Work & Interests

During my career I’ve worked extensively on the topics below amd developed a thorough expertise in these areas. Within each area I developed unique software solutions that riched or surpassed the latest existing benchmarks.

My past and current workprojects are described in more details in the Current and Past Work section of the website.

Neuro Inspired Computing and Deep Learning

Applications of Generative Models

I brought generative models to production for weather predictions and transliteration:

With JUA, we adapted Fourier Neural Operators (FNO), Graph Neural Networks and Generative Adversarial Networks (GAN) to help JUA operate the currently most accurate precipitation weather model.

I am also currently working on porting the FNO architecture to text processing, a practical investigation on the possibility to process language as a continuous process.

For Interscript, I built generative model leveraging transformer architecture for not only transliteration but also semi automatised code production and data generation.

High Dimensional computing

Learning and computations can be accelerated using sparse binary vectors. I worked on various technologies exploiting the benefits of the HD approach:

  • combine semantic vectors with symbolic algebra
  • simpler operations mean compression and more efficient compute
  • graphs and rules encoding/modelling
  • Interplay HDC / Deep Learning

This work was applied in multiple business cases like text processing, patient profiles modelling as well as fuzzy GPS trajectory modelling and matching.

Applications of Language Models

I am interested in the theory, training, usage and integration of language models into my workflow as well as some of the systems described below.

Systems

Conversational Agents

Since 2016 I’ve been working on conversational agents, starting with my R&D work for Constellation AI.

IVA was developed with Markus Thora and Denis Okyere as an end to end IO app allowing speech interactions with a logic compressed into a brain using HD computing combined with a language model.

With the Humbaba projects and some of my work with Interscript, I’ve been working on the idea of software designed as a “Union of conversational Agents”. Ideas of this kind are becoming more realistic with the development of the newest language models (gpt’s, …).

The aforementionned approach has can potentially allow people with a little experience in software engineering to generate complex software as well as the design of truly smart systems.

I am working on:

  • conversational interfaces
  • graphical description of software (with simple technologies like lucidchart, etc…) and its automatised parsing and building
  • software design automation combining the above ideas with advanced language models

Collaborative Systems

TaxiQ is an app developed to give London’s cab drivers a real time overview of the taxi ranks. I also helped faxi (now kinto join) with recommendations, GPS processing algorithms and community evaluations via an app. highlights are IMO:

  • optimisation and behavioural simulations
  • efficient GPS modelling and retrieval
  • real time city modelling in c++

Fuzzy retrieval

High dimensional computing and encoding was used to model and retrieve matching GPS trajectories, text snippet and medical data.