Projects & Research

A building block repository for ongoing technical explorations.

Applied Mathematics Research on Markov Dynamics, Stochastic Processes and Emergent Behaviour at the Mathematics Applications Consortium for Science and Industry UL ↗

From 2023 to 2025 I took part in a Maths Modelling Programme under the supervision of Professor James Gleeson at the Mathematics Applications Consortium for Science and Industry at the University of Limerick. I have retained an interest in using mathematics to understand the wider world employing a multidisciplinary translational approach to science, applying the principles of one scientific discipline to further another. Over the course of two years I took part in two year long group projects, deriving different insights from both of them. In my first year in the programme, I modelled forestry planning in the interest of the development of a self sustainable carbon sequestration model, this was one of my earliest experiences with social good innovation. During the programme my team and I modelled the life cycle from initial investment to maturity projecting more than 30 years into the future, accounting for initial investment cost, planting density, plant survival, market value per acre to outline how such a programme could be organised. As I previously had experience as the Secretary & Treasurer of my Greens School Committee, and had acted as a student representative in discussion with Green Party TDs this was an important experience in translating my appreciation for social impact to innovation. In the subsequent year the technical challenge increased as I tackled the design of stadiums for optimal emergency evacuation clearance time. During this time I gained an understanding of stochastic dynamics, noise that occurs when multiple actors occur independently but within the same environment, and dynamical systems that now form the basis of my work with Mesengenic AI. In crowd dynamics despite the apparent chaos, there are underlying potentials which govern movements in complex multifactorial environments, similarly in protein design I was interested in how the Evolutionary Density of a protein family, acts as a probability field biasing the mutational flow to regions of the fitness landscape that are high function but unsampled by nature, thus this experience was critical in helping me form the ground truth thesis of Mesengenic AI. It was also important in helping me contextualise evolution in a way which permitted me to arrive at this thesis - viewing evolution as a constrained stochastic walk. This was an important experience for me in contextualising my perception of the world, it demonstrated to me that through innovation I could make advancements in regards to areas that were of value to me, and was critical in teaching me to deploy my pre-existing skill sets to interpret the world around me, particularly when navigating unfamiliar technical concepts, utilising this novel inferential approach I was required to deploy as my edge and strength.

Python
PyTorch
VAE

Mesengenic AI - Some Highlights: Custom PDB File Parser for Structural Screening and Thermodynamic Stress Testing Integration Pipeline & Adopting a Hypothetico-Deductive Approach to Protein Design designing with KL Divergence, beta hyperparameterisation, and reparameterisation as the priority ↗

The most difficult technical challenge I’ve faced recently was building an integration pipeline between two core components of my Mesengenic industrial suite pipeline: Radar (structural screening via OpenMM) and Sisyphus (thermodynamic stress testing via ESMFold). Coming from a background in elegant science I appreciate the necessity of rigorously screening outcomes. However, running thermodynamic and steric screening on novel candidates under the time pressure of a hackathon is entirely different. The core bottleneck was raw structural data. Standard PDB files are littered with crystallised salts, excess hydrogens, and spatial artifacts that crashed our simulation environment. As no off-the-shelf solution existed, the pipeline stalled. I succeeded by writing a custom parser to systematically scrape and clean the PDB inputs.I approached this crisis with operational pragmatism—compartmentalising the problem and executing the required data engineering. This reflects my core character: I don’t just want to design elegant mathematical models; I am willing to build the gritty, unglamorous plumbing required to translate theory into reality. Within the generation of the candidates themselves one of the areas I have focused on most is the hyperparameterisation - the manner in which I can train the model based on negative results. I deploy candidates I synthesise through structural testing with ESM fold screening for structural hubris, and steric electron cloud clashing, as well as placing my molecules through thermodynamic testing at 350K to ensure that the epistatic peaks identified are stable under high temperature and pressure to ensure that they have a high preference to their folded conformational active state. However the most challenging aspects related to the adjustments of weights with reparameterisation and reconstruction loss, and ensuring my latent space remained regularised and sufficiently navigable for my explorative walkers. With reparameterisation the primary concern was ensuring that the results of each pass of the VAE circuitry provided the valuable feedback to adjust the weights used in the parameterisation of proteins to the latent space, without poisoning the model with “dirty” hallucinatory feedback. I achieved this by running each instance of the VAE in a different agentic sandbox ensuring the hyperparameterisation data from each pass of encoding and decoding could be isolated and screened before incorporating into the wider parameterisation process of the models. Managing the KL divergence was critical for ensuring that my fitness landscape was continuous, facilitating the interpolation required for the identification of epistatic peaks, whilst also providing me with a way to directly manipulate my vectorisation of each coordinate. The function of KL divergence is to punish the model for being overly explorative or exploitative, preventing the presence of any gaps in my manifold whilst also ensuring the presence of peaks - to identify the high importance regions. Taking this into consideration with the combined 128 dimensions of my latent space; managing this aspect of hyperparameterisation was likely one of my most conceptually and spatially challenging when developing the model architecture of Mesengenic. I overcame this by comparing the probability distribution of each of my vectorised coordinates to that of a standard normal distribution ensuring both my mean and variance of each coordinate was sufficiently high to create the rugged yet continuous landscape I required. To favour exploration I biased the parameterisation with KL divergence to favour slightly higher variance to facilitate greater overlap of latent space coordinates. In consideration of this, I structured my model to punish highly when hallucinated epistatic peaks were identified. Keeping my beta-hyperparameterisation low and monitoring my Categorical Cross Entropy to discourage the model from forming overlaps where they were not present.

Python
PyTorch
VAE

Mathematical Model of the Efficacy of a Mass Vaccination rollout during the Covid-19 pandemic considering the non-additive interactions of preventative measures. ↗

At the peak of the COVID 19 Pandemic in 2021, I like much of the world was confined to my home. With an interest in Medicine and how health policy can modulate health outcomes, I had for a long time a precursory interest in epidemiology. Consequently the pandemic, an event which was actively shaping the conditions of my childhood, offered an excellent opportunity for me to launch my largest project at the time, centering my thesis on the efficacy of a mass rollout of Covid-19 vaccines. I expanded my study to a sample group of 8 nations, selecting based on a range of variables such as rate of infection, diversity of viral strain, death rate and recovered and vulnerable population size, I accounted for preventive measures such as handwashing, mask wearing, social distancing and the relative risk reduction they extended to populations. Whilst it may appear intuitive that a mass rollout of vaccines would be effective in reducing infections, I found the period of time over which immunisation occurs and variations in the rate of infection that occur as programmes are put in place to be of greater significance than one might imagine in the efficacy of such programmes. This experience was particularly formative to me, as it acted as my initiating event. It proved to me that I had the capacity to use my knowledge to interpret, and understand the world around me and that I could tackle pertinent challenges and offer solutions. I worked on the project over a period of months, using the past year's data from 2020 to 2021 as a precedent for my model and used the modulating factors I identified to determine the current state of the pandemic if vaccines and preventative measures had been deployed from January 2020 onwards. Part of the reason why I didn't expand the model from being retrospective to speculative was in part due to a challenge I had not yet learned to fully overcome but now see as one of the largest pain points I am addressing in my new approach to protein synthesis. That simply being the isolation of variables. The effects of using masks, social distancing, administering vaccinations and handwashing are non-additive, being that when concurrent the combined effect of two or more variables will be less than that of the sum of the values in isolation. I was aware of this shortcoming and resultantly controlled for it by accounting for a reduction in the efficacy of my preventative variables when combined than when in isolation, this sufficed whilst I had the truth prior of the observed data when studying the past, but the when modelling into the future with case of both future total case speculation and anticipated protective effect, my outcomes would likely be the products of hallucinations. The same problem exists in protein design and is why I deem many of the current models to be biology “blind”. Current synthesis models are unable to determine that in amino acid pairs 1+1≠2. This variation in proteins is known as epistasis - the capacity for two amino acids when adjacent to lead to the possibility of both a major increase or decrease in function. The inability to recognise this is what keeps current models trapped as overly exploitative, only sampling known regions of the sequence space and dead-ended at local optima. To protect against this inability to navigate epistasis, they engage in greedy Markov Walks, only selecting for new sequences that also result in an increase in function. This however prevents the models from traversing the regions of uncertainty in the fitness landscape, and reaching the epistatic peaks - regions that they do not even have the capacity to conceive as existing - the global optima..

Python
PyTorch
VAE

TVEC - An Oncolytic Virus Innate Immune System CD8+ T Cell Activation via GM-CSF mediated by Dendritic Cell Maturation ↗

The research I conducted on Oncolytic Viruses namely TVEC and its applications to cancer treatment were important for communicating to me the applications of innovation to health and lifesciences, the bioengineering process and how natural processes and precedents can be a source of inspiration when designing for new to nature functions. In the years prior I had some exposure to immunotherapy treatments attending seminars and events relating to oncology, with a particular interest in CAR-T Cell treatment at Trinity College Dublin. I was most interested in the capacity for the virus to combat cancer indirectly by stimulating an inflammatory cascade within the host’s immune system that resulted in the destruction of local cancer cells. I was fascinated by the capacity for the immune system, what I recognise as one of the largest organisation problems, to demonstrate the capacity for collective-behaviour and self-organisation. These are areas I believe to be current crisis points in the construction, maintenance and continued growth of Neural Networks and wider artificial intelligence. I would argue that if a model does not have the capacity for self regulation, organisation and both unified and disparate actions under local control, then it cannot be deemed truly intelligent. This is an area I would like to advance, as I progress Mesengenic AI and validate the target agnostic architecture expanding to synthesis in single cell biology. In the following months I was invited to University College London’s World Cancer Day Event “Transforming Cancer Care: UCL Research Innovations and Discovery”, during which I learned about the applications of AI in Medicine, particularly in its application in the development of immunotherapies and the manners in which biological systems can be engineered to achieve new behaviours.

Rust
WASM
AWS

F1 in Schools - Project FireTyre National Finalists ↗

F1 in Schools was one of my earlier experiences operating in an enterprise style competition, and I took it on as I was interested in how innovation occurs when it is influenced by wider market dynamics such as the availability of funding and development resources. I operated as the enterprise and resource manager of the project, establishing budgets and managing finances, developing and managing sponsor relationships, and sourcing development resources such as access to 3D Printers for prototyping, acquiring modelling software for development and aerodynamic testing, and identifying manufacturing partners for our competition build. We progressed to the National Finals, competing at the 2023 Finalists Exhibition at the University of Galway. What I enjoyed most was the “team of builders” energy we expressed as we worked together, confirming my appreciation for the structure and operation of start-ups as my academic curiosity surrounding the problems they solve..

Python
PyTorch
VAE