NLP Israel x Team8 - September 2025 Meetup

Agenda:
18:00-18:30 - Gathering, food, and drinks
18:30-18:45 - NLP IL & Team8 : Opening words
18:45-19:15 - Eyal & Maayan (Team8) -
Building Winning AI Companies: Short-Term Shifts, Long-Term Horizons
19:15-19:45 - Peleg (Orchid) -
Scaling Static Analysis to Deep Agentic Analysis
19:45-20:15 - Daniel & Orel (April) -
OCR is Dead. Long Live Multimodal LLMs

Abstracts:

Lecture #1 - Building Winning AI Companies: Short-Term Shifts, Long-Term Horizons (Team8)
The AI ecosystem is evolving at breakneck speed, but not every trend creates lasting value. In this session, we’ll explore how to distinguish short-term shifts from enduring technological transformations, and what it really takes to build AI companies with durable moats that can thrive in the long run.

Lecture #2 - Scaling Static Analysis to Deep Agentic Analysis (Orchid)
This talk presents our journey in building a production-grade system to automate non-trivial static code analysis on our customers' remote production servers. I'll explain the approaches we tried and the lessons we learned, from a simple linear prototype using TF-IDF that lacked context to a deep agentic approach, which mimics a human researcher by reasoning about the code and using tools to perform a more targeted and efficient analysis.

Lecture #3 - OCR is Dead. Long Live Multimodal LLMs (April)
OCR has traditionally been solved as a computer vision pipeline: detect, align, classify, and stitch text back together. While reliable, these systems are brittle, hard to extend, and limited by rigid assumptions about formatting.
This talk presents a new approach—reframing OCR as a multimodal language problem. By applying Large Language Models directly to visual-text inputs, we can collapse multiple steps (layout detection, classification, normalization) into a single reasoning task. The result: higher accuracy, lower latency, easier adaptability, and faster iterations.
We’ll cover a full solution pipeline, share evaluation framework and early benchmarks comparing classic OCR to LLM-based methods, and discuss lessons learned.