Gilad Shalom — AI Builder & Data Analyst
I build practical AI tools, analyze data, and turn ideas into working products and solutions.
Seven years running high-stakes operations taught me to ship under pressure. Now I do it with code and data.
AI tools · Automations · Data analysis · Full-stack
About
I build practical AI tools and data products, and I take them the whole way: from the question, through the data and the code, to something a person actually uses.
Seven years running high stakes operations at the Israel Ministry of Justice, a food science degree where I took a product from concept to prototype, and now the Google × Reichman Data Analyst and AI Program. I learn a system fastest by building something real with it.
- 9 — Products and analyses shipped end to end
- 6 — Full-stack systems and automations coded
- 7 yrs — Leading high-stakes operations
- 200+ — Daily operations coordinated
Featured work
Dr Chassid Eye Clinic (2025)
A private eye clinic runs its whole day on this: scheduling, patient files, clinical eye records and prescriptions, on web, iPhone and Android.
Problem. A private ophthalmology practice was running on a paper diary and a general-purpose calendar. Appointments, eye records, prescriptions and follow-ups lived in separate places, so nothing could be checked against anything else and nobody could see which patients were overdue.
What I built. A complete clinic management system. A real scheduling calendar with five colour-coded visit types, per-doctor conflict detection and tokenized links that let a patient cancel without an account. Patient files with per-eye clinical records for sphere, cylinder, axis, add and IOP, and generated prescription PDFs. Role-guarded logins for doctors and receptionists on short access tokens. And an Insights view that scans the practice and surfaces follow-ups coming due plus every intraocular pressure at or above 21 mmHg worth a glaucoma work-up.
My role. Sole developer and product designer. Scoped the clinic’s real workflows with the practice, designed the schema, then built the API, the web app and the mobile app.
Result. Chasing a follow-up and spotting a pressure worth a second look no longer depend on somebody remembering. One system covers the whole working day instead of four tools, and web, iOS and Android run off a single API in a Turborepo monorepo, so a change to the shared schema reaches all three at once.
Built with: React, TypeScript, Fastify, PostgreSQL, Prisma, Expo, Docker
Israeli Supermarket Analytics (2026)
Six months of one supermarket branch turned into six moves management can act on, starting with the finding that its checkout pressure is a scheduling problem rather than a hiring one.
Problem. A supermarket branch believed it was understaffed at the tills. Hiring is the expensive answer, and nobody had the evidence to say whether it was the right one.
What I built. An end-to-end analysis in BigQuery over millions of in-store location pings and tens of thousands of receipts: a data audit, customer-versus-staff segmentation inferred from behaviour alone, demand modelled by day type and hour with linear regression, and a bilingual Looker Studio dashboard a manager can explore without needing an analyst.
My role. Solo, end to end: framing the question, auditing the raw tables, writing the SQL and Python, building the dashboard, and presenting the recommendations to the client.
Result. One Friday lunchtime hour turned out to be the only hour needing the full checkout line, against a roster sized for it all week. That reframes the ask from more people to different hours, and five further moves came out alongside it. Every regression was computed twice by independent routes that have to agree.
Built with: BigQuery SQL, Python, Linear regression, Geolocation data, Looker Studio
FitApp (2026)
A Hebrew nutrition tracker built on the actual supermarket shelf: 34,650 Israeli foods, barcode scanning that works on iPhone, and a coach that answers from your own log.
Problem. Nutrition apps in Israel run on crowd-sourced food databases, so searching for the product you just bought returns somebody else’s guess at it. In English, laid out left to right.
What I built. A Hebrew, right-to-left tracker over 34,650 Israeli foods, assembled from the supermarket chains’ own public price files and resolved against laboratory-measured Ministry of Health figures. A scanline EAN-13/EAN-8/UPC reader written from scratch so scanning works on iPhone as well as Android. Photograph a label or a plate and a vision model fills in the rest, with calories always coming from the database rather than the model. Mifflin-St Jeor targets, a day of Israeli meals scaled to hit them, and a coach that answers from the real log.
My role. Solo build: the data pipeline, the barcode decoder, the app, the AI features and the MCP server.
Result. A search returns products actually on sale rather than crowd-sourced uploads. No Node, no build step, no database and no account required: signed out, nothing leaves the browser. It also ships an MCP server, so Claude can query the same food data directly.
Built with: Vanilla JS, Python (stdlib), Hebrew · RTL, Barcode decoding, Open data, Claude API, MCP
Agentic OS (2026)
Type a one-line mission and a Manager agent delegates it to a roster of specialists — a working multi-agent system, not a mock-up.
Problem. Multi-agent AI demos are usually either a video, or they need a pile of API keys before they will run at all. Neither lets somebody watch the orchestration actually happen.
What I built. A personal command center. A Manager agent takes a one-line mission, decides how to split it, and delegates to Researcher, Critical Thinking and Creative agents whose cards report what each one is doing — wrapped in live markets, FX, weather, calendar, Notion and Slack panels, plus an Agent Village where the same crew moves between assignments. Installable as a PWA.
My role. Solo build: the architecture, the agent orchestration layer, the Node service and the whole interface.
Result. One clean seam makes it both a public demo and a real tool. The mission service ships in two build-time flavours: a mock that turns the whole thing into a static site with no backend and no secrets to leak, and a live one that posts to a small zero-dependency Node server running Claude through the local CLI. Either way there is no API key to manage.
Built with: React 19, TypeScript, Zustand, Framer Motion, Node.js, PWA
More work
ZenCity Bike Rental (2026)
SQL analysis of 16,585 Austin bike-rental trips that found a 77.3% revenue dependency, grown into a dashboard that forecasts Q2.
Problem. ZenCity wanted to plan for Q2 growth without knowing which riders its revenue actually depended on, or what the summer would do to demand.
What I built. Advanced SQL in BigQuery over 16,585 real transactions — CTEs and window functions for cleaning, segmentation and station performance — then an interactive React dashboard with a predictive Q2 simulator whose regression calculator re-forecasts demand as you change the scenario.
My role. Analysis and SQL solo; the dashboard built together with Michael Geiner.
Result. 77.3% of revenue traced to a single rider segment, a concentration risk worth de-risking before summer. Four Q2 moves came out of it, each mapped back to the hypothesis it answers.
Built with: BigQuery SQL, Python, React, Google Sheets
ShopIL Dashboard (2026)
An e-commerce dashboard where every number on the page can be traced back to the SQL query that produced it.
Problem. Dashboards get trusted or distrusted as a whole, because there is usually no way to ask where one specific number came from.
What I built. Six months of orders as a single self-contained page with no dependencies and no build step, whose date-range, category and channel filters genuinely recompute every KPI and chart. Alongside it, a companion SQL file with every widget’s query written out and organized by section.
My role. Solo: the data model, the SQL, the page and the export path.
Result. Every figure traces to the statement that produced it, so a number can be re-run instead of taken on faith. The verified result is committed beside each query as a comment. The page opens offline from a double-click and hands back the current filter state as CSV.
Built with: SQL, SQLite, Data visualization, Vanilla JS
CV Helper (2026)
Reads a CV against a job posting and shows which of your own lines are actually earning their place.
Problem. Tailoring a CV to a posting is guesswork. You can tell the match is imperfect, but not which of your own lines are doing the work and which are filler.
What I built. A browser tool that reads a CV as PDF, DOCX, TXT or Markdown, pulls the weighted terms out of a pasted job posting, scores the match, splits those terms into what the CV already covers and what is worth adding, then ranks the CV’s own lines by how well each one answers the role.
My role. Solo build. Written for one job hunt and kept because it was faster than the alternative.
Result. The matched terms are named under every line, so the suggestion is arguable rather than magic. One HTML file, one stylesheet, one script: the analysis runs entirely in the browser with no key, and an optional Claude API key kept in local storage unlocks a full rewrite pass.
Built with: Vanilla JS, Claude API, PDF/DOCX parsing, No build step
StatsTutor AI (2025)
A statistics tutor that adapts its explanations, hints and practice to how each learner is actually doing.
Problem. Statistics is the course everyone dreads, and a fixed question bank cannot tell the student who understands the concept from the student who guessed right.
What I built. An AI-powered learning platform where feedback and analytics drive what comes next, so explanations, hints and practice adapt to each learner instead of following a set order.
My role. Solo build: the adaptive logic, the AI feedback layer and the interface.
Result. Practice targets what a learner is actually getting wrong rather than whatever is next in the list. The project that first put the analytics work and the user-experience work in the same place.
Built with: Python, AI, Analytics, UX
High-Protein Pancake Mix (2025)
A high-protein pancake mix taken from concept to a working prototype — where the end-to-end habit started.
Problem. High-protein baking mixes trade texture for protein content, which is why most of them get bought exactly once.
What I built. A full product development cycle for my B.Sc. capstone at Tel-Hai: formulation, market research, feasibility, regulatory review, and iterative optimization of the final mix.
My role. Led the project end to end, from concept through to a prototype.
Result. A working prototype you can hold and eat, and the habit of owning a build from the first question to the finished thing. It is the same instinct that goes into the data and software work now.
Built with: Formulation, Market research, Regulation, R&D
Certifications and education
- Data Analyst & AI Program — Google × Reichman Tech School (Dec 2025 — Sep 2026)
- B.Sc. Food Science — Tel-Hai University (2021 — 2025)
- Generative AI: Introduction and Applications — IBM · Coursera (Jun 2026)
- Introduction to Data Analysis Using Python — Google · Coursera (Jun 2026)
Contact
This site needs JavaScript for its full version. The plain text of
everything on it is also at /llms.txt.