Srdjan Marinovic

I am a computer scientist with startup experience. My areas of interest are distributed streaming systems, information retrieval, runtime verification, formal security models, non-monotonic AI, graph databases, and applications of ML to graph analytics.

My software architectures are inspired by Parnas' information-hiding principles and layered virtual machines. My teams build iterative models and deploy fast: building complex systems is a process of exploration and discovery.

I am currently at PwC where I lead the development of Signal Graph data platform (see below) as well as engage with our clients in designing and implementing cloud-based data and analytics platforms. I also serve as an ML advisor to Futurae on fraud-detection and adaptive-authentication models. In 2020, I worked with Fibr Bank on implementing an AWS-based data lake solution with strong integrity and forensics requirements.

I was the technical co-founder and the CTO of SignalFrame (acquired by PwC in 2021, founded in 2015). My team of ~20 engineers (UI, mobile, backend, ML) and I built a streaming temporal graph platform. Our system collected data from ~1 billion IoT devices per month, and ingested ~100 million vertices with ~1 billion edges per day. The platform built graph embeddings in a streaming mode for temporal and spatial predictive analytics. My technical work was mostly focused on:

• Temporal graph analytics
• Streaming framework with feedback loops (Golang)
• Distributed graph query system (Golang)
• Graph-pattern classifiers (Py and Scala)
• Knowledge-base with Markov(-ish) semantics
• Monitoring and automated recovery of Kubernetes services

Prior to SignalFrame, I was a senior researcher at ETH Zurich's Security Institute:

• Database Security and Privacy – preventing probabilistic attacks on SQL DBs. Our algorithm used a Bayesian network synthesized from a probabilistic logic model and attack priors.
• Distributed Systems Monitoring (with Google) – metric temporal logics and streaming algorithms that handle missing and compromised system logs.
• Decentralized Authorization Systems (with Kaba Security) – tools to formally specify and verify authorization policies in decentralized settings. Our language is functionally-complete and captures ABAC, delegations, and compositions.

I received a PhD in Computing from Imperial College London. I developed a soft symbolic AI model (based on multi-valued logics) to ensemble unreliable and mutually-conflicting inputs. We applied this model to verify adaptive security models that enforce break-glass access control.

A few talks and papers

• SignalFrame's GeoGraph COVID-19 Model (Whitepaper; 2020)
• Modelling Temporal Graphs and Embeddigns -- A Practicioner's Approach (DC DataSci Meetup; 2020)
• Declarative Concurrent Programming in Streaming Systems (DevFest DC; 2019)
• Commune - Shared Ownership in an Agnostic Cloud (IEEE Trans. Information Forensics and Security 13(12) 2018)
• Go For Programmers (One-day school) (Washington DC; 2017)
• Securing Databases from Probabilistic Inference (CSF 2017)
• Monitoring of Temporal First-order Properties with Aggregations (Formal Methods in System Design 46(3): 262-285 (2015))
• Go Patterns for Cassandra (Golang Meetup DC; 2015)

A full list of my research papers is on DBLP.

Patents

• Validating human mobile users with wireless signals
• Locating or tracking devices using proximal groupings of devices
• Generating and updating proximal groupings of electronic devices