Same two products, three audiences. Pick the framing that fits how you actually think about software: a casual conversation, a business document, or an entrepreneur's pitch deck. The underlying tech is identical — only the words change.
You upload a short video clip — anything, a TikTok, a sunset, a basketball highlight. Cortex shows you a 3D brain that lights up in the regions that your brain (well, an average human brain) would use to process it. Then four different "AI critics" describe what just happened in their own voices: a chatty ISU freshman, a Northwestern neurologist, a WBEZ reporter, and a Google ML engineer.
Picture a movie theater with 20,484 seats, where each seat is a tiny patch of your cortex. The movie is whatever you uploaded. Some seats lean forward when faces appear, others when music plays, others when something moves quickly. Cortex shows you which seats lean forward, and four different people in the back row whisper their take on the show.
Mercury is an AI assistant that lives on Soumit's own RTX 5090 in
Chicago, Illinois — not in some data center in Virginia. You can talk to
it three ways today: through a terminal, through a Discord bot called
@abmsnowy, or through a phone-friendly web page (over a
private VPN). Same agent, same memory, every door.
The point is: nothing leaves the building. Your messages don't get sent to OpenAI or Google or Anthropic. The model that answers you is Gemma 4, running locally on the same desktop that runs Cortex. If the local machine is busy, the request quietly hops to a MacBook on the same VPN. If both die, it falls back to OpenRouter's free tier — still no per-token bill.
Cortex is a research-software platform that converts arbitrary media stimuli (video, audio, image, text) into per-vertex BOLD-signal predictions across the 20,484-vertex fsaverage5 cortical surface, paired with population-targeted natural-language interpretations at four reading levels.
Use cases under exploration: research-grade fMRI explainer for undergraduate teaching, clinical-communication training (translating neurology jargon for patients), and content-perception analytics for educational media.
Mercury is a fork of NousResearch's Hermes Agent framework, configured for single-tenant deployment on consumer GPU hardware. It exposes one agent instance through three transport surfaces with shared session state and memory backed to a local SQLite database.
mercury command), Discord bot (@abmsnowy), web dashboard at :9119 over TailscaleTarget buyer: small teams with regulated data (legal, medical, financial) who need agentic AI but can't send conversations to a hyperscaler model API.
Cortex is the proof of a thesis: research-grade neuroscience demos can run on consumer hardware for the cost of a coffee. Every public scan on the live demo costs about $0.011 in electricity on the 5090. At ten scans a day, the entire brain-analysis platform runs on $3/month of Chicago electricity. No AWS bill, no OpenAI invoice, no per-token math.
Why this matters as a small business:
Mercury is the local-first answer to the hyperscaler-API agent problem. Most agent products today route through OpenAI / Anthropic / Google Cloud and pass through every per-token fee to the customer. Mercury inverts that: a one-time hardware cost (~$2,500–3,000 for the RTX 5090) replaces five years of API bills for a single user.
Why this matters as a small business:
Mercury is the orchestrator; Cortex is the specialty tool. From inside Mercury (in any of its three surfaces) you can hand a media file to Cortex and get the brain analysis + four narrations back as a structured response. Same RTX 5090, same Tailscale network, same Ollama inference pool. From a small-business standpoint that means one piece of hardware monetizes two distinct demos and powers any future agent skill that wants to call into a brain-scan endpoint.