into a structured DecisionPacket — all locally.
03 Decide
Policy Network
An MLP policy network (13K parameters, RL-trained) determines mode, intent, topic count, creativity, and tone from your extracted features.
04 Compose
Fragment Composition
Fragment-based assembly with linguistic connectors, comparison openers, and tone modulation produces precise, varied, natural-sounding answers.
05 Render
KaTeX & Session Memory
Responses render with KaTeX math, formatted text, and session memory for follow-up context — all without a server round-trip.
Full Architecture Details
Features
On-Device AI
Transformer runs entirely in your browser. No data leaves your machine — not even embeddings.
Privacy-First
No accounts, no cookies, no tracking, no server processing. Just you and your browser.
Open Source
Fully transparent MIT license. Inspect the code, fork it, extend it. Built for the community.
No LLMs
Deterministic, explainable responses. No hallucinations, no API costs, no vendor lock-in.
LaTeX Support
Mathematical notation rendered beautifully with KaTeX. Perfect for game theory, statistics, and ML formulas.
Browser-Based
Works anywhere with a URL. No installs, no dependencies. Desktop, tablet, or mobile.

Game Theory Chat

On-device NLU assistant for Nash equilibrium, Shapley value, auctions, prisoner's dilemma, and more. Ask in natural language — get precise, mathematical answers with LaTeX rendering.

Strategy On-device AI LaTeX Mathematics
Launch Chat
Game Theory Chat
How do I find the Nash equilibrium?
For a 2x2 game, check each player's best response. If both play best responses to each other, that's a Nash equilibrium.
$u_i(s_i^*, s_{-i}^*) \ge u_i(s_i, s_{-i}^*)$
Can you compute the Shapley value?

Golden Age Inquiry

Explore the scientific and philosophical discoveries of the Islamic Golden Age. Ask about Al-Khwarizmi's algebra, Ibn al-Haytham's optics, the House of Wisdom, and the scholars who shaped modern science.

History of Science Philosophy On-device AI Mathematics
Launch Chat
Golden Age Inquiry
Who was Al-Khwarizmi?
Al-Khwarizmi (c. 780–850) founded algebra and introduced algorithms. His name gave us "algorithm" and his book gave us "algebra."
He worked at the House of Wisdom in Baghdad, synthesizing Greek, Indian, and Persian mathematics.
What is the Tusi couple?

Interactive ML Tools

Explore the math behind machine learning with six interactive visualizations: neural networks, gradient descent, backpropagation, activation functions, k-means clustering, and decision trees — all running in your browser.

Neural Networks Clustering Decision Trees Activations On-device
Explore Tools
Neural Network Explorer
Input Hidden Output

Data Science Chat

On-device NLU assistant for data science and machine learning. Ask about pandas, NumPy, statistics, classification, regression, clustering, feature engineering, model evaluation, and more — all running locally in your browser.

Data Science On-device AI Statistics Machine Learning
Launch Chat
Data Science Chat
What is logistic regression?
Logistic regression models the probability of a binary outcome using the logistic function $\sigma(z)=1/(1+e^{-z})$, where $z=\beta_0+\beta_1x_1+\cdots+\beta_px_p$.
It's a linear model for classification — despite the name, it's used for predicting classes, not continuous values.
Explain cross-validation
ReLU.chat On-Device Chatbot Builder Toolkit — local intelligence, privacy by design
Premium · From the makers of ReLU.chat
Build your own on-device chatbot

You've seen what local, privacy-first AI can do on this site. The On-Device Chatbot Builder Toolkit hands you the architecture behind it — frameworks, templates, and evaluation guides covering the full pipeline, from raw input to validated response.

Full pipeline: Input → Response Frameworks & templates Retrieval + policy + privacy
One-time purchase · $29
Get the Toolkit
About

ReLU.chat is an open-source experiment by Yunus Emre Vurgun to discover simpler, faster ways to build chatbots — without creating an LLM or relying on pre-LLM era chatbot logic.

The goal is to combine sentence transformers, BM25 sparse retrieval, reinforcement-learned policy networks, and fragment-based composition into a system that runs entirely in the browser. No servers, no API keys, no data collection. The power stays on your device, achieving absolute data privacy for anyone who uses it.

yunusemrevurgun.com Source Code
Install ReLU.chat for offline access and faster loading.