Adam Slucki

Client delivery · retail

Agentic conversational commerce

An agentic assistant that replaced static questionnaires with a conversation. I built its multimodal capabilities end to end: understanding customer-provided images, generating images, and semantic retrieval over the product inventory, exposed as tools to the agent and integrated with the application built by the backend team.

The work combined LLM agents and orchestration, vector search and image-generation models, with evaluation of both visual and textual output.

Outcome The solution was taken to production.

Client delivery · retail

Fresh produce recognition at the point of sale

A computer vision system that recognises fresh produce at the checkout and integrates the result with the store’s point-of-sale system. I was responsible for the vision side end to end: choosing the camera hardware, building the recognition model and developing the inference application connected to the POS.

Outcome Deployed in a real store.

Client delivery · food manufacturing

Industrial visual quality inspection

Technical lead and main technical contact for an inspection system for products on a production line. The system combines sequenced lighting control, segmentation, object cutout and 3D projection on GPU edge hardware integrated with industrial I/O.

Beyond the architecture, I built the annotation tooling, ran annotation together with the client’s quality inspectors, owned evaluation, mentored the engineer developing the segmentation models and coordinated integration with the client’s IT.

Outcome Deployed as a pilot in a factory.

Internal R&D

Multi-agent knowledge system for incident analysis

A research platform for answering analytical questions over a public corpus of workplace incident reports. It uses advanced document processing and indexing, graph-based retrieval, MCP tools and hub-and-spoke orchestration of specialised agents, with automated evaluation of retrieval quality.

Outcome Used internally to develop and test agent and retrieval patterns.

Open source · main contributor

WildNLP: robustness testing for NLP models

A Python library that corrupts input text in controlled, composable ways (misspellings, keyboard typos, OCR-style character errors, punctuation and formatting changes, and more) to measure how NLP models behave on noisy, real-world input, with ready-made pipelines for standard benchmark datasets.

Outcome Published on PyPI as wild-nlp.

Applied research · social media analytics

Predicting audience retention in social media video

A model that scores every fragment of a video by its predicted drop in viewer retention, so creators can find the shots likely to lose their audience before publishing. It learns from aggregated player interaction data, with drops normalised into percentiles so that videos of different lengths can be compared.

The model reads the video as a sequence, combining the frames with the text overlaid on them, extracted with a pipeline I published at ICCVG 2018. Instead of relying on error metrics alone, we validated it against what stakeholders care about: views.

Outcome Videos the model rated as engaging had markedly more organic views than those it rated poorly.

Client delivery · several projects

Computer vision across industries

Long-term hands-on and technical-lead work on image enhancement, product recognition and industrial inspection, combining deep learning with classical vision, 3D and point-cloud processing, and deploying models on device with TensorRT, on mobile and on proprietary hardware.