Vibe data science.#
Train deep learning models entirely from chat.
Three weeks. 169 training runs. 46 architecture versions. One developer, one AI — his only interface was chat.
Get started free → Read the story →
Free for personal & non-commercial use. Commercial licensing →
Chat is your only interface#
No terminal. No commands. No context switching between tools.
Tool calls stream inline as they happen — every file read, every command run, every decision is visible as it unfolds. You're not talking to a black box; you're watching your AI work in real time.
The same interface that answers a quick question can run an unattended overnight session.
Works while you sleep#
Configure an agent with a heartbeat interval and walk away. It wakes up, does its work, and posts results to your channel.
The seed lottery that cracked our volatility model ran twelve hours overnight — unattended, autonomous, supervised from a phone.
Supervise from anywhere#
A real mobile interface — purpose-built for checking in, not just a scaled-down web page.
Check progress, redirect your agent, review results. Real projects get finished because supervision doesn't require a desk.
This is 6am. Run 169 had just hit 0.6033 val_AUC — the first time we'd broken 0.60. The AI noticed before Ben did.
Channels — a new model for human-AI collaboration#
Persistent, named workspaces for ongoing projects. Each channel has its own memory, history, and skills. Context compounds over days and weeks. You pick up exactly where you left off.
This conversation is happening in one right now.
Your context belongs to you.
Switch between Claude, ChatGPT, Gemini, or a local model without losing history or prior work. Your session lives in LIT — not in any AI provider's servers.
Remote web & mobile access.
A real mobile interface. Check in from anywhere, redirect your agent, review overnight results. Real projects happen because supervision doesn't require a desk.
Your hardware, your data, your models.
Everything runs on your machine. Local GPU support and Ollama built in. Sensitive data never leaves your environment.
Multiple simultaneous agents.
Run Claude on one workstream, Gemini on another. Compare results, route tasks by model strength, use a cheap local model for mechanical work and a frontier model for hard reasoning.
10 years of ML guardrails.
Temporal integrity enforced. Lookahead bias structurally impossible. Checkpointing that just works. The platform encodes a decade of hard-won opinions about what goes wrong in real ML projects.
Full visibility.
Every tool call streams inline. Every file read, command run, decision — observable and auditable. Not a black box.
Professional services#
11+ years building production AI systems in healthcare, finance, and enterprise.
Fractional CTO · Strategic AI leadership · $10,000/month · Learn more →
Custom AI development · Deep learning, computer vision, production systems · $375/hour · Get in touch →
Technology licensing · Patents in ML training optimization and neural network systems · Enterprise →
Platform licensing · Agentic infrastructure for teams building AI products · Inquire →