Research & Science
ARCHITECTURE.md · 21 documents
Qwen Image and Edit: Open-sourcing and Local GGUF Generations with Lightning
Documents the Qwen-Image and Qwen-Image-Edit models, covering architecture, training, benchmarks, ComfyUI setup, and prompting techniques for local GGUF deployment.
Smart Implementation Assistant - Complete API Design & Architecture
Defines a phased implementation plan generator that turns medical recommendations into step-by-step action schedules with routine integration and monitoring.
Dense Prediction of Mammography
Describes a ConvNet for dense prediction of mammography abnormalities using CBIS-DDSM data, with training strategies and results.
System Prompt Architecture Research & Findings
Analyzes system prompt architectures from Claude Code, Cline, and Bolt.new to recommend modular, tool-centric improvements for Better-i18n's AI agent.
draft
Describes a permutation-invariant neural network architecture that processes sensory inputs in any order using an attention mechanism.
Qwen3-TTS — Model Reference
Documents the architecture, weights, tokenizer, and inference algorithm for implementing Qwen3-TTS from scratch.
Part 2: Latent Action Prediction Model
Builds a VQ-VAE model to infer discrete latent actions between consecutive Atari game frames for next-frame prediction.
IGT Risk-Sensitive RL Video Script
Scripts a 6-7 minute video explaining a risk-sensitive RL model for the Iowa Gambling Task, comparing it to a baseline DQN and human data.
AI Implementation Guide: CS01 Case Study
Documents how two AI features, Term Sheet Analyzer and DealCraft Negotiation, are built with Gemini, Supabase, and Vercel serverless functions.
HyQCOpt: A Unified Framework for Distributed Hybrid Quantum-Classical Optimization
Describes a three-layer hybrid quantum-classical optimization framework that claims 22.3x speedup on NP-hard problems using Grover amplification and distributed QPUs.
Project Sarpeidon — Warp Core Simulation Engine
Documents a Star Trek warp core simulation engine with AI ensemble, physics engine, and holographic display integration.
Deep Learning-Based Automated Detection of Brain Tumors from MRI Scans
Describes a full-stack architecture for detecting brain tumors from MRI scans using a U-Net deep learning model.
RFC Document and Community Memorandum
Opens a community review process for a decentralized AI training network design, inviting feedback on architecture, consensus, and tokenomics.
Archipelago Alerts - Project Summary for LLMs
Summarises the architecture, stack, and key components of an Archipelago multiworld game tracker with a Flask backend and Android frontend.
Increasing pervasiveness of machine intelligence
Outlines sociocultural, ethical, and political implications of machine intelligence, with links and discussion prompts for a book chapter.
An Image is Worth 16X16 Words: Transformers for Image Recognition at Scale
Introduces the Vision Transformer (ViT) that applies a standard Transformer directly to image patches for classification, outperforming CNNs when pre-trained at scale.
The Hype Surrounding DeepSeek-OCR
Collects and contextualises public claims about DeepSeek-OCR's context-compression potential that motivated a research paper.
MCP-Titan HOPE Alignment & Implementation Plan
Audits a codebase against a research paper, catalogues missing features, and lays out a phased build plan with code examples.
Nested Learning: The Illusion of Deep Learning
Introduces Nested Learning, a paradigm representing models as nested optimization problems, revealing how deep learning compresses context flows and enabling higher-order in-context learning.
FHL Bible MCP Server — Research Notes
Evaluates a local Python MCP server that wraps the FHL Bible HTTP API into 27 MCP tools for Claude Code, then decides not to integrate it.
CMCP: Cognitive Multi-Model Communication Protocol
Defines a provenance metadata schema for multi-model systems to tag outputs with epistemic position, verifiability latency, ground truth access, and validation state.