Notes on AI agents, data platforms, software engineering, and production systems by Itamar Weiss.
For most of two decades, turning a pile of photos into a 3D scene meant running an optimizer . You'd feed a few dozen images into COLMAP, walk away, and come back to find it had matched features, guessed where every camera was, and slowly nudged all of it into agreement. It worked, it was accurate, and it was the thing you reached for. It was also slow, fragile, and prone to just... giving up on a…
A few weeks ago I opened up a dataset of Hezbollah FPV drone-strike videos — every clip I could pull out of OSINT, labelled and kept in one place so anyone building a defense could start from real footage instead of a press release. It worked as a dataset , but it was miserable as something to actually look at . To answer a simple question — "show me the strikes on air-defense sites near the…
I've been working with Apache Iceberg for several years. In the last year I've been helping one of the most data-heavy organizations in Israel deploy it at scale — petabytes of data, multiple ingestion pipelines, constant schema evolution, several query engines reading the same tables. Exactly the kind of environment that stress-tests every assumption you had about the format. Most of the problems…
The cheapest, fastest-evolving threat on Israel's northern border right now is a $500 quadcopter with a warhead zip-tied to it. Hezbollah has been flying FPV (first-person-view) drones across the frontier and posting the strike videos as propaganda — and every one of those clips is also, whether they like it or not, a piece of intelligence. The camera is the weapon's eye. If you watch enough of…
Every "AI for data" demo I've seen does the same magic trick: point a model at a database, ask a question in English, get an answer. It's genuinely impressive for about a hundred tables. Then it quietly falls apart — the model joins two unrelated id columns, invents a metric, or confidently reports a number that contradicts the dashboard three feet away. The usual reaction is "the model isn't…
There's a Hebrew language model called Hebatron that, in a single week, racked up 30,000 downloads and four community quantizations. For an open model that isn't in English, that's a lot. I listened to the ExplAInable episode where the team that built it walked through how it was made — and it turned out to be one of the most honest accounts of actually training a model I've heard: the dead ends,…
I recently listened to a Geekonomy episode with Dr. Assaf Natanzon , Chief Architect at Eon , and it reframed something I thought I understood. Eon started life as a cloud-backup company. Somewhere along the way they realized that the artifact a backup product already produces - a normalized, queryable copy of every piece of data in the organization - is exactly the thing AI has been starving for.…
I keep running into Gaussian splatting — in mapping demos, in VR captures, in the "scan your living room with your phone" apps — and for a while I only had a fuzzy sense of what it actually was. So I sat down to understand it properly, and this post is the explanation I wish I'd had: what problem it solves, how it relates to NeRF, and why it took over so fast. I've built a few interactive pieces…
You've seen it a hundred times. A character stares down a hallway, the camera holds them dead centre at exactly the same size — and yet the walls seem to lunge toward them, the far end of the corridor rushing in or falling away. It's the shot Hitchcock invented for Vertigo , and it shows up every time a director wants to put unease on screen without moving the actor. The dolly zoom . Where you've…
When a team starts building a brand-new product, the most valuable thing it owns isn't in the product yet. It's scattered across customer calls, Slack threads, WhatsApp arguments at midnight, hallway conversations, and half-finished decks. Sa'ar Arbel, who leads an engineering group at monday.com, calls this "shadow data" — and on a recent Startup for Startup episode he described how his team…