Cosmic AI
Retail investors go from research to a confirmed order in one conversation
For Cosmic AI we built MarketVerse, a research and trading app where opportunity, research and trading agents work behind one chat. The agents prepare every trade, and nothing reaches the broker until the investor confirms it.
Financial ServicesMulti-agent investing assistantBuilt from zero
The challenge
Retail investors worked across scattered data, news, filings and a separate broker, and users of the first release got lost among too many panels.
What we did
Specialist agents behind one chat, a separate execution layer that places an order only after the investor confirms it, and onboarding that sets each investor's risk profile.
The outcome
The app went live on Azure and was relaunched around one chat-led flow from research to execution, with every order confirmed by the investor.
- Agents
- Opportunities, research and trading, behind one chat
- Orders
- Prepared by agents, placed only after confirmation
- Broker
- Robinhood, connected only with the investor's consent
- Runs on
- Azure, with separate QA and production
The starting point
MarketVerse is Cosmic AI's market intelligence and trading app for retail investors, and we built it as the company's product team. It serves day traders, analysts and everyday investors with a market dashboard, company research and a chat over AI agents.
What had to change
Retail investors work across scattered data, news, filings and a separate broker, while institutions have tools that bring it all together. After launch, users got lost among too many panels, so the product had to become one chat-led flow from research to a trade.
What could not break
- No order reaches the broker without the investor's explicit confirmation
- The broker is connected only with the investor's consent
- The agents that analyze stay separate from the layer that executes
- Options data stays in QA, out of production, because of its licensing
- Every release passes a separate QA environment before production
The calls we made
We framed it as one conversation over specialist agents, with a hard line between the agents that think and the layer that trades.
- 01
Analysis and execution kept apart
An order touches a real brokerage account, so the agents only research and prepare it, and a separate execution layer places it after consent, with editable parameters and a mandatory confirmation.
- 02
Specialist agents behind one front agent
Finding opportunities, researching a company and preparing a trade are different jobs, so each got its own agent, and a conversational front agent routes between them inside one conversation. We tested that routing across edge cases before release.
- 03
Profile before proposals
Onboarding captures risk profile, time horizon and strategy focus, and recommendations start from that profile together with the investor's watchlist, portfolio and the opportunity agent's latest picks.
- 04
Chat at the center of the product
When live users got lost among the panels, we made the chat the one place to start: it opens automatically, offers suggested questions and voice input, and the app was redesigned mobile first.
Ruled outKeeping the panel heavy layout that users coming from trading apps found cluttered.
- 05
Execution in the broker investors already use
Ahead of the relaunch, we moved order execution and portfolio import from the older execution layer to the Robinhood SDK, the broker the relaunch targeted, and kept the agents out of the order path.
Ruled outCarrying the older execution layer into the relaunch.
How the system fits together
What we built
Specialist agents behind one chat
A front agent routes each question to the opportunity, research or trading agent across five flows: portfolio and macro impact to a trade, a daily opportunity to a deep dive and a risk-based trade, hedge fund activity, congressional trading, and technical analysis of volatile stocks.
Orders prepared, never placed alone
Investors connect Robinhood with explicit consent and import their holdings; the trading agent prepares market or limit orders with editable parameters, and each one waits for review and confirmation before it is placed and logged in the order history.
Onboarding that sets the profile
A seven step onboarding walks through market signals, the AI assistant, watchlist and portfolio, asset selection, and risk and strategy, so recommendations start from the investor's own profile.
Plans with agent access per tier
Free and premium plans set which agents each account can use, with a trial for new users, Stripe subscriptions and a guided flow for each account type.
A chat-led, mobile first relaunch
The chat-centered redesign opens the chat automatically, adds suggested questions and voice input, and reworks the app for mobile first use.
How it works
Connect
The investor finishes onboarding with a risk profile and strategy, then consents to link the broker and import holdings.
Ask
A question in chat goes to the front agent, which hands it to the opportunity, research or trading agent.
Research
The research and opportunity agents weigh the portfolio against macro trends, news, filings and market signals.
Prepare
The trading agent drafts a market or limit order based on the investor's risk profile, with every parameter editable.
Confirm
The investor reviews and confirms, and only then is the order placed with the broker and added to the order history.
A closer look
What changed
MarketVerse went live on Azure, and the relaunch moved investors into one chat-led flow from research to execution. Agents did the research and the preparation, and the investor kept the final say on every order.
Under the hood
- Datics role
- Product built by Datics
- Agents
- The chat and the hand-offs between agents run on LangGraph and CopilotKit with OpenAI GPT-4o, and Crew AI and Autogen were also part of the agent work.
- Execution boundary
- The Robinhood SDK handles holdings import and orders, behind explicit consent, editable parameters and a mandatory confirmation, with options data kept in QA because of licensing.
- Platform
- Azure VMs and app services, Azure SQL, Azure AI Search and Azure Functions for scheduled jobs, with Gunicorn behind Nginx and separate CI/CD for front end and back end.
- Accounts and plans
- Clerk handles passwordless sign-in with Apple, Google or an email link, and Stripe subscriptions set which agents each plan can use.
AI and models
Cloud and infrastructure
Apps and APIs
Also in the stack
How it shipped
Design and first build
Design came first, then a first build with research, news, filings, sentiment and knowledge graph APIs, watchlists, chat and a dashboard, with CI/CD and a QA environment on Azure.
Own domain and launch
The app moved to its own domain with a marketing site, and after load testing and server tuning it went live at a launch event.
Portfolio and macro
We added a portfolio tracker with institutional ownership, chart analysis, pricing and macro trend graphs.
Relaunch around chat and the broker
Execution moved to the Robinhood SDK, Stripe payments and a trial went live, and a revised chat agent shipped with price targets and macro trends in a chat-centered redesign.
Onboarding and plans
A guided onboarding, free and premium plans, hedge fund holdings data and validation sessions with users followed.
The venture's product team
We worked as the venture's product team, from design and agents to data, QA and releases, with product and design calls made in weekly syncs with the founder.
Changed while live, tested before release
We swapped the execution layer and recentered the app on chat while it already had users, and every change passed a separate QA environment with its own CI/CD before production.
What this unlocked
We then proposed a B2B version for investment firms: the same agents behind an authenticated API, deployed inside the firm's own Azure subscription with single sign-on, Key Vault for secrets, an admin kill switch and human in the loop settings.
Related work
Have a product with the same kind of constraints?
Tell us what already exists and what you want it to do. We will tell you what it takes.