Applied Scientist, Context Graph Team

About Skan Skan automates automation. We capture how work actually happens across an organization and autonomously turn that observation into automation.

We are now building toward agentic automation of that work: systems that operate over observed process data and generate Agentic Operating Procedures, structured and auditable playbooks mined from direct observation that allow agents to understand the steps, decisions, branches, and exceptions that define how work is actually performed.

We are a small and passionate team of researchers, engineers, and domain experts working on the next generation of agentic automation: self-discovered AI agents that can understand and operate real enterprise software.

About The Role

Before an agent can reliably operate software on a user's behalf, it needs to understand what should happen next and why: the underlying process, its steps, decision points and governing logic, branches, variations, dependencies, and long-tail edge cases. In practice, this knowledge is rarely represented explicitly.

It is distributed across messy event logs, unstructured documents, screen recordings, UI interaction traces, historical executions, and the tacit knowledge of domain experts.

As an Applied Scientist on this team, you will work across teams to develop process-agnostic, scalable learning systems that recover the latent structure, decision logic, and dynamics underlying complex process execution trajectories from sparse, noisy, highly correlated observations spanning substantial behavioral heterogeneity and organizational scale.

You will take heterogeneous data such as action logs, noisy screen traces, semi-structured documents, and variant-heavy workflows, and build systems that recover canonical process structure, meaningful variations, decision rules, dependencies, exceptions, and provenance.

These representations are consumed by downstream agentic execution systems and must preserve enough detail for an agent to act correctly while remaining structured, compact, transparent, and auditable. The same observational data and learned representations also contribute directly to our work on computer-use models for last-mile interaction with enterprise software. The problems here rarely arrive with clean formulations.

You may begin with a branch the system consistently fails to recover, two superficially different execution paths that should collapse into the same process, or an algorithm that works across thousands of traces but fails at millions.

Your job is to turn those failures into tractable research problems: identify the right abstraction, formulate hypotheses, design evaluations that isolate the problem, and develop methods that solve it.

Depending on the problem, this may draw from process and decision mining, graph algorithms, structured prediction, representation learning, information retrieval, program synthesis, model post-training and distillation, or agentic systems. We care less about allegiance to a particular technique than about finding the simplest method that reliably recovers the underlying structure and scales.

Responsibilities

• Process discovery and representation: Build pipelines that discover, extract, and formalize end-to-end business processes from event logs, UI interaction traces, screen recordings, and unstructured documents.
• Develop algorithms for process variant detection, noise removal, and trace consolidation, turning noisy, high-cardinality action sequences into clean canonical representations.
• Design representations that preserve the information downstream agents need while collapsing irrelevant execution-level variation.
• Study the tradeoff between abstraction and fidelity: determine when superficially different trajectories represent the same underlying behavior and when small differences encode materially different decisions.
• Build retrieval and grounding systems that incorporate SOPs, documents, and domain knowledge when process understanding requires evidence beyond execution traces.

Evaluation and reliability

Develop rigorous evaluation frameworks for process discovery and extraction, including LLM-as-Judge pipelines, conformance checking against ground-truth processes, and regression suites that catch subtle degradation. Develop evaluation methods for settings where there may not be a single canonical ground truth, including behaviorally equivalent process representations.

Enumerate and systematically test exception paths, conditional branches, loops, rare variants, and the failure modes that naive extraction approaches miss. Make deliberate algorithmic tradeoffs so systems remain tractable across millions of traces and thousands of variants, including choices around data structures, approximation, indexing, and level of abstraction.

Diagnose failures in production and turn recurring failure patterns into new algorithmic or research directions.

Collaboration

Work closely with the computer-use agents team so the representations you produce are expressed at the right level of abstraction, contain the necessary execution context, and are robust enough for downstream autonomous execution. Work across research, engineering, and product to move promising methods from prototype to production and measure their effect on downstream agent behavior.

Requirements

• Ambitious and Hungry: "Far and away the best prize that life offers is the chance to work hard at work worth doing." - Theodore Roosevelt.
• We screen for comparable experience and strong fundamentals, not exact tool matches.
• Equivalent technologies are welcome.

Problem Decomposition and Algorithmic Thinking

You can take a loosely defined problem, such as recovering a reliable process from messy action logs, decompose it into tractable sub-problems, reason carefully about edge cases, and arrive at an approach. You are comfortable reasoning from first principles about data structures, graph algorithms, approximation, indexing, and time and space complexity when algorithms must operate over millions of executions.

ML, DL Breadth

Strong working knowledge across representation learning, sequence modeling, clustering, structured prediction, information retrieval, generative modeling, and modern foundation-model adaptation. You have the judgment to recognize when a problem calls for a learned model, an algorithmic solution, an LLM-based system, or some combination of the three.

Cross-Domain Research Translation

You read research from adjacent areas including frontier AI, computer vision, language modelling, process mining, program synthesis, information retrieval, graph learning, and knowledge representation, and can judge whether it applies.

You can identify the core idea inside a method, understand the assumptions that make it work, determine whether those assumptions hold in your setting, and adapt the idea into a working prototype.

Research Judgment

You can distinguish a difficult engineering problem from a research problem, design experiments that isolate the important variable, and abandon an elegant approach when the evidence suggests a simpler method works better.

Applied Foundation Model Systems

Hands-on experience building reliable systems with modern language or multimodal models, including retrieval, structured generation, evaluation, grounding, and model adaptation. You understand the failure modes that arise when model outputs are used to recover structured process knowledge from noisy evidence.

This May Include Experience With Retrieval design, embedding strategies, chunking, re-ranking, and grounding against source evidence. LLM-as-Judge systems, rubric design, calibration against human labels, and knowing when automated evaluation is insufficient. Getting hands-on experience with LLMs, RL, RLHF, RLAIF, post-training (SFT), evals, graders, synthetic data, model training and production ML systems.

Faithfulness verification, constrained generation, confidence calibration, and methods for reducing unsupported model outputs. Guardrails, input and output validation, graceful degradation, and monitoring for drift in production.

Python Proficiency

Strong software engineering in Python, experience with ML frameworks such as PyTorch, comfort with data processing at scale, and the ability to write clean, tested, production-quality code.

Communication

You can explain an algorithmic tradeoff to both a researcher and a product manager, and you write clear technical and design documents. Preferred Experience A strong research track record, demonstrated through publications in top-tier machine learning, artificial intelligence, reinforcement learning, natural language processing, computer vision, or data mining venues.

Experience translating research ideas from published work into production systems with measurable impact. Process mining: event log analysis, process discovery algorithms, conformance checking, or variant analysis.

Information extraction from documents

NER, relation extraction, table extraction, or structured output generation from unstructured text. Multimodal data processing: extracting signals from screen recordings, UI screenshots, or interaction traces alongside textual data. Foundation-model adaptation, including supervised fine-tuning, LoRA, synthetic-data methods, distillation, reinforcement-learning-based post-training, and evaluation of resulting model behavior.

You Might Be a Great Fit If You Look for the hidden structure inside messy observations before deciding what to build. You are excited by the possibility of discovering how a process actually works without being handed its rules in advance. Enjoy problems where the right abstraction, model, and formulation are not obvious.

You are willing to move between algorithms, learned representations, graph structures, retrieval, and foundation models until the system can explain what it has observed.

Are drawn to the idea of building an end-to-end, process-agnostic machine: one that can observe unfamiliar work, infer its underlying boundaries, steps and decision logic, distinguish meaningful variation from noise, and produce concretized knowledge that gives an agent the full state representation to successfully navigate complex real-world processes.

Read deeply across fields and adapt ideas from process mining, program synthesis, representation learning, graph algorithms, information retrieval, multimodal learning, and agentic systems. You care about the underlying insight more than the category it came from.

Are comfortable moving between theory and reality: papers, hypotheses, experiments, raw traces, screen recordings, production failures, and systems that must work at organizational scale. Want to work with a small, unusually ambitious team where research judgment and engineering execution are inseparable, and where your ideas can shape the architecture of an entirely new kind of system.