Most money apps die for two reasons: typing every expense and numbers that mean nothing.
Momentum answers the one question you actually ask — "how much can I spend today?" — and it does the maths where your data already lives: on your device.
There's a second, quieter reason people quit finance apps: every notification is designed to make you anxious about a badge, a streak, a tier you might lose. Momentum's only feedback is "your pattern is understood, your number is real" — the dopamine of being in control, not the dopamine of being chased.
It's for you if you:
- want to know what you can spend today, not a chart of last month
- are a freelancer who dreads tax deadlines (Italy 🇮🇹, Spain 🇪🇸 and Switzerland 🇨🇭 supported)
- invest and want the return after tax, not the brochure number
- don't want your bank life sitting on someone else's server
Every other finance app has a server. That server is the product: it holds the data, and the data is the business model.
Momentum inverts it. The value comes from never receiving your data.
That isn't a privacy promise bolted on top — it's the architecture. Devices sync directly with each other (WebRTC, peer-to-peer, no signalling server). The AI trains locally. There is no account to create, because there's nothing to create it on.
A competitor who monetises data or cloud cannot copy this. They'd have to delete their own business model first.
"You can spend X today." Weekly budget derived from your monthly one, proportional to real days, envelope carry-over, minus subscriptions about to hit — divided by days left.
Month-end projection uses Holt-Winters on your actual trend (falls back to run-rate, and always tells you which method it used).
The part that turns a budgeting app into infrastructure.
🇮🇹 Italy
- Set-aside engine — every payment received tells you what's yours and what's the taxman's. Flat-rate (forfettario) and standard regimes, real ATECO coefficients, INPS, professional funds (casse), reduced 24% rate for employed-plus-freelance.
- Real e-invoices — FatturaPA v1.2.2 XML, generated on-device. Plus an offline predictor of SdI rejection codes (00400/00415/00417/00422/00423/00427…) so you find the error before sending, not after.
- Multi-line invoices, percentage lines (discount/surcharge), courtesy PDF, real checksum validation of VAT numbers and codice fiscale.
- Periodic VAT settlement, purchase register (deductible VAT), passive invoice import (drop in the XML you received, it books itself).
- Pre-filled F24 with verified tax codes (1790/1791/1792, 4033/4034/4001, 6001-6012, 6031-6034, P10) ready to copy into your bank.
- Missed a deadline? Ravvedimento operoso calculated automatically — reduced penalty by lateness band plus legal interest. Most tools just let the deadline vanish.
🇪🇸 Spain
- RETA + IRPF for autónomos — every invoice tells you what's yours and what to set aside. 15 real cotización tramos (tabla reducida + tabla general) sourced from the BOE (Orden PJC/297/2026), 2026 contribution rate verified at 31.5% (contingencias comunes + profesionales + MEI + cese de actividad + formación profesional).
- Withholding on invoices (retención IRPF, 15% standard / 7% for the first years of activity) — calculated on what actually lands in your account, not on the invoiced amount.
- Only the state component of IRPF is estimated: the autonómica one varies across each of the 17 comunidades autónomas and is never guessed — a stated limit, not hidden in one made-up number.
🇨🇭 Switzerland
- AHV/IV/EO contributions for the self-employed, VAT threshold (CHF 100,000 — many freelancers don't need to register at all, and Momentum says so).
- QR-bill — the payment QR code mandatory on every Swiss invoice since 2022. Payload verified byte-for-byte against three official SIX Group examples; the QRR reference check digit verified against SIX's own published reference.
- Interface in German, French, Italian and English, auto-detected.
Every simulator shows you gross. A 7% ETF is not a 7% ETF.
Momentum shows the return after capital gains tax and stamp duty, with rates verified per country — 🇮🇹 Italy (26% / 12.5% govt bonds / 0.2% stamp duty), 🇩🇪 Germany (26.375% flat + €1,000 Sparerpauschbetrag), 🇫🇷 France (31.4% PFU, 2026), 🇨🇭 Switzerland (0% for private investors — with the professional-trader and cantonal wealth tax caveats stated, never hidden).
Plus: net worth, Monte Carlo projections with declared assumptions, market regime detection, 40 years of drawdown base rates, portfolio import and risk-parity rebalancing.
Never a buy/sell recommendation. The picture, never the order — that's a regulatory line, and it's also what makes it integrable rather than blockable.
Momentum reads real SEC filings — not a marketing claim, a script (bench/fetch-panel-sec.mjs) that pulls XBRL data for 11,304 US companies with reported revenue, enriches 1,500 with their real SIC code, and ships 600 of them in full inside the app.
- Sector percentile ranking — where a holding sits against real peers on revenue growth, margins and more, via a hand-built SIC→sector bridge (no free official crosswalk exists — documented, not hidden).
- Beneish M-Score & Piotroski F-Score — the same academic fraud/quality screens due-diligence teams use, computed on-device from the filings above. Momentum states their known blind spot in the same breath it shows the score: very high legitimate revenue growth can trigger a false positive on Beneish, and it says so every time.
- Causal & comparative analysis, per single stock or crypto — a 777-line statistics engine (regression decomposition, permutation testing) sat unreachable in this repo until the SIC→sector bridge unlocked it for any of the 600 tracked companies; a CoinGecko integration extends the same reasoning to major cryptocurrencies.
- Comparable company analysis (comps) — median EV/EBITDA and EV/Revenue against real peers, implied valuation, CSV export straight into Excel. Ask it in chat ("which companies are comparable to NVDA?") or from the asset screen.
- Crypto derivatives positioning — funding rate, open interest and long/short skew combined (not just listed side by side): real crowding is only declared when an elevated funding rate and a skewed position reinforce each other, checked against the coin's own recent history, not one universal threshold. No key needed (Binance Futures' public API). Ask it in chat ("am I too crowded on bitcoin?") or from the asset screen.
- Real stock price history with zero setup — for stocks without a personal market-data key, Momentum falls back to their tokenized proxy (xStock/Ondo/Backed and similar, quoted as ordinary crypto on the same free CoinGecko API), always disclosed as a proxy, never passed off as the exact exchange price.
- On-device news sentiment — a real DistilRoBERTa model (82.5MB, Apache-2.0, fine-tuned on financial news) reads the tone of a headline in under 100ms once loaded, no server, no API key.
- Peer-shared signals — a device that already computed a headline's sentiment, or already knows a price/rate, relays it (label + score only, never raw personal data) over the same P2P mesh to trusted peers who haven't downloaded the model yet — cross-checked against a second independent peer, or a peer with a track record, before anyone trusts it.
None of this pretends to be a live feed: every screen states exactly when the filing data was pulled. Reproduce it: npm run bench:panel regenerates the sector data straight from SEC EDGAR.
An ensemble that votes, plus an arbiter that learns which of its own models to trust, category by category, from your real corrections.
Technical detail
- NeuralNexus — Naive Bayes + neural net (real backprop, L2, gradient clipping), learns from use.
- Nano (always on) — MLP trained in Python/scikit-learn, ported to JS with verified numerical parity.
- Meso (mid/high tier) — hybrid TF-IDF words + character n-grams, 2 hidden layers, built for dirty bank text.
- Orchestrator — N-way weighted vote, weights modulated by per-category precision measured on your corrections (incremental matrix, Laplace smoothing, neutral when it has no data).
- Hardware tiering — a real micro-benchmark at boot picks Monte Carlo depth (500–10,000) and which engines to wake. Routine transactions never wake the heavy ones.
Reproduce the numbers: npm run bench, npm run bench:vs-llm, npm run bench:cash
Measured co-variation between categories on weekly differences (so a shared trend doesn't invent a link), lag 0 and lag 1, with damped propagation and an explainable path.
Stated plainly in the UI: "this isn't a law, it's what happened in your data."
Deterministic intents computed on your real data — "how much did I spend in June?", "can I afford €50?", "when do I pay Netflix?", "what if I spend more on restaurants?" Typo-tolerant. Speaks answers aloud. When it doesn't know, it says so.
One-tap buttons for habitual purchases, ordered by what's likely right now (measured hour-of-day and day-of-week histograms — coffee at 8am, groceries on Saturday), with the reason shown. Amount memory. Bank PDF import (Intesa, UniCredit, N26, Revolut), CSV, receipt OCR, voice with multi-action sentences.
The screens every user hits — onboarding, the Dashboard's headline number, category names, the import prompt, the "what's new" screen — in Italian, English, Spanish, German, French and Dutch (auto-detected, no toggle to hunt for). Chosen from real data, not a guess: France, Belgium, the Netherlands and Germany are where expense-splitting apps (Momentum's own viral loop) already have the deepest adoption. The rest of the app — settings, analysis, tax screens — is still Italian-only; that's declared above under "What's next," not hidden.
Explicit pairing between trusted devices over WebRTC. Weighted FedAvg, anti-poisoning validated on a local set, and a new device inherits the trained network on first link.
Dual-cache service worker, IndexedDB + localStorage, schema migrations, and a hash chain on transactions that is never rewritten.
Full breakdown, competitor by competitor, with the exact file that backs every claim: ANALISI_COMPETITOR.md (Italian).
The short version: Bloomberg Terminal and Bloomberg Intelligence see the market but never your cash. Revolut and your bank see their own account but not the market — and their business model profits from your spread, not your savings rate. Robinhood profits from your order flow. Copilot and Monarch need your bank credentials on a third party's server (Plaid) just to function. YNAB tracks what you already spent, never what the market is doing right now.
Momentum is the only one of these that puts your real cash position and real market data — sector percentiles, quality scores, sentiment — in the same on-device answer, because it's the only one with no server keeping them apart.
An honest pipeline, not a promise list — full detail and priority order in ANALISI_COMPETITOR.md §5 and PIANO_MOMENTUM.md.
- Extend UI translation beyond onboarding, the Dashboard headline, category names and the import prompt — most of the app (settings, analysis, tax screens) is still Italian-only.
- Mesh discovery beyond an already-exchanged pairing code.
- A second cryptocurrency in a pairwise causal comparison (today only vs. Bitcoin).
- Extend the causal engine to markets once there's enough personal-history data to be honest about the result.
- Deeper SLM reasoning for statistics/physics/algorithms beyond finance — not yet researched with the same rigor as the filings work above.
Don't take the claims. Run them.
npm install
npm test # 4620 tests, node --test src/Every capability above has tests next to the code. The Swiss QR-bill is checked against the official SIX examples; the tax rates carry the date they were verified and the source; the AI numbers regenerate with npm run bench:*.
npm install
npm run dev # localhost:5173
npm test # 4620 tests
npm run build # multi-file PWA in dist/
npm run build:singlefile # single ~575KB HTML fileOne runtime dependency: @huggingface/transformers, for the on-device sentiment model (opt-in, see above). Vite is the only dev dependency. This claim used to be "zero" — updated the day the sentiment feature landed, not left stale.
src/
ai/ NeuralNexus, Nano, Meso, Orchestrator, Q&A engine, calibration
predict/ cash forecast, tax engine (IT + ES + CH), VAT settlement, F24,
ravvedimento, deadlines, causal discovery, subscriptions, BNPL
invoice/ FatturaPA XML + SdI rejection predictor, passive import,
Swiss QR-bill, fiscal-ID checksums, per-country registry
alpha/ net return after tax, net worth, portfolio, market regime,
factors, drawdown base rates, verified data sources
mesh/ WebRTC signalling (no server), federated peer, sync CRDT
core/ vault (IndexedDB + hash chain + migrations), auto-update
split/ shared expenses, optimal settlement, invite crypto
i18n/ language detection + UI strings
import/ bank PDF, CSV, receipt OCR, notification parser
voice/ multi-action voice parser
307 source modules across 15 domains (find src -name "*.js" -not -name "*.test.js" | wc -l).
Trust is built by what a project admits, not by what it claims.
- A PWA cannot read other apps' notifications (iOS or Android). Direct reading needs a native Android shell (
NotificationListenerService). On iOS nobody can — there, the route is Open Banking. - iOS doesn't support Web Share Target for PWAs.
- The causal graph measures co-variation, not causality — and says so in the UI too.
- Momentum cannot transmit an invoice to the Italian SdI for you: that needs accreditation as an intermediary, which is a corporate process, not code. It prepares the correct file and walks you through the real portal.
- The Swiss QR-bill produces a correct, scannable code, not yet the fully compliant printable payment slip layout.
- Below CHF 60,500 the Swiss AHV uses a sliding scale that isn't a simple public formula — Momentum shows the verified minimum and links the official calculator instead of inventing a number.
- Spanish RETA picks the tramo from invoiced revenue, not yet from net income after deductible expenses (the official tramos are based on rendimientos netos): for anyone with significant expenses, the tramo shown can be higher than the real one. Needs a deductible-expense classification feature Momentum doesn't have yet.
- Not tax advice. Estimates on public rates, each carrying its verification date.
- Some AI modules are research, not production.
src/ai/omega.js,neurosym.js,expert-adapter.js,executive.jsandnb-categorizer.jsare written and tested but no production code path executes them — the live classification path isorchestrator.js+expert-bandit.js+trained-categorizer+hashed-logreg. We say this here rather than let the file count imply otherwise: a tested module nobody runs is not a feature. - Automatic tax-rule updates need a reachable source. Verified live: the Agenzia delle Entrate and Normattiva block cross-origin requests, so a browser cannot read them, and turning legal text into rates automatically would be exactly the kind of invented number this project forbids. Rules are verified by hand and published as a signed-validated JSON file the app fetches on its own.
- User data never leaves the device.
- No decorative modules — every claim is measured and tested (
npm test). - Pure functions separated from the DOM; every new module ships with its tests.
- The transaction hash chain is never rewritten.
- Every line of UI text must be understandable by an 8-year-old.
- If the number isn't printed, it doesn't exist. No unverified figure enters a document, a commit, or the UI.
- AGENTS.md — start here if you're picking up development (human or AI): non-negotiable rules, verified code map, declared limits, and the traps already paid for. Every claim in it is checked against the real code, never written from memory.
CLAUDE.mdpoints to the same file, so every tool reads one source of truth. - VERSIONI.md — per-component version manifest; versions are earned with measured leaps, never with labels.
- PIANO_MOMENTUM.md — development plan, phase status, gap list.
- NEUROSYM.md — the reasoning architecture, including what it explicitly is not.
Your money. Your device. Nothing leaves.