Second Site

Inspiration

A business owner opening a second location usually starts with a listing site, sorts by rent, and tours the cheapest spaces. The lease is the smallest surprise. The real money goes into what the listing never mentions: whether the suite has plumbing where a treatment room needs a sink, whether the electrical service can carry the equipment, how many walls have to move, and what it costs to furnish the result.

That information exists, but it is scattered. Lease terms live in CoStar. The inside of the suite lives in a Matterport scan. The roof and the lot live in aerial imagery. The building's plumbing and electrical history lives in city permit records. Nobody joins them for a small business owner, so the decision gets made on rent and a walkthrough.

We heard this first-hand from Dr. Ajay Ohja, a teammate's godparent who owns Comprehensive Dental Clinic in Michigan. "It's really difficult to expand when having to manage another business," he told us. "Often so much time is lost on the interior and sourcing pieces. It's hard to be decisive with too many options and to see the full picture." For a dental office, every operatory needs water, drainage, suction and power, so what is behind the walls matters far more than the rent.

We built Second Site to be the team a small business owner doesn’t have. A realtor who looks past the rent to what the space will cost to renovate. A designer who lays out the rooms inside the actual floor plan. A buyer who turns that design into a furnishing list. One project, from the first listing to a space that is ready to open

What it does

Second Site takes a business owner from a business profile to a furnished floor plan, with every step writing to the same project record.

Profile. The owner describes the business: room program, opening capital, rent ceiling, comparison horizon. Our demo business is a facial studio with three treatment rooms, a reception and storage.
Property discovery. Second Site searches active Los Angeles lease spaces from the CoStar data kit, filtered by the profile's area need and rent ceiling.
A digital profile per property. Each candidate is enriched with the Matterport scan of the listed suite when one exists, Vexcel aerial imagery and detected roof elements, and Los Angeles building permit records matched to the exact address and unit.
Infrastructure feasibility. Each business requirement (plumbing, electrical, HVAC) is assessed against sourced evidence. A historical permit is treated as history, not as proof of spare capacity. What we cannot verify stays UNKNOWN and becomes a question for the broker.
Total cost ranking. Properties are ranked on rent, operating costs and expected build-out over the owner's horizon. In our fixture scenario, a higher-rent Hollywood suite beats a lower-rent Alhambra building over 36 months, because the cheaper building needs roughly $100k more build-out.
Approval-gated outreach. The assistant can prepare a broker call with questions drawn from the open unknowns, but it cannot contact anyone. It creates an approval request; only the owner can release it. The call flow escalates to the owner the moment a deposit, guarantee or commitment comes up.
Design Studio. Your AI interior designer. Second Site lays out your rooms inside the real, measured walls of the suite, then shapes the design around your needs and your aesthetic. You work with it like a designer: ask for a fourth treatment room, a bigger reception or a different finish, and it redraws the space. Think The Sims, but for the location you are about to lease.
Procurement. Everything you see staged in the design is something you can buy. Second Site turns the finished layout into a room-by-room shopping list of furniture, fixtures and finishes with a running total, so the space you designed is the space you open. What is live today: CoStar data kit listings, the Matterport Model API, the Vexcel API and LA permit records, plus the feasibility, finance, ranking, layout and bill-of-materials engines and the web app.

Owner messaging and broker calls: the owner links a Photon channel by texting the agent once, then asks questions, checks status and joins calls from their own phone. Broker calls run through Telnyx with an automation disclosure, questions grounded in the property's open unknowns, and a handoff to the owner by text. What the broker says comes back as evidence and re-ranks the shortlist.

How we built it

We split Second Site into four workstreams, one per teammate, and built them in parallel. Each of us worked with a coding agent, and Ira set up a shared message board on GitHub so one person's agent could hand work to another's without us stopping to sync.

Backend: Python 3.12, FastAPI, PostgreSQL, Redis, LangChain Deep Agents, LangGraph, Google OR-Tools, Shapely, plus a Rust port of the core engines. Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS, react-konva.

Data, agents and backend (Sal). Sal built the core that everything else runs on. It joins four sources on one address: CoStar lease listings, Matterport scans, Vexcel aerial imagery and Los Angeles building permits. A LangChain Deep Agents supervisor then runs the analysis with its subagents. The agents plan and explain, while the numbers come from deterministic engines that check each space against what the business needs, total rent, operating costs and renovation over the owner's time horizon in exact integer cents, and rank the options. Anything we cannot verify stays marked unknown and becomes a question for the broker. Every step is saved to an event log, so one new fact can update the cost and the ranking automatically. Sal also wrote a Rust port of the core and ranking engines, tested against the Python originals.

Conversations (Jasleen). Jasleen built how Second Site talks to people. The owner texts the assistant through Photon 2.0 to ask questions, check status and send design changes. Broker calls run through Telnyx Call Control, with questions drawn from each property's open unknowns. The assistant cannot contact anyone until the owner approves, and the approval is tied to the exact action so it cannot be reused. It hands the call to the owner as soon as a deposit, guarantee or commitment comes up. What the broker says is saved as evidence, which updates the cost and the ranking.

Design Studio (Ira). Ira built the part where the space comes together. A constraint solver (Google OR-Tools CP-SAT) lays out the owner's rooms inside the measured walls from the Matterport scan, and Shapely checks that every room fits and connects. The owner can ask for changes, such as a fourth treatment room, and get a new version without losing the last one. The finished layout becomes a room-by-room list of furniture, fixtures and finishes with a total.

Web app (Mihir). Mihir built what the owner sees. Six pages follow the journey: an overview of the project, a ranked list of properties, a property intelligence view with the lease terms, 3D tour, aerial image and permit history, the Design Studio, the purchase list, and a live activity feed where the owner approves actions and follows calls.

The agents. One supervisor coordinates 14 specialists in four groups:

Group	Agent	What it does
Coordination	Expansion Supervisor	Plans the work, delegates to the specialists and explains results to the owner
Find and evaluate	Business Requirements	Captures the rooms, budget and constraints of the business
Find and evaluate	Property Scout	Searches lease listings and resolves each address
Find and evaluate	Spatial Intelligence	Turns scans and aerial imagery into measured geometry and site observations
Find and evaluate	Infrastructure Feasibility	Checks plumbing, power and HVAC needs against the evidence
Find and evaluate	Financial Modeling	Totals rent, operating costs and renovation into a cost scenario
Find and evaluate	Property Ranking	Ranks the candidates and explains the tradeoffs
Find and evaluate	Evidence and Compliance	Confirms every claim is backed by a sourced record
Conversations	Outreach and Negotiation	Prepares broker outreach and waits for the owner's approval
Conversations	Voice and Handoff	Runs the broker call and brings the owner in when needed
Conversations	Photon Communications	Handles the owner's text channel
Design and open	Floor Plan Architect	Generates layouts inside the measured walls
Design and open	Interior Design	Applies finishes and style to the layout
Design and open	Procurement	Builds the room-by-room purchase list
Design and open	Logistics and Opening	Plans the timeline to opening day

How the work flows. The four pieces run as one pipeline on one shared project record:

The owner describes the business: rooms, budget, rent ceiling.
The backend finds candidate spaces, builds a profile for each, and ranks them on true cost.
Open questions go to the broker by phone once the owner approves, and the answers update the ranking on their own.
The owner picks a space, and the Design Studio lays it out and produces the purchase list.
Every step shows up live in the web app.

Sponsor tracks

CoStar Group: Inside and Out, AI for the Built World. Second Site uses all three CoStar Group layers together, for the same address, and each one does real work. CoStar's market data tells us what a space costs: we read active Los Angeles lease listings and turn quoted rates into monthly rent and operating costs. Matterport gives us the inside of the building: we pull the measured walls and openings of a suite and hand them to a solver that designs the owner's floor plan inside them. Vexcel gives us the view from above: aerial imagery and detected roof elements for each candidate site. The combination produces something no single layer can, a ranking of spaces by what they truly cost to open, and a layout drawn inside the real walls. It is agentic, not a chat wrapper over a property record: the agents reason about what a business needs, plan the analysis, and act by preparing broker outreach for the owner's approval. And it is spatial and visual by nature, because the product ends in a floor plan. It solves a real problem for a real person, a small business owner choosing a second location, and it is something that owner would want switched on Monday.

LangChain: Best use of deepagents. The heart of Second Site is a deepagent. Our Expansion Supervisor is built with create_deep_agent and delegates to four subagents, for requirements, feasibility, financial analysis and evidence, each given only the tools its role needs. We designed it around one rule: the model plans and explains, but it is never the source of a fact. Every tool is a deterministic service that returns record IDs, so each sentence the agent says can be traced to stored evidence. The single tool that reaches the outside world can only create an approval request for the owner. LangGraph checkpoints hold the agent's state, separate from the business data. The result is a deep agent an owner can trust with a decision as large as a lease.

Photon: Build a Messaging Agent using Photon 2.0. Second Site puts an agent where the owner already talks. Through Photon 2.0, the owner texts their expansion project from their own phone: they ask about their shortlist, check status, and send design changes. This is real human-in-the-loop collaboration. When a broker call reaches a question only the owner should answer, the agent texts them, and a one-word reply brings them into the live call or defers it. Context persists across channels, so what happens on a call shows up in the text thread and in the app. We built the safety and fallback behaviors first: the agent never answers an unknown number, never accepts an approval by text, and tells the truth about whether a message was delivered. It is hybrid intelligence in a practical form: the agent carries the workload, and the human makes the decisions that matter. A business owner could run it tomorrow with nothing but their phone.

Capital One: Best Usage of Nessie. Opening a second location is a financial decision before it is anything else, so we brought the bank into the product. Second Site uses the Nessie API in procurement, the last step of the journey. When the Design Studio turns a layout into a purchase list, Second Site checks that list against the business's account balance through Nessie, so the owner sees what they can afford before ordering, and records furnishing purchases as transactions against the account. It reimagines a small corner of the banking experience: the owner's bank account is part of the expansion plan, sitting next to the rent, the renovation estimate and the furniture total, which helps owners expand without overextending.

Amazon: Best Practical AI Innovation. Second Site is AI that solves a real problem for real people: small business owners who want to grow and do not have the time. It shows agentic behavior from end to end. The agents reason from the owner's own requirements, plan a multi-step analysis across listings, scans, imagery and permits, and take action by preparing broker calls and building a design. It adapts: when a broker shares a new fact, the cost and the ranking update, and when the owner asks for a design change, the layout is redrawn. We worked to make a powerful system feel simple, with six pages, plain-language explanations, and clear labels on which numbers are quoted and which are estimates. The idea is original, the need is real, the engineering is tested, and the experience is one an owner would come back to.

Peri: Forward Deployed Challenge. We built Second Site the forward deployed way, by talking to a real business first. We consulted Dr. Ajay Ohja, dentist and owner of Comprehensive Dental Clinic in Michigan. He told us: "It's really difficult to expand when having to manage another business. Even the small things that change time are very important to people like us." He also told us where the hours go: "Often so much time is lost on the interior and sourcing pieces. It's hard to be decisive with too many options and to see the full picture." Those two problems are the product. Second Site's agents handle the search and the broker conversations, so the owner keeps running the first location, and the Design Studio and purchase list give them one complete picture in place of endless options. His reaction to agents handling management and communication: "a great idea." It addresses a real-world need he described in his own words, it asks nothing of him but a description of his practice, and it runs end to end from property search to a furnished floor plan.

Challenges we ran into

Putting four parallel builds together. We worked on separate branches for most of the weekend, so integration was the moment things could have fallen apart. Conflicts did come up when we merged. Two things kept them small and quick to fix. First, we agreed on shared contracts at the start, one definition of every piece of data that passes between the backend, the agents and the web app, so each of us was building against the same shapes. Second, our coding agents talked to each other on a shared message board, so when one teammate's agent changed something another depended on, the other agent heard about it right away. When we merged, the pieces fit, and the conflicts we hit were resolved in minutes instead of hours.
The data was not what we planned for. We expected to query CoStar through a live API and designed around that. What we received was a data kit: a snapshot of listing files with documentation. We rebuilt our property search to read the kit directly and learned its rules for turning a quoted rate into monthly rent. It also taught us to label where every number comes from, which became one of the product's strengths.
Real floors are not rectangles. Our layout solver was written for simple rectangular spaces, but a Matterport scan of a real suite has angled walls, columns and odd corners. We had to find the largest usable rectangle inside each measured floor, design the layout there, and then map the rooms back onto the real floor plan. Getting a layout to sit correctly inside the actual walls of a real listed suite was a hard technical moment of the weekend.

Accomplishments that we're proud of
Using all three CoStar Group data layers together. CoStar listings, Matterport scans and Vexcel aerial imagery all describe the same building, and we got them working as one. For a single address, Second Site knows what the space costs, what it looks like inside down to the measured walls, and what it looks like from above. We added city permit history on top, so an owner sees one complete picture of a property in place of four separate searches.
A web app that makes a complicated decision feel simple. Opening a second location involves leases, renovation costs, permits, broker calls and furniture. Our interface walks an owner through all of it in six clear pages, from a ranked list of spaces to a live feed where they can follow and approve everything the agents do. Someone who has never read a commercial lease can use it.
A Sims-style Design Studio for a real space. The owner sees their own rooms laid out inside the actual walls of a real listed suite, taken from its Matterport scan, and can ask for changes and watch the design update. Seeing a business take shape inside a space before signing the lease is the moment the whole project comes together.

What's next for Second Site

A bigger catalog. Scrape more stores and more products, so the Design Studio can furnish a space from a much wider range of real furniture, fixtures and finishes.
A decorating agent that learns. Have the interior design agent learn from each owner's feedback and past choices, so its suggestions get closer to their taste with every revision.
Beyond Los Angeles. Bring Second Site to more cities, starting with Houston, where our property search already supports listings.