DawnSecurity is building a next-generation cybersecurity platform. At its core is a hard applied-research problem, and this role owns it: leading our Data Science and Data Engineering team, directing algorithm research, building the models behind the platform, and driving the company's intellectual property.
What makes this role different
This is not a "train a classifier on labeled data" role. The core challenges are:
- Learning without labels. The events that matter most are too rare to train on. Models must be built from structure, statistics, and synthetically generated signal rather than labeled examples.
- Probabilistic modeling over graphs. The problem domain is inherently relational — entities, hierarchies, and connections — and the models must reason over that structure, not just over flat feature vectors.
- Adversarially robust learning. An adversary can influence the data our models learn from. Robustness to manipulation and governed model updates are first-class design constraints.
- Explainability as a requirement. Every score must decompose into a sentence a domain expert can act on. Opaque models that outperform explainable ones by a margin still lose here.
- LLMs as components, not oracles. We use LLMs inside pipelines with statistical quality gates and human-in-the-loop escalation — knowing when not to trust model output is the skill.
Core Responsibilities
Research & Modeling Leadership
- Own the research agenda: from hypothesis through prototype, empirical validation, shadow-mode deployment, and production.
- Design probabilistic and statistical models for sparse, high-cardinality event data — scoring, smoothing, calibration, drift detection.
- Lead graph-based modeling: entity resolution across heterogeneous sources, inference over large graphs, relationship discovery with tiered confidence.
- Combine statistical models with rule- and constraint-based reasoning where the two must agree.
- Define the evaluation discipline: synthetic data, shadow-mode comparisons, calibration audits, and precision targets tied to user trust.
IP Creation & Protection
- Drive patentable inventions from the research roadmap: identify novel mechanisms, validate them empirically, and work with patent counsel on claims.
- Distinguish what to file, what to publish, and what to keep as trade secret — and sequence research disclosures accordingly.
Engineering & Operational Excellence
- Architect the sandbox-to-production path: streaming and batch pipelines, feature stores, model serving with low-latency inference at high event throughput.
- Establish MLOps: versioning of code, data, and models; automated validation; monitoring that triggers governed retraining.
- Balance startup velocity against platform sustainability; keep compute costs proportionate to value.
Team & Cross-Functional Leadership
- Hire, grow, and retain a small, senior DS/DE team in a competitive market.
- Translate algorithmic tradeoffs into risk/reward language for founders, board, and customers.
- Partner with product and platform engineering so research lands as product, on a startup cadence.
Required Qualifications
- Statistical depth: MSc/PhD in a quantitative field, with demonstrated ability to design models under label scarcity — probabilistic modeling, calibration, uncertainty quantification.
- Graph & entity modeling: hands-on experience with entity resolution, graph inference, or knowledge-graph construction at scale.
- Adversarial thinking: experience with robustness, data poisoning, or security-relevant ML — or a demonstrated instinct for "how would an attacker game this model?"
- Applied LLM engineering: building production pipelines that use LLMs with quality gates, not chat products.
- Explainable ML: track record of shipping models whose outputs domain experts accepted and acted on.
- Production ML systems: streaming/batch data pipelines, model serving, MLOps — strong Python; distributed processing at large scale.
- Research-to-product leadership: managed teams through the uncertainty of research-heavy cycles while shipping on product deadlines; product-first bias toward the 90% solution that ships.
- Communication: translates algorithmic tradeoffs for non-technical stakeholders.
Strong Advantages
- Prior patents or publications, and experience working with patent counsel on claims.
- Background in an adversarial domain: security, fraud, abuse, anti-cheat.
- Experience with human-in-the-loop systems and governed model updates.
- Familiarity with formal methods (SMT, constraint solving) alongside statistical ML.
You apply on Workable, where this job was posted. Peter Nowak isn’t the employer and doesn’t check listings.