ElmWater AIPioneering Financial AI Research
Our R&D is not an academic pursuit. It's a focused engineering discipline aimed at solving one problem: building the most reliable, capable, and secure AI for the financial world. We translate the latest breakthroughs into measurable business advantages for our clients.
Our Research Principles
Finance-First, Not Model-First
We never chase the largest models, but rather seek and create the architectures and technologies best suited to solve specific financial problems.
Rigor over Novelty
Robustness and reproducibility take precedence over pure theoretical novelty.
Translational Research
Our ultimate goal is not to publish papers, but to create value in our clients' business processes.
Core Research Pillars
Our research is organized around three fundamental pillars that drive our engineering discipline.
Pillar 1: Specialization & Adaptation
Research Goal: To efficiently transform general-purpose large models into "domain experts" proficient in finance.
Core Technical Paths
Instruction Tuning with Financial Corpora
Fine-tuning using massive amounts of high-quality financial instruction data (such as financial statement analysis and risk report generation) to enable the model to master financial thought processes.
Reinforcement Learning from Human Feedback (RLHF)
Employing reinforcement learning based on human feedback, using financial compliance regulations and ethical guidelines as optimization targets to ensure that the model's output is safe, compliant, and in line with human values.
Constitutional AI
Embedding financial regulatory regulations and ethical guidelines as a "constitution" into the model's reasoning process to achieve hard constraints on model behavior.
Parameter-Efficient Fine-Tuning (PEFT)
Applying LoRA and its variants to achieve efficient and lightweight model customization for clients.
Pillar 2: Reasoning & Scalability
Research Goal: To enhance the model's ability to handle complex financial logic and extremely long documents.
Core Technical Paths
Long-Context Understanding
Research and integrate Sparse Attention mechanisms (such as FlashAttention-2) to enable models to process hundreds of pages of financial statements and legal contracts without loss of quality.
Agent-based Reasoning
Explore multi-agent frameworks to decompose complex tasks into collaborative processes performed by dedicated agents, improving the reliability of complex reasoning.
Chain-of-Thought (CoT) and Program-Aided Reasoning
Guide models to generate intermediate reasoning steps or executable code, improving the accuracy of complex numerical computation and logical reasoning.
Pillar 3: Validation & Assurance
Research Goal: Establish a "quality control center" for financial AI to ensure the model's safety and reliability in production environments.
Core Technical Paths
Financial Benchmarking
Build and maintain internal financial capability benchmarks, covering dimensions such as numerical computation, logical reasoning, and factual accuracy.
Red Teaming & Safety Evaluation
Systematically conduct adversarial testing, construct attack sets for financial scenarios, and proactively discover and fix model defects.
Uncertainty Quantification
Developing techniques to quantify uncertainty, enabling models to estimate the confidence level of their outputs.
The People Behind the Financial AI
Garron Karamitas
Director, AI Model Customization & Product
Bridges AI capabilities with financial outcomes. Helps leaders define AI needs and builds solutions aligned with workflows and risk constraints.
James Grodsky
Founding Engineer, Platform & MLOps
Builds the production backbone for training, evaluation, and deployment. Focused on reliability, observability, cost efficiency, and the tooling teams need to ship models safely at speed.
Christy Tanenblum
Head of AI Safety & Trust
Builds the safety and reliability layer for everything we ship. Designs guardrails, evaluation, and monitoring so models behave predictably in real workflows and under real-world edge cases.
Research Collaborations & Dialogues
We are committed to advancing the field of financial AI through open collaboration and knowledge exchange. Our research is shaped by ongoing engagement with the global scientific community, including technical dialogues with leading labs such as:
These engagements focus on scoping research challenges in graph neural networks, reinforcement learning, and federated learning for financial applications.
Collaborate With Us
Interested in collaborating on research? We partner with leading academic institutions and financial firms to advance the state of financial AI.
Get in Touch