London AI, ML, and Computer Vision Meetup - July 27, 2026

Join our in-person meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

Data curation is often described as drudgery, but an agent can turn it into an interactive, puzzle-solving process. This talk walks through a data-centric computer vision workflow built on the FiftyOne platform and the FiftyOne Agent, which uses skills (step-by-step instructions for tasks) and the Model Context Protocol (MCP) to handle dataset management, model inference, and brain computations through natural language — querying metadata, visualizing class distributions, and running complex operations without writing manual code.

The demo applies the workflow to a crop and weed dataset of 8,034 images and more than 112,000 labels (bounding boxes, keypoints, and segmentation masks). The agent diagnoses severe class imbalance and duplicate frames, builds interactive histograms and dashboards to surface domain shifts like lighting variation, finds invalid annotations so they can be corrected directly in the FiftyOne App, then automates exporting the curated dataset, fine-tuning a YOLO model, and evaluating its performance. The result is a repeatable pattern — ingest, explore, improve, fine-tune, evaluate, iterate — with a human in the loop through an intuitive, agentic interface.

In this talk, we'll explore reasoning about images from a material-centric perspective, namely through the lens of material understanding. Materials distinguish themselves by their response to light, which is governed and modelled through physical properties like roughness or gloss - however, understanding such properties is a non-trivial task for current algorithms and models.

We'll see how we can select materials similar to a given query material, significantly improve selection fidelity and eventually even venture beyond 2D, to enable selection in the 3D domain.

Lighting has a strong influence on visual appearance, yet understanding and representing lighting in images remains notoriously difficult. UniLight introduces a joint latent space to unify previously incompatible lighting representation - environment maps, images, irradiance and text descriptions.

Modality-specific encoders are trained contrastively to align their representations, with an auxiliary spherical-harmonics prediction task reinforcing directional understanding. Our joint lighting embedding enables applications such as retrieval, example-based light control during image generation, and environment map generation from various modalities.

A candid lessons-learned from running an agentic engineering consultancy with clients ranging from federal governments to early-stage AI startups. I'll cover what's held up under real production pressure, what I tried and abandoned, and the approaches that are quietly dead but still being sold. Expect specifics, opinions, and a few uncomfortable conclusions.

Tokenization is fundamental to generative modeling and especially important for autoregressive 3D generation. However, current 3D shape tokenizers rely on geometric level-of-detail hierarchies that are token-inefficient and poorly aligned with semantic structure.

We propose Level-of-Semantics Tokenization (LoST), which orders tokens by semantic salience so early tokens produce complete, plausible shapes and later tokens refine detailed geometry and semantics. LoST is trained with Relational Inter-Distance Alignment (RIDA), a semantic alignment loss that matches relationships in 3D shape latent space to those in DINO feature space.

Experiments show that LoST achieves state-of-the-art reconstruction and efficient high-quality AR 3D generation while using only 0.1%–10% of the tokens required by prior methods.