Founding Applied Scientist

About this role

Sphere's focus is novel reasoning over proprietary research. We help scientists and engineers compress months of literature review into seconds, pulling from disparate academic fields that may be only loosely related. At the core of this is understanding what data is valuable: which data we retrieve, what data we make available to retrieve, and, over time, a value score that can be used to price data itself. You will own the systems that make that reasoning accurate, grounded, and trustworthy, and help shape how we measure and value the research flowing through Sphere.

What you'll own

• Reasoning quality over proprietary research, from retrieval through grounded, cited synthesis
• Post-training methods that specialize models for research reasoning: fine-tuning, LoRA and other PEFT, preference optimization (DPO / RLHF), and distillation
• Retrieval, ranking, and reranking that surface relevant work across loosely related fields
• Chunking, parsing, and representation quality for complex scientific and technical documents
• Benchmarking and evaluation frameworks that reflect how researchers actually work
• Data valuation: which data we retrieve, what we make available to retrieve, and a value score that can price data
• Citation fidelity, groundedness, and hallucination reduction

What you'll do

• Apply and combine post-training methods (fine-tuning, LoRA / PEFT, RAG, DPO / RLHF, distillation) to make reasoning over research sharper and better grounded
• Improve retrieval and cross-domain synthesis so a query surfaces the right work even from loosely related fields
• Design and run evaluation frameworks grounded in real research workflows, and define what "better" means in measurable terms
• Build the data-valuation layer: score which sources are worth retrieving and worth acquiring, growing toward a price signal for data
• Test and refine chunking, parsing, embedding, reranking, and selection strategies over scientific corpora
• Ship production improvements alongside product and infrastructure, not just experiments

What we're looking for

• Strong applied data science or applied ML background across information retrieval, ranking, NLP, or reasoning systems
• Hands-on with post-training methods: fine-tuning, LoRA and other PEFT, RAG, preference optimization (DPO / RLHF), or distillation
• Ability to design evaluation frameworks, not just run existing benchmarks
• Comfort turning fuzzy questions like "what data is valuable?" into measurable models and metrics
• Strong engineering instincts and willingness to ship production code
• Comfort with messy scientific and technical documents and real-world constraints

Strong pluses

• Scientific, technical, or multimodal research corpora
• Retrieval-augmented reasoning, agentic retrieval, or grounded-answer systems
• Citation-grounded evaluation, provenance, or source attribution
• Data valuation, pricing models, or value-of-information and active-learning approaches
• Optimizing training or inference under constrained hardware

Your first 90 days

• Establish a clear evaluation baseline for research reasoning, retrieval, and citation quality
• Ship at least one post-training or retrieval improvement that moves a measurable quality metric
• Stand up the first version of a data-valuation signal: which sources are worth retrieving and acquiring
• Lay out a practical roadmap toward a value score that can inform how we price data

Compensation

Benefits

• Health, dental, and vision insurance
• Flexible paid time off
• Company-provided equipment

As an early-stage company, we're committed to building a benefits package that grows with us.