An early-stage Israeli fintech venture wanted to show investors and partners a mobile app that pulls a person's pension, provident and study-fund policies from the national pension clearing house and turns them into a single, readable picture. I designed and built the proof of concept end to end: a secure API layer, a data mapper for a deeply nested Hebrew payload, and a mobile-first, right-to-left dashboard.
Most Israeli employees hold several long-term savings products: a pension fund, one or more provident funds and a study fund, often with different providers. The national clearing house can return all of them in one response, but that response is built for institutions, not people. One sample answer for three policies holds 684 data points, with Hebrew field names, nested tables, numbers stored as text with thousands separators, and the same field arriving as a list in one policy and an object in another. A saver who wants to know “how much do I have and how is it doing?” usually ends up logging into each provider separately.
The diagram is wider than the screen: swipe sideways to see all of it →
For the demo, the dashboard reads a recorded clearing-house response through the same mapper the live data will use. Switching to live data is a change in one service file; nothing in the UI changes.
The clearing-house key and the personal-data request never touch the browser. The NestJS layer adds JWT auth, CORS and Swagger docs, so the front end and any future mobile client share one documented contract. Trade-off: one more service to run, which is cheap compared with leaking a regulated credential.
All knowledge of the payload's Hebrew keys, nested tables and inconsistent shapes lives in one mapper. Components only see a 9-field PensionFund type. Trade-off: the mapper drops most fields for now, but adding one is a single line in one file.
Live clearing-house access needs licensing and the saver's signed consent, which a one-week POC can't wait for. The server routes for the live flow were built first; the UI runs on a recorded response through the same mapper. Trade-off: the demo shows real data shapes, not real-time data.
A chart library was installed and tried, but for a first screen a saver needs three honest numbers: total, liquid amount and a balance-weighted return. Charts were left for the next phase. Trade-off: less visual flair in the demo.
A one-week POC still has real problems. Most of them came from the shape of the data.
The JWT guard on the clearing-house routes failed at start-up: the token service was registered in the auth module but not visible to the rest of the app.
Exported the JWT module from the auth module so every guarded route resolves the same configuration (1-hour tokens).
The investment-track details came back as a list in 2 of the 3 sample policies and as a single object in the third. Code that assumed a list would have crashed on a real customer.
The mapper normalises both shapes before reading the track's net return.
Projected balances arrived as text with thousands separators, other amounts as plain numbers, and dates as day-month-year strings.
One parsing path for every amount and date, with safe zero defaults, so a missing field shows ₪0 instead of “NaN”.
A simple mean of fund returns let a small fund with a high return outweigh a large pension fund. On realistic balances the headline return was overstated by more than 2 percentage points.
Returns are now weighted by balance, separately for short- and long-term horizons and for the total.
Each policy has its own validity date, so a single “as of” label could be wrong for half the cards.
Each card shows the latest validity month of the policies it sums, in Hebrew month names.
Lint warnings and leftover scaffolding had built up during a fast week.
Fixed all lint warnings, removed unused code, and made the production build fail on any lint or type error.
A synthetic three-policy portfolio (no client data). Switch between the two ways of computing the headline return.
Free scoping call · 30 minutes · No commitment
Write a few lines: the process or system, what hurts today (time, errors, cost or an AI pilot that stalled) and who uses it. After a short call you get a written plan and a fixed quote within ≈3 days.
Opens a pre-filled email · reply within one business day
The model reads. The code counts.
All projects