AI Research Scientist · Technical AI Governance · Trustworthy & Responsible AI
I'm Dr. Syeda Shafia Rubbani, an AI PhD (Heriot-Watt University, 2025) and FHEA (UK). My doctoral work applied Shannon's information theory to edge-AI protocol optimization. My current work is the technical governance of multi-agent LLM systems: coordination risk, evaluation, privacy-preserving ML, and responsible deployment.
η = R / C
Achieved rate over channel capacity — the efficiency metric from my thesis, and the habit of mind I bring to AI systems: measure against the bound, not against the last result.
Four threads, one question: what can we actually know about a system before we deploy it?
How autonomous LLM agents interact, fail, and should be evaluated before deployment — grounded in building multi-agent pipelines end to end.
Differential privacy (DP-SGD, Rényi accounting), federated learning (FedProx, secure aggregation), and the trade-offs they impose on utility and latency.
Designing evals that measure what matters: hallucination scoring, high-stakes accuracy subsets, and LLM-as-judge pipelines with explicit failure taxonomies.
Shannon capacity as a lens for protocol efficiency in constrained systems — a doctoral foundation I now bring to reasoning about the limits of AI systems.
I treat deployable systems as part of my research identity: claims about AI governance should be grounded in what real systems actually do.
A five-agent decision-support platform (Interpret → Retrieve → Ideate → Evaluate → Recommend) that turns complex organizational problems into structured, evidence-grounded recommendations in roughly 40 seconds, with citation-grounded retrieval and multi-tenant vector namespaces.
Privacy-preserving federated platform for phishing and business email compromise detection: TinyLlama with LoRA multi-task heads trained under Opacus DP-SGD (ε < 5), federated via Flower FedProx with SecAgg+, plus an adversarial defense layer and threat-intel retrieval.
A benchmark testing whether general-purpose AI systems give voters accurate election-logistics answers — deadlines, ID requirements, ballot return rules. Three-way verdict scoring separates outright errors from safe redirects to official authorities, and distinguishes suppressive errors that cost someone a vote from recoverable ones. Reports accuracy on a high-stakes subset and official-source citation rate.
Clinical monitoring dashboard for the ICBHI 2017 Respiratory Sound Database, built for the Edge-AI Smart COPD Mask project — FHIR R4 patient summaries with explainability via SHAP and Grad-CAM, developed to IEC 62304-aligned practice.
A brain–computer interface to LLM thought-communication prototype: an EEG signal pipeline (bandpass and ICA, FFT/P300 features, EEGNet intent classification) translated to natural language via an LLM, designed for patients with severe motor impairment.
Agentic clinical framework generating evidence-grounded medical reports from sparse inputs: a nine-agent LangGraph pipeline with ICD-10 and SNOMED CT mapping, per-claim hallucination detection, PubMed and NICE citations, and human-in-the-loop physician review.
Defined a protocol efficiency metric η = R / C normalized against the Shannon–Hartley bound, and validated it across CoAP, MQTT, AMQP and HTTPS through neuroevolutionary optimization and MIMO frameworks — raising measured capacity utilization from a 50–70% baseline to 70–90% with neuroevolutionary search on resource-constrained edge devices. A second strand addressed security in distributed sensor networks: attack-surface identification, detection-test design, and quantifying failure modes under adversarial conditions.
A five-paper series is in preparation from the thesis, all single-authored. Targets below reflect current submission plans; statuses are stated honestly and updated as they change.
The η = R/C framework and cross-protocol empirical study for M2M edge communication.
Evolutionary search over protocol configurations against capacity, power, memory and handover objectives.
Theoretical treatment of capacity bounds under the mobility and MIMO regimes studied empirically in the thesis.
MIMO-based analysis of throughput, handover time, and stability during mobility.
Extends the thesis line toward learning-theoretic guarantees for remote estimation over constrained channels.
Three Research Square preprints on neuroevolutionary capacity optimization, data-stream harmonization, and Shannon's lens for edge-based machines.
Four peer-reviewed conference papers, including an IoT testbed for smart-city sensing, cluster-based routing in mobile ad-hoc networks, and a defence mechanism against internal attacks in wireless sensor networks.
Advising organizations on applied and generative AI; designing evaluation and validation protocols for LLM and multi-agent systems; building and deploying the multi-agent, privacy-preserving and clinically-applied systems above; continuing single-authored research toward publication.
Information-theoretic protocol optimization and security research on distributed sensor networks; real-world edge AI and IoT experimentation. Secured AED 50K research funding (Expo 2020 Dubai); elected Postgraduate Research Representative for 100+ doctoral scholars, serving on the Research Coordination and Research Degrees Committees; winner, Three Minute Thesis Competition.
Commercialization feasibility for AI-driven respiratory healthcare: 55+ structured interviews across clinicians, patients, manufacturers, regulators and government; competitive benchmarking of 20+ rival technologies; feasibility report identifying a £300M opportunity that informed an external investment decision.
Designed and delivered graduate courses aligned to UKPSF standards; supervised 30+ final-year projects, four subsequently published or presented at conferences.
Taught 650+ students across core computer science including information security and data mining; supervised 35+ student projects; co-authored conference papers with students and faculty.
Open to research, evaluation and technical governance roles in Switzerland, the UK, the EU and the UAE, and to consulting engagements on multi-agent architecture, evaluation design and responsible deployment.