Projects & Contributions

Multi-model/provider client for running inference against various LLMs. Handles API inconsistencies between models, allows replay of chat threads (including tool calls) against various providers. Future work includes a grid optimizer for tuning prompts, models, and hyperparams.

llama.cpp NIF extensions (via Rust/Rustler) for loading and running inference against models directly from Elixir. All public non-async endpoints for Session and Models are exposed. Auto-infers best chat completion format (ChatML, Zephyr, LLama2Chat, etc.).

Robust Elixir client for OpenAI endpoints. More feature-complete than the GenAI client for OpenAI-specific work: fine tuning, audio, threads, assistants, and more.

Elixir macros for creating and querying Weaviate vector database records. Define schemas with a DSL and interact with the Weaviate API using native Elixir structs.

Conventions and system prompts for improving predictability of model output and behavior, with a focus on middleware system reliability. Defines in-fill syntax, agent definitions, clip continuations, and more.

Framework for composing multi-model/multi-agent/multi-channel virtual workgroups. Detailed architecture and documentation for multi-agent collaboration.

Advanced implementation of multi-agent systems. Further along in implementation compared to NoizuTeams with more complete logic setup.

SMart As Hell - Command line GPT tool for Linux servers. Creates user profiles with system metadata and experience level for contextualized assistance. Predated Microsoft and Anthropic's similar solutions.

Thoughts on composable AI systems, arrow of time/time decay activator functions, and a draft paper on dynamic runtime model tuning via additional QLora input nodes.

PHP library for fragmented key cache management. Efficiently handles cache invalidation across complex dependency graphs.

Java port of the FragmentedKeys library for managing cache key dependencies.

Implementation of fragmented keys for managing data stored across various iOS cache providers.

Protocols for DB/Config driven runtime rule engines. Great candidate for hooking operations to GenAI and other models to drive runtime behavior.

Scaffolding for managing domain objects and repositories with consistent patterns.

Scaffolding for managing large pools of long-lived working processes with built-in health monitoring.

Base records and protocols that drive other Noizu scaffolding libraries.

Domain-driven design objects for PHP applications.

Gherkin/Cucumber extension for PHPUnit. Uses reflection to route BDD-style test lines to handler methods. Supports data-driven tests with Examples tables and regex matching.

Library for defining reusable custom asserts for partially comparing objects, including verification that values are within or outside of specified ranges.

Embedded C streaming parser that tokenizes JSON input byte-by-byte. Built for devices that can't parse/store inbound data fast enough to avoid watchdog triggers. Uses nullable bit-aligned union data structures. Blazingly fast.

C# CLI tool that generates precompiled trie data structures for extremely efficient UART parsing and JSON tokenization. Supports compressed binary blobs, arrays, and structs. Blazingly fast.

Custom XML tags for creating and tearing down PostgreSQL enum types with Liquibase.

UML to Swift code generation support for Sparx's Enterprise Architect.

GitHub API client for Elixir, built for letting AI agents access/insert tickets and comments on projects.

Extended Mnesia wrapper with callback hooks for telemetry, local cache with cluster sync, RocksDB support, compression, and a database emulator for unit/system testing.

Fork with added project switching capabilities for Google DataStore. Useful for synchronizing environments.

Heavily extended with dynamic template (Handlebars) support, sub-version selector for dynamic messages, and template version query/update calls.

Optimized timezone data handling for high-load environments. Moved lookups from ETS into FastGlobal/persistent_terms for billions of TZ operations per minute.