Data Scientist

R&D Phase Data Scientist, Structural Intelligence & Predictive SystemsScier is building a decision architecture for organizations — a system that models how decisions cascade through a company and its markets, automates the artisan data science that today depends on a handful of rare experts, and compounds into flywheels and ambient systems that run in the background. The premise: enterprise data encodes the physics of how decisions, risk, and advantage propagate — dynamics that conventional schemas and content-based models discard. We surface those dynamics through deterministic graph computation, then either expose them directly or fuse them into model weights, and deliver structurally-grounded output. Meanwhile the market fills with strategic slop — LLM output that sounds like thought but contains none, plausibility mistaken for truth. As that output commoditizes, mechanistic, structurally-grounded methods become the one durable source of competitive advantage left. This role sits at the core of building them.WHO WE'RE LOOKING FORThe profile We are not hiring for a single archetype. The problems here sit at the intersection of network science, machine learning, causal inference, and applied domain intelligence — and the people best equipped to work on them tend to resist clean categorization. We value depth of thinking over breadth of credentials. If you are extraordinarily good at one thing — spectral graph theory, LLM fine-tuning, knowledge representation, systems programming — and you can apply that depth to real analytical problems, we want to talk. Equally, if you operate comfortably across the technical and applied layers and can translate between mathematical foundations and domain-specific questions in geopolitics, finance, or M&A, that profile is just as valuable here. There is an adjacent profile we are just as keen to find: engineers who can rethink the cloud integration and orchestration layers — how computed insights get embedded, served, and kept in sync across systems and APIs. That may be a separate role from the science above, and that is fine; if you live mainly on the infrastructure side, we still want to talk. Someone who can do both — the structural computation and the systems that deliver it at scale — is rare, and especially valuable here. Seniority is flexible. What is not flexible is the ability to think independently, work on problems with no established playbook, and hold rigor and pragmatism in the same hand.THE WORKThis is not a production data science role. You will be building Client components of a multi-layer pipeline spanning mathematics and statistics, traditional ML, and graph computation, with LLM fine-tuning and generative output as one layer among them rather than the center of gravity. The focus is embedding knowledge from real-world data and computing it in Client ways in Python, then serving the results through systems and APIs that deliver those insights with precision and depth — generative AI is one way to access or present an insight, but only one tool in the kit. The architecture below reflects our current working assumptions — the specific models, tools, and techniques named throughout are starting points, not commitments, and part of the R&D is discovering where the outcomes lead us to revise them — and where to focus our research, offerings, and product next. The challenge is making each layer work precisely, validating (or falsifying) Client theoretical assumptions against real data, and ensuring every component is reproducible and auditable. The draw is the application: you will aim genuinely cutting-edge algorithms — spectral methods, diffusion, community detection, graph-aware fine-tuning — at business and strategic problems most people never realize can be approached this way, and turn the results into something a decision-maker can actually use.Build and validate the graph Laplacian computation layer (L = D ? A) using cuGraph and NetworkX; maintain versioned spectral snapshots as first-class KG nodesBuild out the Analysis Matrix — a layered library of graph and statistical operations spanning dimensions such as structural, temporal, semantic, distributional, causal, and topologicalDesign multi-task fine-tuning pipelines using Unsloth + PEFT/LoRA on Qwen2.5-32B and equivalent base models; implement spectral regularization to preserve Laplacian geometry across retraining cyclesBuild the feature-gating and importance-scoring layer — gradient-boosted trees with SHAP (XGBoost) are one option, training NNs or GNNs for stronger output is another, and determining which approach wins is part of the work; measure where and by how much network topology features outperform content features in engagement and contagion prediction tasksConstruct and maintain the Knowledge Graph provenance ledger — entities, Laplacian nodes, feature matrices, prediction objects, and model checkpoints as versioned, queryable graph nodesImplement the tensor ontology hypergraph with typed hyperedges (Requires, Precedes, Produces, Constrains, Specializes, Incompatible, Informs, Contextual) and integrate the OntologyConstraintEncoder into LLM system promptsPrototype Rust compute kernels for the hypergraph database with PyO3 bindings; validate against existing Python workflows before committing to full persistence infrastructureIngest and harmonize multiple external data sources — OSINT feeds, alternative datasets, and client corpora — into the knowledge graph, resolving entities and reconciling schemas so the structure is clean enough to compute on; turn the harmonized substrate into MVP APIs and apps that put structural intelligence in front of real usersEngineer features and reduce dimensionality across heterogeneous inputs — market and economic indices, text corpora (via entity and relation extraction), transaction and network data — to represent knowledge and mechanistic influence in forms standard pipelines miss, producing Client structurally-grounded signal that gives clients and the product insight no competitor hasExecute pipeline deployments across geopolitics and global risk, FX/EM contagion, M&A target scoring, and brand influencer network domainsProduce Generative BI output — native HTML analytical documents generated from embedded model forward passes — and validate interpretation stability against conventional dashboard approachesREQUIRED BACKGROUNDGraph theory Spectral analysis LLM fine-tuning Community detection cuGraph / cuDF XGBoost + SHAP GNNs PyTorch + CUDA PEFT / LoRA / TRL Rust (optional) Knowledge graphs Complex systems Graph databases Mathematics / statistics Classical ML Topological data analysisDeep familiarity with graph-theoretic methods: Laplacian spectrum, community detection (Leiden, Louvain), betweenness and bridging centrality, random walk metricsHands-on LLM fine-tuning experience — not API prompting, but actual weight adaptation via LoRA on large base models (30B+ parameters), including quantization-aware trainingGPU-accelerated data science: cuDF, cuGraph, PyTorch on multi-GPU configurations; working knowledge of VRAM constraints and CUDA kernel-level limitationsBreadth across the graph-database landscape, mainstream and emerging — our working assumption is that the well-known engines (Neo4j, TigerGraph) are not enough on their own, so the persistence layer will combine them with nimbler, less-legacy technology — KGs such as Turing KG, TypeDB, or Pometry — and may need to be rethought or built from scratch. We want people who know how far each option can be pushed and where it breaks — and how far graph stores can be combined with vector databases (Qdrant and the like) before either gives outRigorous experimental discipline: TimeSeriesSplit cross-validation, reproducible seeds, clean notebook hygiene, pickle-based caching, W&B loggingComfort working on architecturally Client systems with limited prior art — you will not be able to Google most of what you build hereAbility to read and apply academic literature directly: complex systems, network science, spectral graph theory, topological data analysis, causal inference in observational dataPREFERRED DOMAINSUseful applied context We work across four production domains. Prior exposure to any of these narrows onboarding time considerably, but none is a hard requirement.Geopolitics & global risksFX / EM contagion & financial networksM&A target scoring & co-mention graphsBrand influencer networks & engagement predictionWHAT THIS IS NOTThis is not a frontier-model training shop. LLMs are strong at language and weak at computation, ranking, and aggregation — so we build hybrid Python-plus-LLM systems where the math is computed deterministically and the model is the interface. The edge comes from knowledge structuring and engineering — computing insights once and storing them so they can be recalled, reused, and return the same answer every time — not from retraining frontier models, which is neither efficient nor where the advantage livesThis is not a sprint-driven product team. R&D cycles are iterative and exploratory; some weeks are deep theory, some are debugging CUDA kernel errors at 2amThere is no large team absorbing risk — decisions you make on architecture propagate directly into