PhD in political science. 15 years in the Knesset, government and field. Integrating AI to bring out the best of the tools and the data.

An empirical analysis of 24,225 Private Member Bills from the 19th through 25th Knesset. The project breaks the raw count down into new ideas and re-submissions based on the self-declarations Knesset Members recorded in their explanatory notes (divrei hesber). The picture that emerges: 31% new bills, 69% re-submissions — and of those re-submissions, roughly 79.2% cross party lines.

Context

Academic and journalistic criticism of the “flood of Private Member Bills” relies on a raw count that inflates the number by a factor of two to three. The re-submission layer has never been separated out systematically, because it is buried in free text rather than a structured field. This project brings it to light.

Approach

A daily pipeline that mirrors the Knesset database, classifies each bill based on the explanatory notes the Knesset Member wrote themselves, and exports results to an interactive site. The classification engine relies on explicit self-declarations that Members record in their bill text — naming whether a bill was submitted before and in which Knesset. This approach is grounded in the original text rather than algorithmic inference, and makes it possible to trace chains of re-submissions that cross party lines and administrative terms. Linking bill families is also supported by embedding-based document similarity on the explanatory notes, as a fallback for detecting relationships not explicitly declared.

Data

• Scale: 24,225 Private Member Bills (Knessets 19–25)
• Source: Knesset OData API + scraping of explanatory notes
• Format: Parquet + DuckDB (storage), JSON (API — 11 endpoints)
• Coverage: 1,632 bill families (18%) cite predecessors from Knessets 1–18

A full web presence for the book “First Reading: 61 Small and Big Stories about the Knesset” (Tal Elovits and Itay Setat): a marketing homepage alongside an interactive source archive where every footnote in the book is preserved and archived — live at knessetbook.co.il. The book publishes in September 2026.

Context

A book resting on hundreds of sources — news articles, protocols, Knesset documents — is exposed to link rot: without proactive archiving, a significant share of the references would break within a few years of publication. Beyond preservation, a book heading to print needs a home on the web: a place where readers find the story, book a lecture, and sign up for a launch reminder.

Approach

Two layers. First — active archival: every source is preserved on the Wayback Machine or as a local copy (PDF and screenshot), with a preview image for each footnote and numbering that matches the printed book one-to-one. Second — the public site: homepage, a three-part table of contents, per-chapter archive browsing, and lecture-booking and launch-reminder forms handled in-house, with no third-party services. Under the hood: structured data and full indexing infrastructure, so the book is found in search.

Data

• Scale: 61 chapters in three parts, 267 footnotes numbered to match the printed book
• Archival: 249 of 249 external URLs preserved (100%), 247 preview images
• Publication: the book publishes September 2026

An interactive globe of Knesset Members’ overseas travel, based on Ethics Committee publications from the 16th through 25th Knesset: 2,964 flight records, 331 Knesset Members, approximately 1,195 unique funding bodies, 98 destination countries. The display includes a timeline playback, a heatmap of the 10 top destinations, and a 2D map alongside a 3D globe.

Context

The Ethics Committee publishes travel approvals quarterly as PDFs. The data is public but scattered across formats that vary by Knesset — six different formats across the 16th–25th Knessets. There was no unified database that allowed filtering by funder, destination, or Knesset.

Approach

The data was published as PDFs in six different formats — one per Knesset. The processing pipeline is a multi-stage data normalization pipeline: it extracts each format separately, corrects Hebrew text issues that arise from reading PDFs, and normalizes in three stages: matching Knesset Member names against the official list, grouping 1,731 variants of funding body names into 711 canonical entities, and mapping destinations to unified country names. Funding body normalization was done in collaboration with AI — a hybrid process where manual curation and AI verification complemented each other.

Data

• Scale: 2,964 flight records, Knessets 16–25
• Knesset Members: 331 (100% matched to the official list)
• Funding bodies: ~1,195 unique (from 1,731 raw variants)
• Destination countries: 98

A research mapping tool for the CARRD (checks and balances) lab at Tel Aviv University, of which I am a member. The tool classifies proposals to reform Israel’s governance structure into 25 categories of democratic dysfunction — from protecting judicial independence to flaws in the electoral system and press freedom. Output: JSON for exploration and an HTML dashboard embedded in the site.

Context

The CARRD lab collects democratic reform proposals from research, podcasts, and policy documents. The challenge: the proposals are written in free language and cannot be directly compared. Building a typology that research can work with requires classifying each proposal along predefined dysfunction axes.

Approach

A typology of 25 categories, developed in dialogue with lab members and adapted to Israeli discourse on governance and democracy. Each proposal goes through prompt-engineered classification with a fixed taxonomy: classified separately against each of the 25 categories, with a confidence score per classification. Output: a queryable database, with an HTML interface embedded in the lab’s site.

Data

• Framework: 25 categories of democratic dysfunction
• Source: Reform proposals from research, podcasts, and policy documents
• Output: Queryable JSON + embedded HTML interface
• Documentation: Accompanying bibliographic review across all 25 axes

A system for structured extraction from policy documents — Hebrew government support documents (calls for applications and support criteria) — built for המארג הישראלי. The system includes an NGO database, an eligibility matching engine, and a dashboard for budget analysis. Running at hamaharag.teca-it.com.

Context

המארג הישראלי, which monitors government support grants for nonprofits, was working with dozens of support documents each written in a different format. Manual extraction of fields such as funding ceilings, eligibility criteria, and submission deadlines — from a 40–80 page document — takes hours. The system automated this process.

Approach

Each support document goes through AI processing in three stages: mapping the topics in the document, extracting structured fields against a fixed predefined schema, and filling in critical fields that remained missing. After extraction, the system connects to government databases to verify budget data and match support programs to the relevant nonprofits.

Data

• Documents processed: 52 calls for applications and support criteria documents
• NGO database: 45,590 organizations from GuideStar and 8,679 programs from Gefen
• Extracted fields: funding ceilings, eligibility criteria, submission deadlines
• Source: Israeli government ministries, Hebrew-language documents only

A unified system for analyzing 5 years of Estates Committee decisions (2020–2024): 2,374 decisions, 2,302 unique organizations, 459,733,733 NIS allocated in total. All annual totals are verified against the signature pages of the official records. Exists as an independent sub-project within KolKore, and operates as a static viewer and public SQL API at hamaharag.teca-it.com/izavun.

Context

The Estates Committee allocates funds from private estates to public nonprofits. The records are public, but split by year in PDF format — no unified database, no search interface, and no way to examine allocation patterns over time by organization or estate theme.

Approach

The data is spread across annual PDFs from the official public records. We built a structured data extraction pipeline from public records: it scans all the records, extracts and matches allocations, and builds a 5-year activity history for each nonprofit. Adding a new year to the system requires a short configuration cycle, not new development.

Data

• Scale: 2,374 allocation decisions, 2,302 unique organizations (2020–2024)
• Source: Official Estates Committee records — one PDF per year
• Total allocated: 459,733,733 NIS — verified against signature pages
• Output format: Excel, SQLite, and an HTML viewer with a public SQL API

A search interface across 35 transcripts of the special committee handling the Security Service Bill (25th Knesset), built for Israel Hofsheet. All transcripts were downloaded as DOCX from the Knesset website, processed in Python, and presented in a Hebrew RTL interface.

Context

The special Conscription Bill Committee produced a large volume of transcripts over several months. Reading each one independently is not practical — Israel Hofsheet wanted a research tool that lets users locate speakers, positions, and rulings directly by topic.

Approach

All 35 transcripts were downloaded from the Knesset website and processed automatically. AI extracted motions, positions, and rulings from each one. The interface combines full-text search with AI-driven motion extraction — and allows jumping to the original quote, because the team needs full context, not just a summary.

Data

• Scale: 35 transcripts from the special committee on the Conscription Bill (25th Knesset)
• Source: Knesset website — DOCX downloads
• Processing: Extraction of motions and rulings per transcript
• Format: Full text + structured actions for search

A weekly tracking system for social media platform regulation across 39 OECD countries, built for the Brandeis Institute. The pipeline identifies new regulations, classifies them, and sends real-time alerts to the team via Telegram and email. Running at tracker.teca-it.com.

Context

Social media platform regulation is evolving at a different pace in each country. Researchers, policy advisors, and journalists had no single place consolidating developments in real time — which is what the Brandeis Institute team needed for ongoing monitoring.

Approach

A weekly 7-stage agentic monitoring pipeline: discovering new regulations across all 39 countries, retrieving documents, filtering and verifying, AI analysis, and summarizing into a weekly report with Telegram and email alerts. Each regulation receives a reliability score that weights source type, source diversity, and academic documentation — a score ≥60 signals high confidence.

Data

• Scale: 39 OECD countries, ongoing weekly monitoring
• Sources: Government websites, policy organizations, news outlets, academic literature
• Regulations identified: 24 countries with active social media platform regulation
• Frequency: Weekly

TECA’s internal management system for clients, tasks, expenses, and scheduling. Includes AI email triage and structured invoice extraction, Sumit integration (Israeli invoicing), Hebrew PDF expense reports, and a Telegram bot for day-to-day management. No public URL — login only, deployed on a VPS.

Context

Self-managing a consulting firm requires filtering invoices from emails, reconciling to expenses, and exporting to reports — two to three hours of repetitive work per week. The system automates most of this, including receipt OCR and incoming invoice filtering.

Approach

Email syncs regularly; each message goes through AI analysis that identifies whether it contains an invoice, classifies the expense, and generates a Telegram alert for approval. Invoices are automatically linked to Sumit (an Israeli invoicing platform). Monthly Hebrew PDF expense reports are generated automatically and uploaded to Drive.

Data

• Sync: 2 Gmail accounts, synced regularly
• Integrations: Sumit (invoicing), Google Drive, calendar, Telegram bot
• Estimated savings: ~10 hours of repetitive work per month
• Deployment: Private VPS, no public URL

An ETL of Knesset caucus data from the 9th through 25th Knesset: 1,220 rows, 834 unique name groups (unique caucuses), 434 Knesset Members, 14 Knessets.

Approach

An ETL with three-stage name normalization: a pipeline that unifies caucus names written in varying forms across the years into a single consistent entity. Roughly 60 merges were verified by hand, followed by automated completion based on textual similarity. Topical classification into 16 categories covers 83% of caucuses. Output: a local dashboard for exploring caucuses across all Knesset terms.

Data

• Scale: 1,220 rows, 834 unique caucuses
• Coverage: Knessets 9–25 (14 terms)
• Scope: 434 Knesset Members
• Topic classification: 16 categories, 83% coverage

An ETL toolkit for collecting and processing Israeli government public data. Three modules to date: 25,800+ government resolutions from 1993 to the present (classified by 12 action types and 17 domains); explanatory notes for resolutions with dedicated schematics; and structured extraction of Knesset Ethics Committee decisions from PDF transcripts. The project is also a practical illustration of how I work as a PM: the template established in the first module — DEV_GUIDE, pipeline architecture, Cloudflare bypass guidance — is the starting point for every subsequent module.

Context

gov.il is a SPA behind Cloudflare — standard requests gets a 403. Scanned Hebrew PDFs require OCR. Content categories are not defined in structured fields. A developer joining in 2026 has to work through all of these obstacles from scratch — unless they are already documented.

Approach

The Israeli government website does not offer an open API — the data is accessible through an interface that obscures its data structure. We built a toolkit that discovers the hidden access points, extracts the data, and applies multi-stage classification (rules + AI) on government documents, with multi-modal OCR for scanned PDFs — across 12 action types and 17 domains. A developer guide is written before each module’s code — so a developer joining reads only the differences from the previous module.

Data

• Scale: 25,800+ government resolutions, 1993 to present
• Source: Israeli government (gov.il) + Knesset Ethics Committee transcripts
• Classification: 12 action types × 17 domains — ~60% rule-based, ~40% AI
• Output: Structured Excel per module

A harmonization pipeline for 40 waves of the “Israeli Voice Index” (IDI), covering 12/2022 through 03/2026.

Approach

Each wave is a standalone Excel file with Hebrew column headers that vary from wave to wave — questions were reworded and weight variables use different names. Multi-wave harmonization across all 40 waves builds a canonical variable registry and a wide panel that enables longitudinal analysis of Israeli public opinion over more than three years. Output includes 35 analysis blocks, 560+ charts, and an interactive viewer.

Data

• Scale: 40 survey waves, 12/2022 through 03/2026
• Sample: ~750–800 respondents per wave (~600 Jewish, ~150 Arab)
• Source: “Israeli Voice Index” — Israel Democracy Institute (IDI)
• Output: 35 analysis blocks, 560+ charts

An interactive dashboard of biographies for 1,098 Knesset Members — education, prior occupations, birthplaces, across all Knessets.

Approach

Data was collected from the Knesset’s official API and processed through cleaning and enrichment stages. AI subagents performed structured field extraction from free-form biographical text — converting education and occupation details into structured fields. The output is a locally served dashboard filterable by Knesset term, party, education level, and occupation — no server required.

Data

• Scale: 1,098 Knesset Members
• Source: Knesset official API
• Coverage: All Knesset terms
• Format: JSON, presented in dashboard.html

A PDF-to-structured-JSON ETL of the 2026 state budget book, covering 29 PDF files across government ministries and the defense establishment.

Approach

The documents were scanned and converted into a unified data structure extracting four types of information: verbal explanations, budget changes with justifications, a time series of budgets from 2017–2026, and employment data by gender. The output serves as a preparatory stage for future budget analysis projects.

Data

• Scale: 29 PDF files — President/Knesset, government ministries, defense establishment
• Time range: Budget time series 2017–2026
• Output: budget_database.json — a single unified budget database
• Chapters: A–E (verbal explanations, changes with justifications, time series, employment)

A lightweight OData client for exploring Knesset open data via the official OData interface (ParliamentInfo).

Approach

A small local server that bridges the browser to the Knesset’s open API, making it possible to browse Knesset data directly from a browser without any technical configuration. A simple browsing interface enables visual exploration of the various data entities. Used as a base layer for other Knesset projects such as caucuses and MK biographies.

Data

• Source: Knesset official OData interface — knesset.gov.il/OdataV4/ParliamentInfo
• Content: Parliamentary entities: MKs, caucuses, committees, bills, and more
• Usage: Base layer for Knesset projects — caucuses, MK biographies