Gunnar Hoffman
Resume (MD)Resume (PDF)

AI & architecture case study

Agentic statement spreading for commercial lending

Baker Hill, commercial lending customers

Business problem

Lenders receive large document packages: financial statements, tax returns, and supporting files. Analysts historically re-keyed figures into spreading tools by hand.

Statement periods had to balance; when they did not, teams traced line items to the right chart-of-accounts entries through slow, manual research. The work consumed skilled analyst time and invited transcription and mapping errors.

Role

Major individual contributor to the agentic architecture: how the system reasons like an experienced analyst, reading narrative context, matching and verifying figures, explaining decisions, and producing an audit trail that shows its work. Also defined engineering practices for non-deterministic AI in this domain and built test tooling and frameworks for shift-left detection and regression prevention.

Stack & approach

C# on Microsoft Azure AI Foundry. The agent framework is model-agnostic; production tuning uses GPT-5.6-Sol.

The pipeline loads a structured spread definition as the target shape, combines it with source documents, and asks the agent to emit JSON that conforms to a defined schema. A deterministic math-verification step validates balances; failures return as feedback so the agent can correct itself. A final engine loads verified data into the database and confirms the load succeeded.

Pipeline

Statement spreading agent pipelineSource documents and a spread definition feed an agent on Azure AI Foundry. The agent produces schema-conformant JSON. A deterministic math verifier accepts the result or sends feedback for correction. Verified data loads into the database; analysts perform final verification.Source documentsFinancials, tax filesSpread definitionTarget structure fileAgentAzure AI FoundryModel-agnostic (GPT-5.6-Sol)Structured JSONSchema-conformantMath verificationDeterministicLoad engineDB write + verifyAnalyst reviewPassFeedbackIssues → agent corrects
Agent output is bounded by schema and math checks; failed verification returns structured feedback until balances pass, then data loads into the spreading database for analyst confirmation.

Tradeoffs

LLM steps are non-deterministic, so the design pairs the agent with schema-bound output, deterministic math verification, and a correction loop instead of trusting a single pass. Analysts still review results, but the heavy lifting moves from manual entry to verification. Engineering standards and test frameworks account for regression and drift in model behavior over time.

Outcome

Document packages that once drove hours of manual spreading and reconciliation can load in minutes with balanced statements and math already verified. The analyst role shifts to brief verification: confirming results, refining a few line names or placements, instead of building the spread from scratch.