Powered by NVIDIA Nemotron
Money Mirror
Plenty of apps will tell you a net worth number. Very few will tell you where it came from, and almost none teach you anything you can act on afterwards.
Money Mirror counts everything you own: the cash, the furniture in your apartment, and what the labor market will pay for what you know. Every figure arrives with its reasoning attached.
Your mirror, right now
- Net worth
- Loading…
- Market potential
- Loading…
- Items scanned
- Loading…
The problem
A number you can't explain is a number you can't decide with
Financial literacy doesn't fail because people can't add. It fails because the tools hand over a total and stop, so nobody learns which inputs moved it or what to do next.
- Budgeting apps end at the balance. They categorize what already happened. They never explain what a figure is made of, so there is nothing to learn from.
- Most of what people own is never counted. A laptop, a bike, a couch. Real resale value, invisible to every account-linking tool because it isn't in an account.
- Earning power is the largest asset most people under thirty have. It's also the only one no personal finance app tracks, priced, or tells you how to grow.
The math
Transparent on purpose
financial assets + physical assets − liabilities = estimated net worthMarket potential sits beside that number rather than inside it. Your earning power is real, but it isn't cash, and folding it in would make the total a lie.
What it does
Three kinds of capital, one mirror
Financial
Accounts, investments and property tracked against loans, credit and other liabilities. The arithmetic is ordinary and deliberately boring: no model touches it.
Open Financial →Physical assets
Photograph something, or pan a short video around a room. A vision model picks out the items, you tap one to isolate it, and it is priced against real eBay comparables with Nemotron explaining the result. Anything you confirm joins your inventory and rolls into net worth.
Scan a possession →Human capital
Upload a resume. Nemotron reads it into a structured profile, matches it to occupations, and pairs it with real Bureau of Labor Statistics wage data. You get a compensation range, the skills the market is asking for, the ones you're missing, and the better paying roles already within reach.
Upload a resume →The model
Nemotron does five different jobs here
Not one job five times. Each is a separate prompt with its own schema, its own failure handling, and its own re-prompt when the reply doesn't parse.
- Profile extraction
- Unstructured resume text into a typed profile: roles, dates, skills, education, publications.
- Occupation matching
- That profile into candidate occupations, each with the key skills it asks for.
- Compensation estimation
- A structured salary estimate for roles where BLS publishes no series.
- Asset valuation
- A resale price for a detected object, with written reasoning you can read and disagree with.
- Explanation
- Plain language prose explaining what a computed range means and why it varies.
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning, served from
NVIDIA NIM.
The design rule
Nemotron supplies judgment. C# does the arithmetic.
The model is never asked to add up your net worth.
FinancialNetWorthCalculator, BlsCompensationAggregator,
SkillGapAnalyzer, OpportunityFinder and
MarketValuationCalculator are plain deterministic code, and
they own every number on the screen.
The explanation prompt is the clearest case. It hands Nemotron a range already computed from real BLS observations and tells it, in so many words, not to propose numbers of its own, only to explain the evidence it was given. That keeps the prose educational instead of confidently invented.
Where the numbers come from
Real sources, named on the page
Bureau of Labor Statistics
The Public Data API v2 time series endpoint, queried directly for occupational wage observations. Suppressed values are dropped rather than guessed at, and if nothing usable comes back the app reports insufficient data instead of inventing a range.
eBay comparable listings
Real eBay listings for the same item are the evidence behind a
resale figure. Every comp is kept with its price, title and
condition, and MarketValuationCalculator takes their
median rather than asking a model to pick one. Under three usable
comps the figure is flagged low confidence instead of quietly
presented as fact.
NVIDIA vision
meta/llama-3.2-11b-vision-instruct finds the objects
in a photo or a video frame. Non-JSON replies are re-prompted, and
a saturated endpoint retries with backoff instead of failing the
scan.
MobileSAM, on your device
Tap an object in the camera view and the cutout mask is computed in your browser on WebGPU, with a WASM fallback. Nothing but the crop you chose is ever uploaded.
Built to be honest
When it isn't confident, it says so
- No fabricated ranges. Insufficient evidence is an explicit result everywhere it can happen, on both the wage side and the valuation side.
- Low confidence gets flagged, not hidden. Scanned items below the confidence threshold carry a warning on the figure itself, not in a footnote.
- Every provider failure is inspectable. Errors come with a Show provider response button, so you can see what actually came back.
- It runs on your machine. One Docker command brings up the app and its database. There is no account, and your resume and photos never leave it except to reach the model endpoint.
See what you're actually worth
Start with whichever you have on hand. Both take about a minute.