dennis

webpage

Inspiration

Our friend recently had his final-round interview for a quant internship. Instead of just praying for an offer, he decided to sell his friends unemployment insurance. If he got rejected, his friends paid him $5k each. If he got hired, he paid them $5k each.

Bro literally created a financial derivative on his own life.

If he gets the offer, he's employed. If he gets rejected, his friends soften the blow. That's +EV, so a win-win (depending on how you look at it).

Naturally, we wondered: Why stop there? Why can't we short our friends' dating lives, exams, and internship applications?

Everyone in the group chat wants in, but nobody wants to write contracts, manage spreadsheets, or Venmo-request five people. But we realized something else. Every group chat has that one friend who does everything nobody else wants to do. The one who actually plans the trips, coordinates people and purchases, finds the deals, and makes sure things happen.

What if every group chat had that friend? The responsible mom friend AND the absolute menace?

That's how dennis was born.

What it does

dennis is an always-on social AI agent that turns everyday conversations into opportunities.

Unlike a traditional chatbot, you don't have to constantly prompt dennis. He understands what's happening in conversations you've opted into, recognizes opportunities, talks to other people's agents, and brings back something useful.

He's got two personalities:

The Mom Friend: Financially responsible, socially unstoppable.

1. Your agents find deals you didn't know existed.

Imagine you casually text your friend:

"bro i need a new monitor mine's cooked"

Meanwhile, someone in a completely different conversation mentions they're shopping for the exact same monitor. Neither of you knows the other is looking.

But your dennis and their dennis do.

The agents independently recognize your purchase intentions, discover the overlap, compare group-buying options, and come back with a proposal that could save you both money. No searching. No organizing a group order. No awkward "anyone else wanna buy a monitor?" message.

For our demo, dennis discovers a simulated bulk deal that saves two friends $45 each. Both independently approve joining the group purchase and authorizing their contributions.

2. The trips finally make it out of the group chat.

"we lowk should go to cancun"

Normally, that's where the plan ends.

dennis can turn these conversations into actual plans: researching destinations, comparing options, organizing potential costs, and helping friends coordinate who's doing what. He's the friend who messages each friend independently, does the research, drafts the itinerary, and remembers that somebody still hasn't paid up.

The Menace: Your life is now a prediction market.

3. ShortMe — Short yourself. Back your friends.

Your friend texts: "i'm so cooked for my meta technical tomorrow"

dennis says: "getting rejected sucks. you should hedge that risk. wanna short yourself?"

Suddenly, there's a private ShortMe market:

Will Alex pass his Meta technical?

Friends can BACK or FADE the prediction and bet on it. Alex can hedge against his own success. It works for internship interviews, exams, college milestones, and even your friends' romantic lives.

Everything happens within your friend group. The person the market is about must approve it before publication, and results are confirmed before settlement.

It's like Polymarket, but with the lives of your closest friends.

The real magic: Agent → Agent → User

The most exciting part of dennis isn't just what he does, but how he does it. AI chatbots wait for you to ask a question. dennis works in the background of authorized conversations.

Your personal agent understands your life. It coordinates with other people's agents using limited, permissioned information. Together, they discover opportunities neither user knew existed. Then dennis comes back to the humans with a proposal.

Our desktop dashboard makes this whole process visible with live ShortMe markets, group-buying opportunities, personal and shared agent activity, consent controls, and an iMessage-style conversation simulator.

How we built it

Underneath the fun, dennis is a network of collaborating agents with a real backend.

Photon — The group chat

Photon connects dennis to messaging through its Spectrum integration. dennis can receive messages, recognize social context, and send responses through connected conversations. It's what makes dennis feel like another member of your friend group instead of a separate app or separate chatbot.

Claude — The brain

We use Anthropic's Claude API to understand casual conversation, extract structured intentions, research possible plans, and generate responses in dennis's personality. But we don't let the model do everything. Deterministic backend code controls purchase matching, calculations, approvals, accounting, and payment authorization.

dennis can be a menace, but he still needs to be safe.

Capital One Nessie — The money

We developed and tested a Nessie banking adapter for synthetic accounts and transfers. During live API testing, Nessie created transaction records without the expected balance changes. Rather than pretend the transfers completed successfully, we built a clearly labeled simulated banking fallback for the working group-purchase demo.

The app itself:

Next.js + React: Our desktop dashboard, including markets, purchases, chat, and activity visualization.
TypeScript + Hono: Our API and multi-agent orchestration.
SQLite + Drizzle ORM: Persistent users, conversations, intentions, opportunities, approvals, and transactions.
Zod: Shared contracts and validation between the agents and backend.
Python evaluation tooling: Local tests of intent recognition, opportunity matching, and privacy behavior.

Challenges we ran into

1. Teaching AI to understand brainrot.

Real group chats don't sound like carefully crafted prompts. People say things like "im cooked," "am i chopped," or "we lowk need to lock in on iceland." dennis needs to distinguish meaningful intentions from jokes, hypotheticals, and random messages without turning every conversation into an unwanted financial proposal.

We also encountered cases where the model's structured JSON responses were silently truncated by small output budgets. Fixing those limits was essential for reliable extraction.

2. Making agents social without making them creepy.

For two agents to discover a shared purchasing opportunity, they need some information about their users. But we don't want someone's entire private conversation to be handed to another person's AI.

We built consent checks, private views, and minimal matching information so agents can discover compatible needs without broadcasting the original messages. Every group participant must opt in, and financial actions require explicit approval.

3. Building a real financial workflow, not a fake success animation.

Our Nessie integration exposed a problem: an accepted transaction API call didn't necessarily mean the account balances changed. We had to verify results, handle provider failures, and keep simulated transactions clearly distinguishable from live API outcomes.

4. Keeping four parts of dennis from fighting each other.

Messaging, multi-agent coordination, ShortMe, banking, and the frontend all needed to work together. Building them concurrently required strict shared interfaces, clear module ownership, and extensive integration testing.

Turns out coordinating AI coding agents is its own social optimization problem.

Accomplishments that we are proud of

We got independent agents to discover opportunities together.

Our purchasing flow starts with separate conversations, recognizes compatible purchase intentions, finds a potential shared deal, requests approvals from both users, and proceeds through simulated settlement. Nobody needs to explicitly ask dennis to match them with another buyer.

We made ShortMe more than just a joke.

The prediction-market feature includes subject consent, private market publication, virtual positions, zero-sum points accounting, outcome confirmation, dispute handling, and friend-group participation.

We made an unhinged social AI safer.

Approvals are single-use, time-limited, and tied to specific proposals. Our backend enforces consent and authorization independently of what the LLM says. We also built privacy-isolation tests, replay protection, and checks for accounting correctness.

And we brought dennis to life.

A desktop experience that brings together group chats, prediction markets, collaborative purchasing, and visible agent-to-agent activity. The goal was to build something that doesn't feel like another productivity tool.

It should feel like adding a very financially literate, slightly problematic friend to the group chat.

What we learned

The interesting part of AI isn't always having a smarter chatbot. It's what happens when personal agents start collaborating. One person's agent can understand their needs. Two agents can discover opportunities neither person knew existed.

But autonomy also creates responsibility. The hardest part was making sure he knew when to ask permission, what information he could share, and which decisions should remain entirely in human hands.

We learned that the best agent doesn't take control away from you.

Instead, it does the annoying work so you don't have to.

What's next for dennis

We want dennis to be a better mom friend AND a better menace.

We want dennis to become the personal AI friend who's always looking out for you and your social circle.

Next up:

Real-world group deals: Connect live merchant offers and make shared purchasing genuinely useful.
Trips that actually happen: Move beyond planning to approved bookings, shared itineraries, and coordinated expenses.
Smarter agent networks: Discover more opportunities involving moving, events, subscriptions, and everyday purchases.
Better evaluation: Integrate Chalk and expand testing of real-world agent behavior.
More ShortMe markets: More private friend-group markets, funnier outcomes, and leaderboards.
More reliable integrations: Complete live Photon validation and improve banking-provider verification and recovery.

We think the future of personal AI is social. Not one assistant answering one person's questions, but a network of personal agents collaborating to make everyone's lives easier, cheaper, and a lot more fun.

Good friends should be invested in your life.

dennis makes sure they can.