Frontier AI for decisions where evidence, accuracy and trust matter.
AIDF works with financial institutions and technology partners to turn difficult industry problems into research-grade AI systems. Our engagements span autonomous financial research, multimodal credit intelligence, investment decision support, causal analytics, graph-based forecasting, adaptive RAG and interpretable machine learning — combining methodological innovation with the domain constraints of real financial workflows.
Every engagement begins with a hard question, not a model.
Before any architecture gets built, a partner brings us a question they can't yet answer with the tools they have. Below are nine of those questions — and where we took them.
We've been finding the answers — one industry engagement at a time.
See how we solved them →Not one model. A portfolio of research capabilities.
Industry problems rarely fit a single algorithm. Across the nine projects on this page we combine specialised capabilities into end-to-end research architectures — from document understanding and evidence retrieval to causal structure, forecasting, verification and human decision support. Select a node to see how it maps to our engagements.
AI ResearchCoreAgentic
Agentic AI for financial research workflows
Specialised agents coordinate retrieval, reasoning, extraction, verification and report generation — an AI research workflow that can execute complex knowledge tasks while preserving evidence and expert oversight.
- Flagship: Development Bank of Japan — automated public-company research and credit-memo generation
- Workflow: discover → understand → verify → reason → generate
- Research edge: multi-stage AI with explicit verification and human approval
University depth. Industry tempo. Research that can survive scrutiny.
Corporate–university collaboration is most valuable when it creates something an internal delivery team would not build alone. AIDF combines frontier methods with financial-domain framing, independent evaluation and translation into workflows where accuracy, governance and human judgement matter.
Research, not repackaging
We start from the unresolved decision problem and design a technical research question around it — rather than forcing the problem into a pre-selected product or model.
Finance-native intelligence
Models are designed around financial evidence, document structure, risk logic, market signals and the operational realities of regulated decision-making.
Verification by design
Where the use case is high-stakes, we make validation, interpretability, source reconciliation and human review part of the architecture — not an afterthought.
Translation with a research edge
The goal is a prototype that teaches both sides something new: a differentiated capability for the partner and a reusable research insight for the wider financial AI ecosystem.
Autonomous credit research, from source documents to a verified credit memo.
A six-stage agentic workflow built for the Development Bank of Japan: AI collects public and official documents, reasons over visually complex filings, extracts material financial numbers, verifies them, synthesises credit-relevant evidence and drafts a memo for human review.
Collect Evidence
Annual reports, regulatory filings, official corporate materials and public evidence.
Visual Reasoning
Read tables, charts, statements and complex financial-document layouts.
Cross-Check
Reconcile sources and validate important numbers before downstream use.
Credit Analysis
Identify material events, risk signals and evidence relevant to credit judgement.
Draft Memo
Convert structured evidence into a professional credit memorandum.
Human Approval
Keep analysts in control of interpretation, review and final decision.
Industry research portfolio
Each project begins with a concrete partner problem and is pushed toward a technically distinctive research architecture. The emphasis is on capabilities that go beyond standard enterprise AI: domain adaptation, agentic reasoning, multimodal inference, causal structure, model distillation, synthetic data and evidence-aware automation.
Autonomous Credit Research & Verified Credit-Memo Intelligence
An agentic AI pipeline that autonomously discovers and collects public and official corporate documents, reasons over visually complex filings and extracts financially material numbers. A separate AI verification layer checks those numbers before they enter analysis, and the system drafts a full credit memo — while a human-in-the-loop control point keeps final judgement, review and approval with the analyst.
LLM-Native Investment Idea Ranking
State-of-the-art LLM methods rank trading ideas using the full decision context available to an investment team — fundamental information, technical signals and asset-pricing factors — rather than evaluating each signal in isolation. The architecture synthesises heterogeneous evidence and reasons across potentially conflicting indicators to produce a disciplined relative ranking of investment opportunities.
Causal Geo-Temporal Graph Forecasting for Rent & Price
A simultaneous rent-and-price forecasting framework combining graph neural networks, causal geospatial-temporal modeling and multi-horizon prediction. Locations, neighbouring markets and time-varying interactions are modelled as a connected system rather than independent observations, with synthetic data enriching experimentation where observed market histories are sparse.
Adaptive Self-Improving RAG Tutor for Finance Exams
A retrieval-augmented generation tutor designed for financial professionals preparing for high-stakes examinations. Rather than a static chatbot, the research centres on an adaptive, self-improving learning loop that retrieves relevant domain material, responds to the learner's current knowledge state and continually refines how it supports practice and explanation.
Multimodal Private-SME Fraud Detection & Credit Intelligence
A multimodal AI framework for private-SME risk assessment that unifies fraud detection and credit scoring in one analytical stack. The system is designed for environments where the most informative risk evidence is fragmented across heterogeneous private-company data, learning across modalities to surface inconsistencies and produce richer credit intelligence for SMEs underserved by traditional data pipelines.
Transparent Credit Scorecards through LightGBM Distillation
A transparent, interpretable credit scorecard distilled from a high-performing LightGBM model, addressing a core challenge in digital lending: retaining the predictive power of modern machine learning while making the resulting credit logic easier to inspect, explain and govern within a regulated decision process.
Insurance-Specialised LLM & Causal Sales Intelligence
A two-stream insurance AI programme: one stream fine-tunes an LLM on insurance-specific language and terminology, creating a more domain-aware foundation for downstream tasks; a second builds an AI-enabled sales KPI dashboard in which the LLM interprets sales trends and works alongside causal analysis to identify the factors most plausibly associated with increases or declines.
LLM-Guided Feature Selection for SME Credit Rating
This collaboration applies LLMs to the feature-selection problem in SME credit rating, using the model's ability to reason over financial concepts to identify which candidate variables are most decision-relevant for risk assessment — bringing semantic, domain-aware reasoning into a modeling stage often treated as purely mechanical.
LLM-Powered Automated Financial Report Writing
This project explores the use of LLMs to automate financial report writing, turning the model from a general text generator into a domain-oriented reporting engine — with the longer-term goal of reducing repetitive drafting effort while making report generation faster, more standardised and more responsive to new information.
Bring us the problem your current AI stack cannot solve.
The best collaborations begin before the technical solution is obvious. If the challenge involves fragmented evidence, high-stakes reasoning, verification, causal questions, multimodal data or expert workflows, we can help turn it into a rigorous research programme and a differentiated prototype.
Discuss a Corporate–University Research Project →Decision Problem
Define what the organisation needs to decide, what evidence exists and where current tools fail.
Research Advantage
Identify the new method, benchmark or architecture that could create a defensible capability.
Prototype & Evaluate
Test the approach on realistic data and workflows with explicit performance and trust criteria.
Operational Learning
Convert the research into a usable prototype, decision framework and next-stage R&D agenda.