My tinkerings
Since 2023 my focus has mostly been on projects related to LLM-driven software development, with some hardware-related side quests here and there.
I learn through building, so while some of the projects are used in production right now (like llmist or cascade), many are purely exploratory in nature.
Running 5-6 agents in parallel started taking a toll on my cognition, making it the limiting factor, so I decided to build an orchestration layer that exposes only the touch points I care about along the entire SDLC, allowing me to build faster with better quality and not go crazy in the process. It’s open source, integrates a range of agentic harnesses with existing PM/SCM toolsets. I'm using it to develop my other projects (including itself) and introducing other teams to it.
Streaming-first multi-provider LLM client + agentic harness in TypeScript. The agent and tool layer most of my other stuff sits on top of. Home-rolled tool-calling format, with a built-in DAG so independent calls run in parallel. Fully featured CLI and TUI. Budget-aware, supports MCP, remote skills. Three-tier hook system for asynchronous integrations. Supports Anthropic, OpenAI, Gemini, HuggingFace and OpenRouter.
Personal agentic operating system. Integrates email, calendar, meeting transcripts, tasks, documents, time tracking and more. Somewhere between Claude Cowork, Notion and ClickUp. Designed with agents in mind from the ground up, powered by llmist. Supports MCP, skills, Sourcegraph, Google Search, Exa and more. Model-provider agnostic. Agents run ucho's own CLI within E2B sandboxes. I'm using it to organize my daily work across clients and projects.
In an effort to learn how drones work, I decided to build one from scratch, and write a custom flight controller firmware running on an ESP32 platform. Had to learn soldering and 3D printing (LLMs + OpenSCAD made it a whole lot easier). Using etherlink for telemetry and ground station comms on a separate ESP32-C3.
An effort to extract higher-level understanding of large codebases with AST and LLMs to automatically generate and expose top-down view of the source code.
Squint performs static analysis of existing codebases (using tree-sitter) and adds a semantic, LLM-provided understanding layer. Annotates every symbol and relationship, generates architectural modules, detects cross-module interactions, traces user-journey flows and groups flows into product-level features. Agent-first CLI allows agents to drill down from overview to a single symbol without prior knowledge and laborious find & grep sessions. Built-in web visualization of the codebase. Still quite experimental.
In my consulting work, before introducing teams to Agentic Engineering, I like to understand their current process and bottlenecks. This simple tool syncs data from GitHub and JIRA into a local SQLite database and exposes standard metrics like flow efficiency, per-cause gate breakdown, PR cycle time, review health, sprint say-do ratio, and others via CLI.
Agents then use it to configure and query the metrics, as well as include team input to add more context beyond just raw numbers so that I can identify issues and how they can be solved.
I got interested in SysML v2 as a potential candidate to describe complex systems with, prior to my work on squint. I wanted to experiment with building SysML v2-based documentation with LLMs, but there was no parser readily available, so I built my own CLI allowing the agents to deterministically read and manipulate the format and deterministically validate complex multi-file markups.
Parses both KerML and SysML v2 layers, validates references and types, runs pattern queries, and edits models in place. C11 with a PackCC-generated PEG parser, arena allocator, Clang-style diagnostics.
Browser automation gadgets for llmist agents, built on the Camoufox anti-detect browser instead of vanilla Playwright — so the agent stops getting fingerprinted off the open web. Ships as a BrowseWeb subagent: one line of config and any llmist agent grows a pair of hands. Per-profile dev/research/production presets with documented memory tradeoffs for running fleets.
Lightweight binary protocol for ESP32 telemetry over BLE (Nordic UART Service) or serial UART. 4-byte framed, CRC-8, 250 B payload, zero-copy callbacks, transport-agnostic core — ships as three ESP-IDF components so a project pulls in only the transport it needs.
Decrypts Victron SmartSolar BLE advertisements (no GATT connection, no pairing) into both a CSV log and a Prometheus endpoint scraped by Grafana Alloy into Grafana Cloud. Solar W, battery V/A, daily yield, charge state — all labeled per device. Lives at my off-grid shipping container cabin.
Agents meet versioned rich text. Spine was an experimental project that combined the lexical.dev stack with llmist, enabling agents to work on a versioned set of rich-text documents along with the user, performing various tasks like research and fact finding. User could discuss with the agent selected portions of a document with a full change/intent audit trail.
Event-sourced, fully async backend on Node + Hono. Append-only event store, optimistic concurrency, command bus, Postgres projections, SSE realtime, version-history diffing.
AU enabled the agent to run through the codebase to extract higher-level understanding of it and embed it in SysML v2 definitions via my CLI. Turned out to be not deterministic and not efficient enough. I continued my experiments in the area with squint, which starts from the AST and ultimately drops the SysML support for a simpler structure.
One of my dogs (Jack Russell Terrier) loves chasing his ball and when he lost his sight, I experimented with putting an ESP32 into his ball and collar, along with some motion sensors and piezo to make finding it easier for him. Next, I played with Time-of-Flight sensors and vibration motors so his collar could warn him of incoming obstacles. Fortunately Jack got so good at using his remaining senses, that I abandoned the project unfinished, yet happy.
I led the R&D and implementation of an early version of shotgun.sh.
Terminal CLI for creating comprehensive and consistent specifications of new projects and features for later development with coding agents. Python event-driven server runs the agent runtime and event bus; a React + Ink TypeScript CLI and SDK talk to it over a versioned WebSocket protocol. SQLite-backed persistence for sessions, agent states, and spec artifacts. Composable agents built from modular memory / tool-use / LLM behaviors. Indexed arbitrary codebases with tree-sitter into a Kuzu-based graph database. Launched on Product Hunt September 2025. Since rewritten and released as open source.
Proofs was an early agentic platform, including harness, sandbox execution infrastructure and tooling enabling end-to-end proof-of-concept builds for sales-engineering teams with a particular focus on headless content management and e-commerce platforms.
Designed and developed the prototype of a coding agent in late 2023 to raise pre-seed on top of and hire an amazing team to build it.
Enter a prompt, get frontend and backend designed and implemented from scratch in a cloud-hosted Docker container, tested, pushed and deployed. Also included advanced website analysis (“copy that website to my headless CMS in a nice structure through the API and pixel perfect FE implementation in my framework - I will be back in an hour!”). Leveraged GitHub code search and other sources of information to have the agent succeed with unfamiliar stacks and technologies, as well as learn from experience (i.e. failures). Piping browser events and network traffic back to the testing agents for fun and profit. Agentic Execution Environments (boot them a container and they will come) with gRPC-based access to a tmux server for them to code in. Semi-deterministic Agent Execution with pre-created plans including pre-defined points of flexibility so the agents could take on really complicated projects effectively. Like programs for agents. Fits on a floppy.
A Ruby REPL coding agent from 2023 — point it at a repo, ask in plain English, it parses the model's response into edit actions and applies them, with git reset --hard HEAD~1 as the undo button. The shape that Claude Code, Aider and friends arrived at a year later.
A Ruby framework for composing LLM apps as ordered components — input parser, context retrievers, prompt, model — all sharing one execution context. Shipped a CLI and a Sinatra server in 2023, before LangChain ate the world. Same shared-context mental model I keep finding in every serious agent harness since.
A pip-installable CLI that runs one prompt against OpenAI, Anthropic, Vertex AI and HuggingFace in parallel, with Jinja templating, a watch mode for live iteration, and shebang-executable prompt files. Shipped in mid-2023, before "prompt engineering" had a Wikipedia page or LLM gateways were a category.
Type a root query, OpenAI spiders a graph of related concepts, ReactFlow renders the result with drill-down tooltips. A 2023 sketch of LLM-as-graph-explorer UX before the category had a name.
Whisper, ChatGPT, and ElevenLabs walked into a bar. A Ruby voice assistant that records via sox, transcribes locally with whisper.cpp (or the OpenAI API), pipes the text to ChatGPT, and speaks the reply back through ElevenLabs — with a naive prompt-stitched memory across turns. A couple of hours of hacking in early 2023; the loop still feels right.