An experiment lab for applied AI
Speedy Labs builds custom software and AI tools that people rely on every day, and runs the slightly crazy experiments that might become the next ones. One lab, from first prototype to daily use.
Speedy, the lab's flagship workspace. Interface recreated with demo data.
The lab
Some ideas deserve years of careful engineering. Others deserve a weekend and a wild guess. We do both, and the best experiments graduate into products.
Products
Custom software for founders and teams, held to production standards: tested, documented, deployed with rollbacks, and used every day.
Experiments
Applied-AI bets that sound slightly unreasonable: software that upgrades itself, reading lips from silent video, agents trading real markets. Some fail. The good ones ship.
What we build
Assistants, coding agents and multi-agent systems connected to your tools, with model routing, cost controls and guardrails that hold up in daily use.
SwiftUI for Mac and iPhone, Kotlin for Android. Menu bar utilities, custom keyboards, widgets and background services that feel at home on the device.
Dictation, transcription and meeting notes that handle code-mixed speech such as Hinglish, learn your vocabulary and clean up the text for you.
Knowledge graphs, retrieval and integrations with Docs, Drive, Notion, Linear and Slack, so assistants work from what you actually know.
Self-deploying pipelines, test harnesses, browser automation and internal tools that take repetitive work off your team.
Non-custodial agents, trading infrastructure and smart contracts on Sui, Ethereum and Stellar, with risk limits enforced in code.
Flagship product · Speedy
In use dailySpeedy is a self-hosted AI workspace that runs on its owner's keys across Mac and iPhone. It brings together chat, coding agents, research, documents and memory, and it ships its own improvements.
Describe a change in chat and Speedy hands it to Codex or Cursor on a separate branch. Documentation, tests, type checks, a production build and a browser suite all run before it publishes, restarts and health-checks itself.
The Code tab runs real coding sessions on your own machine. Pick a folder, an agent and a permission mode, then watch every command and edit as it happens. The chat can also delegate work to an agent and report back.
Videos up to 64 GiB upload straight into a chat in resumable chunks and play inline. Speedy turns long, silent screen recordings into a visual transcript: what happens on screen, segment boundaries, on-screen text and key frames.
Studio creates and edits images, 3D models with Blender and CadQuery, and editable PowerPoints. Brain connects a personal knowledge graph of people, projects and plans, so every chat starts with context. Google Docs and Drive are one connector away.
OpenAI and Anthropic models are listed live from each provider, and AWS Bedrock adds Claude, gpt-oss, DeepSeek, Qwen and Llama on your own AWS account. Auto routing picks a model per task and shows the token usage it spends.
The same server powers both devices, so chats, projects, memory and coding sessions are identical everywhere. Scan a QR code to sign in the native iPhone app, then approve agents and pick up conversations on the move.
Products
Each began as a problem in our own day and grew into a tool. The same standard goes into every client project.
wispr-custom
·Custom dictation
In use dailyHold a key, speak, let go. Two speech models transcribe in parallel, a cleanup pass removes fillers and self-corrections while keeping technical terms, and the text lands wherever your cursor is. Built for code-mixed English and Hindi, and it learns your vocabulary as you correct it.
speedy-meetings
·Meeting intelligence
ExperimentalRecord locally, or send a self-hosted notetaker bot into Google Meet. Calls are transcribed with Hinglish support, cleaned without dropping a voice, summarised and synced to Speedy, with no third-party bot service.
MacPulse
·System monitor
ExperimentalA native dashboard reading temperatures, fans, power draw, memory pressure and swap straight from the hardware, with plain-English insights and one-click fixes. Its storage advisor finds what is safe to clear.
Brain
·Knowledge graph
In use dailyA Git-backed graph of projects, people, links, plans and ideas, captured by chat, voice or terminal and browsable on the web. A planning agent turns it into a day, and Speedy reads and writes it with revision checks so every conversation starts with context.
Experiments · Lab notebook
Every experiment starts with a question. Here is what we tried, and where each one stands.
Can we recover what people said from a video with no sound at all?
Visual speech recognition reads the speaker’s mouth frame by frame, then a context-aware language model reconstructs readable sentences from that noisy signal, using what else is visible on screen.
Heading into Speedy
Visual speechLLM reconstructionWhat if an app could take a feature request in chat and ship it safely, with nobody at a terminal?
A coding agent builds the change on an isolated branch. Docs, tests, types, a production build and a browser suite must pass before an atomic release switch, a health check and automatic rollback.
Lives in Speedy
Coding agentsRelease engineeringCan an AI agent manage real money without ever having the power to take it?
Language models plan; the allocation is computed and signed inside a trusted enclave; the chain verifies the signature, caps and receipts before funds move. Deployed on Sui mainnet across three lending protocols.
HexLiquid Yield
SuiMoveTEEUnder identical prompts and risk rules, which model actually trades best?
Models from several labs trade perpetual futures with stop losses enforced in code, every decision logged, and a realtime dashboard comparing their reasoning and results.
Does merging two speech models beat either one alone on code-mixed Hinglish?
Both models transcribe the same audio in parallel; a cleanup pass merges them, fixes self-corrections and keeps names and technical terms. It became the engine of our voice keyboard.
Lives in wispr-custom
SpeechHinglishHow quickly can one engine react to two of the fastest chains at once?
A Rust collector-strategy-executor pipeline watching DEX-to-DEX routes at 10 ms block times alongside exchange-versus-onchain and funding-rate opportunities.
arbx-dual
RustMEVApproach
You work directly with the people building your product. We use the same AI-assisted workflow we build for others: agents write code quickly, and nothing ships until it passes the checks.
We judge our work by whether people reach for it every morning, not by how it demos. Latency, empty states and edge cases get the same care as the headline feature.
We default to self-hosted, local-first designs. Credentials live in the Keychain, models run on your own accounts, and nothing leaves your machine without a reason.
Every change passes real checks: unit, type, browser and health checks. Rollbacks are designed in from the start, and the documentation ships with the code.
We move quickly to working software, then slow down on the details. Fewer features, finished properly, beat a long list of almost.
A conversation and a short written brief: the problem, the people, and what success looks like.
Working software in days, not slides, so decisions are made with something real in hand.
Frequent releases you can use, with tests, documentation and a clear changelog.
Source, docs and deployment in your own accounts, with ongoing support if you want it.
About
Speedy Labs is an umbrella for two things that usually live apart: dependable custom software for founders and teams, and applied-AI experiments run purely to see what is possible. Each makes the other better. Experiments keep the products ambitious, and the products keep the experiments honest.
Most of what you see on this page runs every day on our own machines, which is the fastest way we know to find what actually matters.
We are comfortable across the stack, from SwiftUI interfaces and Android keyboards to agent orchestration, speech pipelines and smart contracts, and we are happiest where those meet.
Stack we reach for
Contact
A product that needs building properly, or an experiment you are not sure is possible yet. Tell us what you have in mind. We read every message and reply personally.