SellAnyBiz.ai

The Challenge

Buying or selling a business is one of the largest transactions most people ever make, and it runs on the least infrastructure. Listings are scattered across brokers and classifieds. Asking prices are guesses nobody can check. Buyers spend weeks on deals that were never a fit, sellers wait months for a serious enquiry, and brokers do the same manual matching by hand every time. Then, at the end, two strangers have to move a large amount of money between them and trust that it lands. Every one of those steps is work a person currently does because no system does it for them.

The Solution

Bun as both runtime and package manager, running Next.js 16 and React 19 through a custom server so WebSockets sit alongside the app. Supabase provides Postgres, auth and storage, with pgvector behind semantic search. Redis and BullMQ drive nine families of background queues. The API is ElysiaJS mounted behind a single catch-all route, which keeps types flowing end to end from the database to the client. On the front end TanStack Query owns server state and Zustand owns client state, so the two never blur together. Roughly 60 page routes sit on top of about 240 backend route modules.

Agentic, not a chatbot bolted on

Most products add AI as a support widget in the corner. Here it holds tools. There are 67 tool definitions across 24 agent prompt modules, and they aren't all read-only: the model updates companies, writes financials, adds documents to a data room, grants and revokes listing access, and drafts outreach. Which tools a model can even see is filtered by the acting user's permission tier, and every write is re-checked server-side, because a filtered tool list shapes behaviour but a route is what actually guards the data.

The loop behind it is built to survive the real world. It runs on a queue, snapshots message state at every step boundary, resumes from that snapshot if a job stalls, and can be cancelled mid-run. It advances one step at a time on purpose, so a provider outage replays a single step instead of the whole conversation, with a circuit breaker and a step budget that raises an explicit error instead of quietly returning truncated work. One agent can invoke another: company research runs as a sub-agent that can take minutes and redacts personal contact details from what it hands back.

Where the AI ends and the maths begins

Valuation is deliberately not an AI guess. Underneath sits a deterministic engine implementing eight methods, including discounted cash flow with projected years, discount rate and terminal growth, and comparable sales, deriving SDE and EBITDA from raw financial years with different addback sets per method. It recomputes authoritatively on the server and writes a new versioned row only when the headline number changes. The AI sits above that: it gathers inputs, reads uploaded profit-and-loss and tax documents, and produces a ranged indicative report carrying an explicit disclaimer that it isn't a certified valuation. Numbers a seller will price against shouldn't come out of a language model.

Letting people ask before they sign up

The homepage opens on a chat box, not a signup wall. A visitor describes what they're looking for and gets real listings back before creating an account. That shapes the architecture, because the AI layer has to work for anonymous visitors and authenticated users alike, with conversation state that survives the transition when someone does sign up.

Three audiences, one platform

Buyers, sellers and brokers want different things from the same listing data. Buyers search and evaluate. Sellers build a listing, price it and reach real buyers. Brokers run mandates across a portfolio, with commission and success fees calculated by country tier and written into e-signed agreements. Keeping those flows distinct while they share one data model is the recurring design question.

Keeping it maintainable at this size

The codebase splits into client, server and shared, and those boundaries are enforced instead of merely suggested. Client and server each import from shared and never from each other, and shared stays pure with no React and no server dependencies in it. Features live in self-contained folders, and one feature reaches another only through its published api surface, never by reaching into its internals. On a codebase this size that discipline is the difference between adding a feature and excavating one.

Where it's going

The platform handles the deal up to the point money moves: mandates, fee calculation, e-signature, data room and diligence. Closing that last step, so a transaction can complete safely inside the platform, is roadmap and not yet built. The wider goal is that the agent handles progressively more of the work at each stage, which is why so much of the engineering effort goes into giving it real tools and real guardrails instead of a better prompt.

The codebase is private, so there's no repository link, but the platform is public and you can try the AI search on the homepage without an account.

Key Features

• ▸Agentic AI with a real tool surface: the model reads platform data and writes to it, gated by the acting user's permissions
• ▸A nested sub-agent for long-running company research, plus separate agents for listing building, valuation, outreach and data room work
• ▸Durable multi-step agent loop that snapshots state, resumes after a stall and fails over between providers mid-run
• ▸Conversational search on the public homepage, with AI turning plain language into real listing filters
• ▸Eight-method valuation engine covering SDE, EBITDA, revenue multiples, asset-based, cap rate, DCF and comparable sales
• ▸AI reviewer that gates listing publication and fails closed, resetting to draft instead of auto-publishing
• ▸Data room with NDA gating, folder structure, audit logging and an AI copilot that reads uploaded financials
• ▸Deal and mandate management with automated broker commission and success-fee calculation
• ▸Outreach sequences with AI reply classification and suppression handling
• ▸Identity verification, real-time messaging, notifications and multi-tenant workspaces