According to Business Insider, Barry Duong, former portfolio manager at Balyasny Asset Management and analyst at Citadel, has transitioned to lead AI strategist for public equities at professional services AI startup Hebbia. His team has processed over one billion pages of information for clients, equivalent to approximately 3,000 years of reading and 2,000 years of analysis. Duong explains that after leaving Balyasny under a noncompete agreement, he began testing early-stage AI tools and became convinced the technology’s potential far exceeded expectations. At Hebbia, his team develops custom prompts and workflows for financial services clients, conducting training sessions for hundreds of employees while focusing on enhancing capabilities across all seniority levels rather than just automating menial tasks. This career shift highlights the growing convergence between traditional finance expertise and artificial intelligence capabilities.
Sponsored content — provided for informational and promotional purposes.
The Technical Architecture Behind Billion-Page Processing
The scale of processing Duong describes—billions of pages across thousands of client workflows—requires sophisticated distributed systems architecture that goes far beyond simple API calls to large language models. Systems handling this volume typically employ multi-modal retrieval-augmented generation (RAG) pipelines that combine document parsing, vector embeddings, and semantic search before even reaching the generation phase. The computational requirements for processing financial documents—which often contain complex tables, financial statements, and regulatory filings—demand specialized preprocessing to maintain data integrity. Unlike consumer-facing AI tools, enterprise financial AI must maintain strict data lineage and audit trails, requiring sophisticated version control and model governance frameworks that can track every decision point in the analysis chain.
Domain Expertise as the New Critical Infrastructure
What makes Duong’s approach particularly insightful is his emphasis on hiring from “tier-one financial services firms” rather than traditional Silicon Valley technical talent. This represents a fundamental shift in how we think about AI implementation in regulated industries. The domain-specific knowledge these professionals bring—understanding credit agreements, merger documents, and complex financial instruments—becomes the essential scaffolding for effective AI systems. When building prompts for financial analysis, the difference between a generic instruction and one crafted by someone who understands covenant analysis or deal structuring can mean the difference between useful insights and financial nonsense. This human expertise serves as the crucial validation layer in systems processing thousands of documents, ensuring that AI outputs align with financial reality rather than statistical probability.
The Evolution of Financial Workflows
Duong’s description of “automating a PowerPoint deck or a financial model” undersells the profound workflow transformation happening in finance. We’re moving from sequential manual processes to parallel AI-assisted workflows where junior analysts become orchestrators of multiple AI agents simultaneously. A single analyst can now oversee due diligence, financial modeling, and presentation preparation in parallel rather than sequence. This requires new skill sets around prompt chaining, output validation, and agent management. The most successful implementations I’ve observed create feedback loops where human corrections improve model performance over time, essentially creating customized AI assistants that learn the specific analytical style and preferences of each financial institution.
The Future Frontier: Beyond Language Models
Duong’s mention of “foundational quant models that could do math a bit better than maybe the LLMs can” points to the next wave of financial AI. Current large language models excel at textual analysis but struggle with the precise mathematical reasoning required for complex financial modeling. The next generation will likely combine symbolic AI systems for mathematical operations with neural networks for pattern recognition, creating hybrid systems that can both reason about financial concepts and perform precise calculations. We’re already seeing early research into financial-specific foundation models that understand accounting principles, regulatory frameworks, and market dynamics at a fundamental level rather than as statistical patterns in training data.
Implementation Challenges and Organizational Risks
The transition Duong describes isn’t without significant technical and organizational challenges. Financial institutions face model drift risks where AI systems that perform well during testing may degrade as market conditions change. There are also substantial data governance concerns—ensuring that sensitive financial information remains protected while still enabling the AI systems to learn and improve. Perhaps most challenging is the cultural transformation required: moving from traditional hierarchical financial organizations to flatter, more collaborative structures where junior staff manage AI systems that perform work previously done by mid-level professionals. Success requires not just technical implementation but comprehensive change management that addresses both workflow redesign and career path evolution.
