Elaheh Rashedi

Today’s guide by Reza Bahmanzadeh, AI Engineer on our Professional Services Team, teaches the essentials of serving large models with vLLM: https://lnkd.in/eHK9nwZm It builds from foundational neural network concepts, like transformers and attention, to introduce practical inference workflows, explore useful vLLM features and offer hands-on, operational guidance for efficient deployments. #LLMs #Serving #Inference #TechGuides

35

1 Comment

✨ Imagine this: You’re talking to an AI, and instead of forgetting everything after the chat ends, it remembers. It recalls past conversations, learns from them, and builds context like a trusted partner. That’s precisely the world I stepped into with the LLMs as Operating Systems: Agent Memory course from DeepLearning.AI. . 🚀 At first, it felt almost sci-fi: The LLM’s context window became its short-term memory. External storage turned into long-term memory. And the agent itself? It was like an operating system, intelligently deciding what to keep, what to forget, and when to retrieve information. 🧠 The highlight for me was experimenting with MemGPT-style agents. These aren’t just chatbots — they can self-edit their memory, update knowledge, and even collaborate with other agents using shared memory. Suddenly, “AI memory” wasn’t just storage — it was strategy. And the “aha!” moment? When I saw how selective memory works. It’s not just about remembering — it’s about remembering responsibly like an AI that recalls patterns but knows when to redact personal details depending on security or authority. That shift made me realize that memory is the key to trusting AI. This course didn’t just teach me a framework — it gave me a glimpse into the future of AI agents as true operating systems. Agents that don’t just answer, but grow with you. I walked away thinking: 👉 The difference between a chatbot and a real agent isn’t just intelligence — it’s memory. And now, I can’t stop wondering — how will memory-driven agents reshape industries like healthcare, education, or even project management? What do you think — are we ready for AI that truly remembers? #AI #AgenticAI #LLMs #ArtificialIntelligence #MemGPT #FutureOfWork

25

1 Comment

With a heavy heart as we mourn the loss of many of our fellow Iranians, I’d like to share a brief professional update. Our paper “GenForge: A Multi-population Genetic Programming framework with Semantic-Preserving Feature Partitioning for classification and regression tasks” has been published in Software Impacts. 🔹 Paper (DOI): https://lnkd.in/ggfdySDb 🔹 Code (archived version used in the paper): https://lnkd.in/gUyhyDZD 🔹 Main repository + documentation/manual: https://lnkd.in/guhSeiZ9 This work was carried out as part of my PhD research at UTS. I’m grateful to my supervisor Prof. Amir H Gandomi, and to my co-authors Navid Yazdanjue, Hassan Gharoun, A/Prof. Mohammad Reza Nikoo, and Prof. Fang Chen, for their contributions and support. GenForge is an open-source Python package for interpretable symbolic modeling via multi-population Genetic Programming (GP). It integrates Semantic-Preserving Feature Partitioning (SPFP) to build complementary, low-redundancy multi-view partitions from single-source data, and then evolves multiple GP populations (one per view) whose outputs can be fused through ensemble mechanisms for robust performance while keeping the final model transparent and parsimonious. 🔍 Key highlights: ✅ Unified symbolic learning for both regression and classification (gpregressor, gpclassifier) ✅ SPFPPartitioner for multi-view partitioning that supports view-driven multi-population evolution ✅ A consistent workflow (initialize → evolve → report → predict) with reproducible example scripts ✅ Built-in diagnostics and reporting utilities (including HTML summaries) to support explainable modeling and reproducible research 📌 Why does this matter? Many real-world problems need models that are both accurate and explainable. GenForge is designed to make symbolic, interpretable modeling practical in modern workflows by combining multi-view partitioning with cooperative multi-population GP, enabling robust learning while preserving interpretability. #GeneticProgramming #SymbolicRegression #ExplainableAI #MultiViewLearning #EnsembleLearning #Python #OpenSource #ReproducibleResearch #SoftwareImpacts #PhDResearch #UTS

44

6 Comments

The original Bf-Tree paper repository is now open. Not my work, but I presented it at Berkeley systems reading group last December. (https://lnkd.in/gqDUnP6X). Typically, pages in the buffer pool are mapped to the size of the unit of disk I/O i.e 4KB, buffer pool page beyond 4KB causes torn writes, not uncommon in Postgres, MySQL, e.t.c. but handled meticulously. Caching a 4KB page waste memory if only one record is hot. On the other hand, caching records alone does not help for range query since there is no *spatial locality*, i.e no guarantee that records in the cache constitute all the records within that range in the leaf page. Caching the entire leaf page i.e 4KB gives you *spatial locality* i.e you know all the content (records) of the page and their true order. Some databases handles this by allowing users keep both record and block cache e.g RocksDB. However, this introduces data duplication, cache coherence complexities and the paper also mentioned static partitioning of memory for each cache which is difficult to tune correctly as one cache might be underutilized while the other is starved. Bf-Tree introduces mini pages (an in memory page mapped to individual leaf pages on disk) which is dynamically sized according to the current workload. Mini page size is initially sized as 64 bytes (cache line size i.e smallest unit of data that can be transferred between DRAM and the CPU cache) and then grows dynamically. Point reads are probabilistically cached (20% by default), if key is not found, phantom record is inserted in mini page to prevent wasteful reads in the future. Insertion are absorbed into the mini pages (a new mini page is created if none exist for the leaf page). Read-Modify check the mini page before checking leaf page on disk, if found, update stays in mini page to serve future reads faster. Tombstone is used for deletion. Mini page size is doubled (beginning at 64 bytes) and when it grows beyond 2KB, it is merged with the content of leaf page on disk to become a 4KB mini-page that mirrors the leaf page (good for range scan). For range scan, the mini page might not contain all the keys therefore it has to load the entire page from the disk, it doesn't immediately cache the entire page in memory to prevent bloating the mini page with a single unrepeated range scan. To mitigate disk I/O, Bf-Trees allows a frequently scanned mini-page to grow to full page size i.e 4KB preserving spatial locality. Some limitations I see (not in the paper): First, the assumption for optimizing range scan is that lots of range scans will visit the same subset of pages but if many range scan access pages stochastically, I/O gets hurt. Secondly, cache misses reads the entire page (1st I/O), the page is read again during merge (2nd I/O) and then write the page to disk during cache eviction (3rd I/O), this would have been two I/O in normal B-Tree. No free lunch :) #database #rust

27

I am excited to share our recent work at Amazon Fulfillment Technologies & Robotics on generative foundation models for multi-robot coordination in Amazon fulfillment centers. What began with a simple curiosity: "can we train a model to predict robot motions autoregressively?" led to DeepFleet, a suite of four foundation models, trained on 5M+ hours of warehouse movement data and ranging from tens of million to nearly a billion parameters. DeepFleet pushes the frontier of large-scale AI into mobile robotics, and more broadly to physical phenomena that evolve over both space and time. This has been a Herculean team effort and we look forward to advancing these models for fleet-scale multi-robot applications. For details, see an overview here: https://lnkd.in/eNi7ZAVD and a technical description here: https://lnkd.in/evs-zG5S.

56

2 Comments

Excited to share my latest project: an Intelligent Hospital Management AI Assistant! The Idea I built a dual-interface AI system (Text & Voice) designed to streamline daily hospital operations. It acts as a 24/7 intelligent receptionist capable of registering new patients, booking/canceling appointments, and providing real-time hospital information. The Problem It Solves Administrative tasks often overwhelm healthcare providers, leading to scheduling bottlenecks and delays in patient support. This assistant significantly reduces administrative overhead and scheduling friction, allowing medical staff to focus on what truly matters—patient care. The Architecture The system uses a highly modular design: - Interfaces: Users can interact via a Text Terminal or real-time Voice Interface. - Core AI Agents: Handling complex workflows using stateful memory for text, and low-latency pipelines for voice. - External Services: Seamlessly integrated with LLM providers and specialized speech-to-text / text-to-speech APIs. - Tools Layer: Agents dynamically use custom Hospital Tools to interact with the database (Booking, Registry, etc.). - Data Layer: A structured relational database layer managing entities like Doctors, Patients, Appointments, and Calendars. Tech Stack - Backend: Python - AI Frameworks: LangChain, LangGraph, Pipecat - Voice AI: Deepgram (STT), Cartesia (TTS) - Models: Groq, OpenRouter, Google Models - Database: PostgreSQL (Dockerized), SQLAlchemy (ORM), Alembic (Migrations) #AI #HealthcareTech #LangChain #Python #MachineLearning #HealthTech #SoftwareArchitecture #VoiceAgent #AgenticAI https://lnkd.in/duAjSkfe

3