OEB Advisory

An AI financial advisor that shows its work, bilingual, grounded in your real numbers, never a black box

Comment

Inspiration

I have worked with businesses across different sectors, some just starting, some with years of experience, and the same pain keeps showing up. Business owners don’t understand their finances. And even when they need clear numbers for a loan, for a decision, for planning, getting there is either time-consuming to figure out themselves or means hiring someone and just trusting the person is doing a good job.

For cases from my own work made this impossible to ignore:

A contractor who couldn’t see which jobs were actually making money. He left his brother’s company to start his own, build a strong client base through referrals and old connections, and position himself for high-end renovations. He invested in mentors at $7K a month, opened an office, and hired an assistant. But his sales weren’t growing at the same pace as his expenses, and he was miscalculating material costs on estimates. Some jobs came in negative. He ended up depending on the next job just to cover cash flow. When one deal fell through after three months of invested time and money, he had to let his team go and cancel everything. He was failing because nobody was watching whether individual jobs were actually profitable while he scaled.

A business owner who had the reports but still couldn't say if he was profitable. They have dashboards, they have processes, and they still only have a vague idea of their actual burn rate and profitability. Having reports and understanding what they mean turned out to be two completely different skills.

A company that couldn't answer a basic question because its own books were wrong. Their CPA made an entry error that duplicated income. Now they need to know exactly how much the company owes a former partner, and the number on the reports is off. The only way to get the real answer is going transaction by transaction, by hand.

A recruiter who couldn't plan around his own cash cycle. He charges $20K per placement and spends it the moment it lands, with no visibility into when the next one will come. He once turned down a stable $5K monthly retainer because, in his head, one $20K placement was worth more, without ever actually calculating what four months of nothing cost him. Five years in, still running it as a side hustle, still in the same cycle

Four different businesses, four different failure modes, one root cause: nobody had real-time understanding of their own numbers.

What it does

OEB Advisory gives business owners an active role in their own finances. You ask a question in plain language, in English or Spanish, and the advisor queries your real financial data, not a guess, not a generic answer- and tells you what's actually happening. Every answer comes with the exact SQL query it ran to get there, so you're never trusting a black box. You can verify it.

This does not replace a CPA, an accountant, or a bookkeeper. It closes the gap between what they tell you and what you actually understand, so you're a participant in those conversations instead of someone nodding along and hoping it's right.

Ask a factual question, "What was my cash flow last month?", and you get a direct answer pulled from real transactions. Ask a decision question, "Can I afford to hire a $7K a month coach?", and the advisor pulls your actual income and burn history, states what it shows, and lays out the tradeoff, without deciding for you.

Because business runs in more than one language, OEB Advisory reasons natively in English and Spanish. Not a translation layer bolted on top. If you've ever used a tool where the Spanish feels slightly off, slightly robotic, you know what that does to trust. In Mexico, we call that kind of thing "version pirata", the version people buy because they can't afford the original, but you can feel it's the fake one, in the materials, in the quality. It's the same with language. A weird translation loses trust immediately. We built it for zero gap between the two.

The business

OEB Advisory runs on a monthly subscription, priced by business complexity; a solo operation pays less than a business juggling multiple entities or higher transaction volume. That is the core recurring revenue.

The go-to-market isn't a cold start. The four cases above aren't hypothetical personas; they're real client situations from my own fractional bookkeeping and operations work, and those same businesses are the first target customers. OEB Advisory is the always-on version of advice I already sell by the hour, built on top of the client relationships and referral network that already exist through that work. The acquisition channel is already there.

As the product matures past answering questions into the proactive reporting and forecasting described below, that becomes the natural premium tier, expansion revenue from customers who are already getting value from the core product.

This project is pre-revenue during the hackathon window. What it isn't is untested. It's built directly from four real, named client situations, for a customer base I already work with and understand, which is a different starting point than guessing at a market.

How we built it

OEB Advisory runs on Databricks Apps, using Gemini for reasoning and tool calling, Genie to generate and execute real SQL against a Delta Lake financial schema, and Lakebase (Postgres) for chat history and session state. When the advisor needs real numbers, it doesn't guess. It calls a tool that hands the question to Genie, which writes the actual SQL, runs it against the warehouse, and returns both an answer and the query itself, which we surface in the UI so every number is checkable.

Financial accuracy isn't a nice-to-have for a tool like this; it's the whole point, and I've seen firsthand in past production work how answer accuracy gets measured and rated in systems like this. That experience shaped the core design decision here: show the work, every time.

Challenges we ran into

Building this meant working at the edge of a genuinely new stack, Databricks Apps, Gemini's current infrastructure, and Google's evolving authentication and billing model, most of it barely documented anywhere. Over the course of this build I hit five separate authentication and billing walls across two platforms: a corrupted API key format, an undocumented shift in how Google issues API credentials, a Databricks secret that only refreshes on a full deploy and not a restart, an organization level security policy inherited from an account I didn't know existed, and a Gemini specific prepay billing layer that is entirely separate from standard Google Cloud credits. Every one of those had to be diagnosed from scratch, with no clear documentation pointing at the actual cause.

Working through each of those meant reading raw error messages, testing one assumption at a time, and tracing every failure back to its actual cause instead of guessing. It ended up mirroring the product itself: the value was never in the first answer on the surface; it was in figuring out what was really going on underneath.

Accomplishments that we're proud of

There aren't many tools built with real fluency for Spanish-speaking business owners. Most were built in English first, and the Spanish feels translated rather than native. Getting the tone right in both languages, not just the words, was a deliberate design choice from the start.

This also isn't a demo built to look good in a five-minute walkthrough. It's built to save real hours and let business owners make decisions from an informed place instead of a guess. For someone like the contractor who scaled past his own visibility, or the recruiter who turned down stability without knowing the real cost, a tool like this changes the decision itself, not just the reporting around it.

What we learned

The technology (a capable model, tool calling, a real data warehouse) was never the hard part. The hard part was the infrastructure holding it all together, and the discipline to keep every answer accountable to real data instead of letting the model estimate. We also learned, the hard way, that a business idea can look "simple" to describe and still represent something genuinely difficult and valuable to build. The gap between "I built an app that answers questions" and "I built a system that queries real data, shows its work, and reasons natively in two languages" is enormous, even if it doesn't always sound that way out loud.

What's next for OEB Advisory

Right now, OEB Advisory answers questions when asked. The natural next step is proactive reporting: automated quarterly, half-year, and end-of-year summaries, revenue by project or client, expense breakdowns by category, and surfaced insights the business owner didn't think to ask for, all generated from the same grounded data pipeline already in place. Beyond looking back, that same data foundation opens the door to forecasting, helping a business owner see where they're headed before the numbers force the question, not just after. That shift, from answering questions to actively flagging what matters, is what turns this from a tool you open when you're worried into an advisor that's always watching.

The current build uses Databricks native authentication, appropriate for this stage of the build. A production version would add a lightweight customer-facing login layer in front of it, so clients access OEB Advisory directly with their own email, never touching Databricks at all. The backend stays the same, just invisible to the end user.

Built With

Share this project:

Updates