Alyssa Mia Taliotis
Mathematician & AI Engineer
Mathematician & AI Engineer
Alyssa Mia Taliotis is a mathematician and AI engineer building intelligent systems that reason, adapt, and act across industrial, scientific, and infrastructure domains. She holds a Master's in Data Science from Harvard University (4.0 GPA) with research conducted at MIT, and graduated as Valedictorian from the University of Manchester with a BSc (Hons) in Mathematics (First Class, 95.1%).
Her work spans deep learning, agentic systems, computer vision, reinforcement learning, embodied intelligence, and statistical inference. She has published, spoken, and built across leading machine learning venues, MIT Media Lab, and frontier AI systems. Her projects range from autonomous robotic architectures and CAD-driven manufacturing intelligence to multimodal neural networks and privacy-preserving AI.
Alyssa has built scalable memory layers for agentic systems, engineered full-stack AI platforms, and applied advanced modelling across logistics, finance, and industrial automation. She is a two-time Exness Fintech Scholar, an Enavsma Scholar, recipient of the Outstanding Academic Achievement Award, and the Mathematics Excellence Award — both presented by Dame Nancy Rothwell, President and Vice-Chancellor of the University of Manchester.
Originally from Paphos, Cyprus, Alyssa brings a disciplined, high-performance mindset shaped by competitive classical ballet and international leadership. Her mission is to build transformative, globally scalable AI systems shaping the next wave of intelligence.
[ICLR 2025 Submission · First Author]
Developed a method for adaptive boundary detection in neural fields, enhancing edge preservation and discontinuity modeling in 2D representations. The approach eliminates the need for explicit meshing, enhancing the efficiency and accuracy of neural implicit representations in computer vision tasks.
[ICML 2025 · Contributing Author]
Demonstrated that fine-tuning reshapes latent geometry and systematically disrupts targeted behavioural interventions — with direct implications for AI alignment and safety.
Analysed how generative AI is reshaping design across the full computing stack — from software and runtimes to hardware architecture and silicon. Research conducted within a seminar series featuring: Ofir Press (Princeton), Amir Yazdanbaksh (Google DeepMind), Sasha Rush (Cursor & Cornell Tech), Martin Maas (Google DeepMind), Suvinay Subramanian (Google), Jenny Huang (Nvidia), Milad Hashemi (Google), Esha Choukse (Microsoft Azure Research), Mark Ren (Nvidia), Richard Ho (OpenAI), and Kartik Hegde (ChipStack) — spanning ML for systems, TPU/GPU architecture, AI-driven chip design, and compute-model co-optimisation at scale.
Systematically evaluated the trade-offs between privacy guarantees, predictive utility, and demographic fairness in ICU mortality prediction models. Implemented and compared output perturbation, DP-XGBoost, and DP-SGD across varying privacy budgets (epsilon), analysing how differential privacy mechanisms degrade model calibration and disproportionately affect underrepresented subgroups — determining the clinical feasibility boundaries of privacy-preserving ML in critical care settings.
Developed an interpretable multimodal model to support clinicians in diagnosing paediatric appendicitis, leveraging concept-based reasoning to enhance transparency, reduce diagnostic uncertainty, and build trust in AI-assisted clinical decision-making.
Researched agentic AI-driven medical intelligence to optimize treatment strategies for patients with complex comorbidities, enabling cross-specialty coordination and personalized care. Conducted under mentorship of Mitsubishi Electric Innovation Centre.
Led ML pipeline development for automated prosthetic socket modification from 3D point clouds of transtibial limbs, in collaboration with Rise Bionics. Designed a multimodal deep learning framework leveraging clinician-generated annotations to produce scalable, personalized prosthetic designs — reducing design time and cost for lower-limb prosthetics globally.
Applied Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO) to train AI agents for strategic gameplay in Gomoku. Designed custom reward shaping and training environments to improve long-term decision-making, enabling the agent to learn competitive, human-level strategies through self-play.
[Python, PyTorch, Gym, Deep Q-Network (DQN), Proximal Policy Optimization (PPO), NumPy]
Built an ensemble learning framework to predict heart disease risk across the European population, incorporating geographical and demographic data. Leveraged model interpretability techniques to identify region-specific risk factors and key clinical predictors, enabling more targeted public health strategies.
[Python, Scikit-learn, XGBoost, Random Forests, Pandas, Matplotlib, Geopandas]
Core developer on the MIT team behind NANDA (Network of Agents and Decentralized AI), the world's first Internet of AI Agents. NANDA defines a standard protocol (MCP) and decentralized memory infrastructure to enable autonomous, modular, and interoperable agent ecosystems. The initiative pioneers AI-native internet architectures supporting real-time collaboration, memory composition, and identity persistence.
[JavaScript, TypeScript, Python, FastMCP, SSE, Starlette, Uvicorn, Claude, LangChain, JSON Routing]
Built an AI-powered logistics planning platform for real-time, multimodal freight routing. Tavi lets users input origin, destination, product, and priority (e.g., fastest, cheapest, most sustainable) and returns optimised shipment routes using a graph-powered backend. The system integrates live transport data (e.g., port status, rail availability) and supports disruption-aware rerouting while preserving original route plans for comparison.
[React, Tailwind CSS, React Leaflet, FastAPI, CrewAI, GeoJSON, OpenStreetMap, Multimodal Graph Routing]
Built a lightweight, modular memory operating system that supports agent interoperability across local and distributed settings. Enables persistent, queryable memory containers with identity tracking and customizable data retrieval for both LLMs and agentic workflows.
[Python, FastAPI, SQLite, JSON]
Built a multi-agent SQL pipeline for structured wearable data: specialised agents determine data requirements, write and execute queries against live databases, synthesise natural language responses, and enforce safety constraints across multi-session streams. Developed in collaboration with WurQ.
[Python, SQL, Multi-Agent Orchestration, LangChain, FastAPI]
Production-scale personalised education platform: users upload their own materials and receive AI-generated flashcards and a Socratic tutoring agent with spaced repetition scheduling based on the FSRS algorithm. Deployed on GCP and Kubernetes.
[Python, GCP, Kubernetes, FSRS, LLMs, FastAPI]
Developed multiple plug-and-play MCP-compliant servers to demonstrate specialized agent workflows:
[Python, FastAPI, Starlette, SSE, Claude, JSON Routing, REST APIs]
Developed a full-stack, open-source web application that enables patients to input their medical history and receive real-time feedback on potential medication interactions. Designed to improve medication safety and accessibility through an intuitive interface and intelligent backend processing.
[Next.js, React, TypeScript, Python, FastAPI, SQLite, Ngrok]
Engineered the backend architecture for an agentic AI system that analyzes patient histories and clinical reports to identify medical conflicts across comorbid conditions. Enabled nuanced detection of drug-condition conflicts and specialty-level insights using fine-tuned language models and structured EMR data.
[Python, CrewAI, LangChain, OpenAIEmbeddings, FAISS, RAG, SQLite, Pandas, JSON Caching]
Developed an agentic AI system for autonomous ultrasound triage and diagnostic guidance in space environments. The system leverages real-time reasoning, multimodal perception, and context-aware agents to assist astronauts with non-invasive diagnostics when direct medical supervision is unavailable. Designed for zero-gravity usability and resilient communication protocols.
[Python, OpenCV, Ultrasound Imaging, LangChain, Multi-Agent Orchestration, MCP, JSON Routing]
Partially developed at Harvard x Anthropic Hackathon 2025
Applied causal inference techniques to assess the efficacy of melatonin in a double-blind, placebo-controlled randomized clinical trial (RCT). Estimated average treatment effects while adjusting for potential confounders, enabling a robust evaluation of melatonin's impact on sleep outcomes in patients with primary insomnia.
Conducted time series analysis on Australian dry white wine sales using the Box-Jenkins methodology. Performed log transformation, seasonal differencing, and model identification to fit ARIMA/SARIMA models. Evaluated model assumptions via ACF/PACF and residual diagnostics, and validated forecasts against 1985 holdout data to assess predictive performance.
[R, ARIMA, SARIMA, ACF/PACF, Residual Diagnostics, Time Series Decomposition]
Analyzed the Old Faithful geyser dataset using Gaussian Mixture Models (GMMs) to uncover latent eruption patterns based on eruption duration and waiting times. Estimated mixture model parameters using the Expectation-Maximization (EM) algorithm and determined the optimal number of clusters using Bayesian Information Criterion (BIC). Compared GMM results with k-means clustering to evaluate modeling differences and cluster interpretation.
[R, mclust, MASS, Gaussian Mixture Models, EM Algorithm, BIC, K-Means Clustering, Data Visualization]