M:3L Lab
Home / Agentic & Conversational AI Systems

M:3L research

Agentic & Conversational AI Systems

M:3L develops tool-using AI systems that retrieve evidence, remember time-sensitive events, coordinate specialized tools, and support grounded conversations.

Overview

M:3L combines language models with retrieval-augmented generation, semantic search, persistent memory, and tools exposed through Model Context Protocol (MCP) servers and APIs. The goal is to build systems that retrieve evidence, keep useful context, and act through traceable workflows rather than only generate text.

For conversational AI and agent memory, M:3L studies event sequences that capture what happened, when it happened, and which outcome followed. These representations support time-aware retrieval, monitoring, next-action prediction, and grounded investigation of changing systems.

Engineering and scientific agents are another central direction. They can coordinate CAD and CAE tools, numerical simulation, finite element models, optimization routines, validation, and iterative design while retaining explicit feedback and verification steps.

M:3L also partners with hackathon.ngo on BitGn PAC, a global challenge focused on autonomous and trustworthy AI agents with safety, reliability, and trust at the core.

Demonstrated M:3L research

  • A foundational event-embedding and Semantic Vector MCP project for time-sensitive retrieval, agent memory, and conversational investigation.
  • Partnership on BitGn PAC, a global challenge for autonomous and trustworthy AI agents.
  • An end-to-end LiteReality-Agent experiment in which an agent reconstructs and iteratively corrects a coded 3D scene from a scan of the M:3L laboratory.

Technical capabilities

  • Tool-using single-agent and multi-agent systems
  • Grounded conversational AI with retrieval and semantic search
  • Event-aware memory for time-sensitive investigation
  • Outcome-aware event and sequence representations
  • MCP and API integration
  • Simulation, optimization, and validation workflows

Event-aware memory and conversational investigation

Useful agent memory must preserve more than isolated text. M:3L is developing event representations that retain semantic meaning, temporal order, and downstream outcomes so an agent can investigate how a system changed and why a result followed.

  • What happened: semantic event identity
  • When it happened: temporal order and spacing
  • What followed: outcome, value, or system-health signals

Agentic Engineering

M:3L treats agentic engineering as the orchestration of specialized tools and models rather than as a text-only assistant. A workflow may connect CAD and CAE systems, numerical or FEM simulation, optimization algorithms, engineering constraints, and model-validation steps.

This is an active convergence of the laboratory's capabilities. It should not be interpreted as a claim that every component is already deployed as one finished production platform.

  • CAD and parametric geometry tools
  • CAE, numerical simulation, and FEM
  • Optimization and inverse-design loops
  • Model validation and programmatic feedback
  • Iterative design with human or automated review

Research directions

  • Agents that orchestrate CAD, CAE, simulation, optimization, and validation tools
  • Grounded conversational investigation with time-aware retrieval and persistent memory
  • Event-sequence representations that preserve meaning, timing, and outcomes
  • Reliable multi-agent systems with explicit feedback and verification
  • Enterprise workflow automation using MCP servers and APIs

Potential applications

  • Research automation
  • Engineering design iteration
  • Simulation campaign orchestration
  • Optimization workflows
  • Enterprise process automation
  • Monitoring, diagnostics, and next-action prediction

These are relevant application domains, not claims that every application is deployed.

Connected research

Related projects and news

Zero-Shot 3D Reconstruction with LiteReality-Agent

We tested LiteReality-Agent end-to-end on a scan of our laboratory to explore how code-driven agents can move 3D reconstruction beyond visual appearance toward structured, interactive scenes. A deterministic base pass first reconstructs the empty room, then an agent edits a single Python scene file containing geometry, materials, and articulation. It compares the generated scene with the physical scan, measures discrepancies, and iteratively corrects the code. This is an exploratory step toward useful scene models for design, robotics, and engineering; some objects still remain meshes rather than full CAD models.

View original post on LinkedIn

Enhancing Event-Sequence Embeddings for RAG Systems

M:3L is developing a foundational event-embedding model that jointly captures event meaning, temporal progression, and outcome/value signals. The work targets time-sensitive retrieval and agent memory across domains, supporting anomaly and churn detection, next-action and value prediction, causal and counterfactual analysis, and grounded conversational investigation through a Semantic Vector MCP. This work was conducted within the scope of the Faculty Research Funding Program 2025 of the Enterprise Incubator Foundation (EIF).