ADML - Abu Dhabi Machine Learning

From Deep Research To AutoResearch -
AI Agents from Evidence to Experiment

Can research agents move from finding knowledge to creating, and reliably validating, new knowledge?

Monday, 21 September 2026, 6:00–9:00 PM
MBZUAI (more information on exact location later...)

Research agents are moving beyond answering questions. They can now search and synthesize evidence, generate hypotheses, write code, run experiments and use feedback to decide what to try next.
This meetup explores the journey from Deep Research to AutoResearch: how these systems are built, where they create genuine value, how their results should be evaluated, and which parts of scientific judgement must remain human-led.
Expect practical architectures, real research workflows, live systems and an honest discussion of reliability, reproducibility and the limits of autonomous research.

Programme:

Talk 1: From Deep Research to AutoResearch: Building AI Agents That Experiment, Evaluate, and Discover

Abstract: Research agents are rapidly evolving from systems that retrieve and synthesise information into agents that can formulate hypotheses, write code, run experiments, evaluate results, and decide what to investigate next. This talk explores the architecture required to make that transition—from evidence to experiment—using NVIDIA's agentic AI stack and an AI Quant Researcher as a concrete example. We will examine how reasoning models, retrieval, agent orchestration, accelerated data science, evaluation, and secure execution can be combined into an iterative research loop: question → evidence → hypothesis → experiment → evaluation → refinement. The session will include a practical architecture and live demonstration, while discussing reproducibility, provenance, hallucination, and evaluation, as well as where human scientific judgement remains essential.

Bio: Amine El Khair is a Senior Solution Architect at NVIDIA, specializing in AI solutions for the Financial Services Industry. He works on advanced AI systems spanning generative AI, agentic workflows, retrieval-augmented generation, model optimization, and large-scale AI infrastructure. Previously a Senior Data Scientist, he has applied machine learning and deep learning to time-series, NLP, and industrial AI problems. His research interests include foundation models, agentic AI, quantitative finance, time-series analysis, network science, portfolio construction, and systematic investing. He holds an engineering degree from INSA Hauts-de-France and studied Machine Learning at Tsinghua University.

Talk 2: Training Deep Research Agents: From WebGPT to Sub-Agent Swarms

Abstract: Frontier deep research agents remain largely closed while open training recipes are scattered across a fast-moving literature. This talk offers a coherent picture of how research agents have evolved, from WebGPT's browser-assisted QA in 2021 through ReAct-style tool use and RL-trained search agents, to today's frontier systems in which orchestrators delegate long-horizon work to sub-agents. We then present a practical recipe for training a deep research agent in the agentic era in three parts: cold-start SFT on synthesized trajectories, reinforcement learning in an offline as well as online retrieval environment and inference-time harnesses for context management and sub-agent orchestration. Along the way, we share the design choices that matter most, grounded in benchmarks such as GAIA, BrowseComp, HLE, xBench-DeepSearch and WebWalkerQA.

Bio: Suhail is a researcher at the Technology Innovation Institute (TII), Abu Dhabi, where he works on large-scale LLM post-training. Previously, he was an AI Engineer at Google, working on AI privacy and security. He received his PhD from the Indian Institute of Technology Bombay (IIT Bombay) and has held research positions at the University of Texas at Austin and HKUST.

Talk 3: Trustworthy Deep-Research Agents: Grounding Evidence with GraphRAG, Multi-Hop Reasoning & Source Attribution

Abstract: Deep-research agents are only as useful as the evidence they can defend. I'll walk through the architecture behind VeritasGraph, an open-source GraphRAG framework where every answer is traceable to a source — combining vector retrieval with a knowledge graph for multi-hop reasoning, end-to-end source attribution, and concrete hallucination-control patterns. I'll cover what it takes to move a research agent from demo to dependable: retrieval strategy, reasoning traceability, human-in-the-loop verification, and reproducibility. Includes a live demo; vendor-neutral takeaways. ~25–30 minutes.

Bio: Bibin Prathap is an AI Strategy Leader, Microsoft MVP, and IEEE Senior Member based in Abu Dhabi, and the creator of the open-source VeritasGraph GraphRAG framework. Over a decade he has delivered enterprise-secure, explainable AI for regulated sectors, including the AI transformation of a sovereign entity's .