Research
Our publications
Selected work by our researchers in machine learning, language models and distributed systems.
No results for these filters.
Publication · 2025
[RE] GNNBoundary: Finding Boundaries and Going Beyond Them
In brief
This reproducibility study examines GNNBoundary, a method for understanding graph neural networks through their decision boundaries. It investigates how generated graphs near those boundaries can reveal what influences a model’s classification.
Why it matters for Traize
Traize connects people, projects and documents through relationships. Research into graph model behaviour brings useful expertise for evaluating how learning methods use such relationships and whether their decisions reflect meaningful structure.
Publication · 2025
DecoDINO: 3D Human-Scene Contact Prediction with Semantic Classification
In brief
DecoDINO adapts pretrained DINOv2 vision models to predict contact between people and objects. It combines targeted fine-tuning with task-specific components, adding semantic labels and improving contact prediction on the evaluated benchmark.
Why it matters for Traize
Adapting strong pretrained models is central to our research direction at Traize. This work contributes hands-on experience with fine-tuning, architecture choices and evaluation, relevant to our goal of developing models specialised for company tasks.
Publication · 2025
DreamerV3-XP: Optimizing exploration through uncertainty estimation
In brief
DreamerV3-XP uses disagreement between reward predictions to give an agent an internal exploration signal. Combined with prioritised replay, it helps the agent learn from useful experiences and reports faster learning on selected benchmarks, particularly when rewards are sparse.
Why it matters for Traize
The connection for Traize is choosing where to look next. This research offers ideas for training agents to follow promising sources and tool calls with less step-by-step guidance, making progress even when feedback on individual actions is limited.
Publication · 2024
Bilingual language model for protein sequence and structure
In brief
ProstT5 learns to translate between protein sequences and structural representations. The model connects two forms of the same underlying information and provides embeddings for further analysis.
Why it matters for Traize
Company knowledge also comes in different forms, including text and structured records. This work brings expertise in training language models across representations, relevant to how we make varied sources of information usable by Traize’s agents.
Publication · 2023
Fluidity: Location-Awareness in Replicated State Machines
In brief
Fluidity adjusts where replicas run as the location of demand changes. It measures network conditions and relocates replicas to reduce response times for most clients in the evaluated scenarios.
Why it matters for Traize
Every search and tool call contributes to the time an agent needs to answer. This research brings expertise in reducing infrastructure delays and adapting systems to usage, supporting responsive AI workflows as Traize grows.
Publication · 2021
Network Federation for Inter-Cloud Operations
In brief
This paper introduces NetFed, an approach to network federation for operations across cloud environments. It addresses the infrastructure needed to connect services across cloud boundaries.
Why it matters for Traize
AI for companies needs to fit into an existing technical environment. Research into cross-cloud networking brings useful experience for connecting services and reasoning about deployment choices as Traize’s infrastructure evolves.
Publication · 2021
SmartStream: Towards Efficient Byzantine Resilient Data Streaming through Speculation and Sharding
In brief
SmartStream studies efficient data streaming with resilience to faulty or malicious system components. It explores speculation and sharding as techniques for improving the efficiency of replicated streaming infrastructure.
Why it matters for Traize
Keeping company knowledge current depends on processing updates reliably. This work contributes systems expertise relevant to the flow of information between connected tools and Traize, where both throughput and dependable delivery matter.
Publication · 2020
Self-optimising Application-agnostic Multithreading for Replicated State Machines
In brief
The paper investigates self-optimising multithreading for replicated state machines. Its focus is improving concurrent execution without requiring application-specific tuning.
Why it matters for Traize
Demand changes as teams and automated workflows use a platform. Research into automatic performance tuning contributes expertise relevant to keeping shared services responsive while using available computing resources efficiently.
Publication · 2018
Resource-Efficient State-Machine Replication with Multithreading and Vertical Scaling
In brief
This work studies how multithreading and vertical scaling can make state-machine replication more resource-efficient. It examines how reliable services can make better use of computing capacity within their servers.
Why it matters for Traize
The cost of AI includes the services that store, retrieve and move information. This research contributes a resource-efficiency perspective to Traize’s goal of making complete AI workflows economical to operate.