We're specifying a text and image annoation web service (tentatively named annotate.io) which spots brand mentions and accurately disambiguates names (e.g. finding Apple-the-company and ignoring apple-the-fruit) and detects brand logos (e.g. spotting the Starbucks logo in an Instagram photo). Please contact us if you're interested in learning more, contact details in the footer.
Working with AdaptiveLab for YouGov we're tagging Twitter and Facebook posts using Named Entity Recognition to mark-up brand discussions with sentiment for brand reach analysis.
Ian has taught a mix of scientific python tutorials in recent years:
Parallel Python at EuroSciPy 2012 covers 5 ways to move towards parallel python across a local and cloud based network, includes src and slides.
High Performance Python 1 at PyCon 2012 covers 9 approaches to making a pure Python program faster including profiling, numpy a compiling to C, includes src and slides and video.
High Performance Python at EuroPython 2012 covers 10 approaches to making pure Python faster (as for PyCon tutorial), includes src and slides and video.
Working with Kasabi (of Talis) we created the OpenPlants plant-label matcher for use at the Royal Botanic Gardens, Kew, London which allows iPhone users to photograph a plant label and get Wikipedia information back in return. This removes the need to add a QR code to the 10,000 plant labels at Kew Gardens.
Our forthcoming StrongSteam.com API exposes many of our tools for use by non-AI developers over a simple web interface. The goal is to let web and mobile developers develop powerful algorithms without having to go through the time consuming process of learning new libraries (along with compiling, supporting and integrating them!), backed by a scaleable web service. Follow us on Twitter via @StrongSteamAPI.
See these example videos showing the first API for noisy OCR (optical character recognition) matching which removes the needs for QR Codes.
At EuroPython 2011 and PyCon 2012 Ian has taught High Performance Python, this material has been published as a free eBook and may be turned into a larger eBook later. Source, tutorial video from PyCon, slides etc available via the link.
We've been working with InvisibleHand to analyse how humans determine if e-commerce pages represent the same product so we can add human-like reasoning to their superb price comparison toolbar.
We've powered the artificial intelligence for text mining behind the social discovery app SocialTies, the app runs on smartphones to help you discover interesting people at events.
Working with RadicalRobot we've created a mobile application powered by a Python web API that constantly models everyone it has learned about, backed by a model of 'conversation space' and a social graph, to efficiently help conference goers meet interesting people.
Professor Paul Fewster's Research Group at PANalytical work with X-Ray Diffraction techniques to analyse complex fabricated structures such as next-generation blue LEDs and laser diodes. Physics simulations are coupled with optimisation algorithms to automatically fit proposed structure models to the actual results that are measured using X-Ray spectroscopy equipment.
Example: The optimisation techniques were improved so that complex fitting problems were solved more quickly and more reliably. Problems faced included long computation times for complex simulations and non-obvious fitness landscapes (since the problems have high dimensionality). Approaches in use include pyCUDA, NVIDIA CUDA with Visual Studio, algorithmic design and optimisation.
As an experiment we visualised Lanyrd's social connectivity graph using Lanyrd.com and Twitter. The JavaScript visualisation runs in browsers and on mobile devices, it lets people see who they know at conferences and who is well connected. The code is open source.
Ian is working with Applied Machine Intelligence to develop low cost robotic systems for the masses. Topics include machine vision and models of human interaction.
With Ambiental We're assisting in the parallelisation of their flood modeling simulations - parallelisation is essential to take advantage of multi-core and multi-cpu resources so time intensive simulations take place quickly. Tools include OpenMP, Visual Studio and NVIDIA's CUDA. Platforms include 32 bit and 64 bit Windows and Linux solutions.
ProCasts wanted to automated elements of their educational screencast production, we identified ways of detecting pauses to automatically cut samples into shorter files which were matched against the original text script. Tools used include text to speech, audio analysers and Python.
Qtara.com are building intelligent virtual humans using technologies like face and voice recognition, natural language parsing and models of human behaviour. We advised on the creation of these agents.
pyFlux allows a Python programmer to talk to the API inside Cedrat's Flux magnetic flux simulation package. pyFlux has scant documentation so requires an inquisitive mind to find the right API calls. For the client I have worked to automate repetitive flux simulation and analysis tasks.
Using a Digital Elevation Map, the customer required a set of optimal routes for an oil pipeline across a mountainous area. Constraints placed upon the routing included maximum elevation limits, length of pipeline, preference to avoid changes in elevation and a desire to avoid unsafe areas.
These constraints represented a trade-off between cost and safety. The system would quickly generate a route between any two points, and these routes could have been used during a cost-benefit analysis for routing over a large terrain. Developed in conjunction with Dr. Justin Butler at Ambiental.
A large volume of news articles needed to be summarised quickly to bring breaking news events to the user's attention, preferably in real-time. After our involvement the demonstration system was extended into the ScoopJack news aggregation service and is also sold by Corpora.
The prototype was developed with Dr. Nick Jakobi during employment at Algorithmix Ltd. (the webiste used to be www.algorithmix.co.uk pre-2006). Algorithmix was bought by Corpora during 2004.
Example: A simulated model of an industrial waste collection operation was developed in conjunction with the industrial partner. Parameters to the model included the amount and types of waste to be collected, number and skills of each waste-truck driver and the location and capabilities of each waste-disposal facility.
An optimisation system was developed to optimise each parameter to produce robust, reliable work-plans for each driver, whilst respecting work-time constraints and final costs. The system has been in use since installation and produces more efficient plans than those generated by the unassisted human planner. Developed with Olivier Trullier and team whilst working with MASA Group.