Research

Research notes, outputs, and selected engineering work around reliable AI systems, evaluation, and research infrastructure.

Research Profile

My research interests sit between reliable AI systems, empirical evaluation, and software engineering for AI. I am interested in how AI systems can be understood, modeled, and evaluated in ways that lead to robust systems and credible scientific insight. That includes certified robustness and randomized smoothing, but also the broader question of how technical metrics connect to practical usefulness, reproducibility, and responsible deployment.

Research Output

  • Rethinking Certified Training Efficiency: A Benchmarking Framework and Empirical Investigation for Randomized Smoothing

    Empirical work on randomized smoothing, certified robustness, controlled benchmarking, and training-efficiency interventions.

    This work studies the practical cost of certified robustness. The core thread is to turn a broad reliability question into controlled experiments: benchmark certified-training methods under shared conditions, then evaluate interventions such as LossProj and Selective Backpropagation with attention to both certified quality and training cost.

    The framing is intentionally evidence-bound. BenchmarkCtrs exposes cost and robustness trade-offs; LossProj is diagnostic and method-dependent; Selective Backpropagation reduces training cost while preserving competitive certified quality under the evaluated settings. The broader lesson is that reliable AI work depends on careful problem formulation, sound evaluation, and reproducible tooling as much as on the model itself.

Projects

  • API Management And LLM-Assisted Mocking Service

    A research-facing internal service for API management and service-interface prototyping with an agentic LLM-based development assistant.

    At the August-Wilhelm Scheer Institute, I developed an internal API management and mocking service with an agentic LLM-based development assistant. The goal was to support faster prototyping of service interfaces and integration workflows for research-facing AI and cloud projects.

    The project combined service API design, frontend work, backend implementation, and LLM tooling. It fits the practical side of my profile: translating open-ended technical requirements into internal tools that make experimentation and integration easier for other people.

  • Odoo-Backed Publication Tracking And Cloud Cost Reporting

    Internal Odoo-backed workflows for publication logging, automated metadata filling through n8n, and reporting cloud costs from heterogeneous providers.

    At the August-Wilhelm Scheer Institute, I worked on Odoo-backed internal systems that connected operational data with research and business reporting workflows. One part supported publication logging and tracking, including automatic data filling through integrated n8n workflows.

    Another part focused on aggregating and reporting cloud costs from heterogeneous providers. The work required modeling infrastructure and publication data, designing data flows that could be understood by non-specialists, and connecting operational visibility with systems that teams already used.

  • Federated Learning Infrastructure

    Deployment and maintenance of a federated learning server instance for distributed ML experimentation.

    I deployed and maintained a federated learning server instance for distributed ML experimentation. This gave me practical experience with infrastructure around privacy-aware and distributed training, beyond the model code itself.

    The work sits close to my research-engineering interests: ML experiments need reliable infrastructure, visible assumptions, and workflows that make the state of an experiment inspectable by the team using it.

  • Enigma Virtual Assistant SaaS

    Backend architecture and product delivery for an NLP-powered virtual assistant SaaS integrated with social messaging channels.

    At EnigmaAI, I architected backend services for the Enigma Virtual Assistant SaaS. The system connected NLP-powered chatbot workflows with social messaging channels, including Facebook Graph API and WhatsApp Business API, and supported tenant-facing SaaS behavior.

    The project combined backend architecture, data-flow design, ML-related implementation, and startup technical leadership. I also led a small engineering team through product design, implementation, and delivery cycles while contributing directly to backend and ML-related development.