HEAVENPATH

Inspiration

Most people who want to start a business aren't business majors. They're someone who perfected a spicy-chip recipe, or wants to open a coffee shop in their hometown. To get started they have to research the market, the location, the money and the legal rules, and those answers live in a dozen different places in jargon most of us don't speak. And the AI tools that could help have a bad habit: they confidently invent competitors, rents and regulations.

We wanted the opposite. HEAVENPATH takes one sentence ("coffee shop in Cumming, GA") and does the research with real data. It explains everything in plain English and tells you what it does not know.

What it does

One required answer. You type your idea. City, savings, loan preference and priorities are all optional.
A first report in about a minute. A short summary, your next three steps, customer groups to test, real nearby areas, and a money picture.
Real data, labeled. Competitors and communities come from OpenStreetMap, population and income from U.S. Census data, and typical rents from published broker reports Q2 2026, all with their source. Every fact is badged as real, yours, calculated, AI, or made-up example.
Ask in plain words. "Show me cheaper locations", "what if I start online?" or "I don't know my budget yet" change the real plan, with an Undo.
A staged action plan. Each step says what to do, why, how, what to save, and who to contact.
Honest about limits. It says it has no live property-listings feed. It warns you when the space you pick breaks your own budget. For non-U.S. locations it skips U.S. data instead of faking it.
Many ideas, safely. Create, switch, and delete ideas, with a confirmation that names the exact one.

How we built it

Next.js 16, TypeScript and Tailwind, deployed on Vercel. The AI is NVIDIA NIM (nemotron-3-super-120b-a12b), called only from the server.

The key design rule: AI interprets, code calculates. Every number comes from deterministic functions:

$$\text{funding gap} = \text{startup cost} - \text{owner cash}$$

$$S^{*} = \frac{F}{1 - v} \qquad (\text{break-even sales, fixed costs } F,\ \text{variable-cost share } v)$$

$$M = P\,\frac{r}{1-(1+r)^{-n}} \qquad (\text{illustrative loan payment})$$

Places are scored transparently, and missing data is never rewarded:

$$\text{fit} = 10 \cdot \frac{\sum_{i \in K} w_i\, s_i}{\sum_{i} w_i}$$

where $K$ is only the factors we actually know (unknown factors count as 0), and we also show the score's best possible range if the unknowns turn out well.

Other pieces:

Structured, validated AI output. Citations the model returns are checked against the evidence it was given, and invented ones are dropped.
A rule-based fallback for every AI feature, clearly labeled "not AI".
A typed refinement engine. Plain requests become a short list of allowed actions that edit the real project. The AI can only choose from that fixed list.
297 automated tests, plus end-to-end browser tests on the live site.

Challenges we ran into

A wrong-city bug. Our first census query silently returned Sidon, Mississippi's income for Austin, because the API quietly ignored one filter. We fixed the query, then added a guard that rejects any data row for a different place.
Unreliable free map servers. Public OpenStreetMap servers failed about half the time. We added several mirrors with retries, a second search method, and a dated saved copy of real data as a last resort. If all of that fails, the app says so instead of inventing anything.
"Real property listings." No free commercial-listings API exists. Rather than fake it, we say so, link to the listing sites, score spaces users add, and use a clearly labeled market-average stand-in.
Small towns. Our demo city, Cumming, GA, initially returned zero neighborhoods because small towns map their areas as "villages" and "hamlets". We built a wider fallback search.
The world isn't the U.S. We caught that "Mumbai" was silently treated as a U.S. city. Now census data, U.S. rents, and U.S. agencies and loan programs are never applied abroad.
Reasoning-model quirks. The AI sometimes spent its whole token budget "thinking" and returned empty answers, so we turned that off for structured tasks.

Accomplishments we're proud of

A tool that mostly refuses to guess. Our own tests caught several honesty bugs, such as AI text speculating about rent and a "preferred" label that implied a recommendation, and we fixed them.

What we learned

For beginners, trust matters more than cleverness. Labeling data as real, estimated or made-up is a product feature, not a footnote. We also learned how much free, official data exists if you go and look for it.

What's next

Live commercial listings and rents through a licensed data partner, review and rating data, local zoning lookups, and more countries with proper local rules.