Geddy Dukes

Built a GPT-style language model from scratch including RoPE positional encodings, RMSNorm, and SwiGLU activations. Designed a custom continual learning system with replay buffers and evaluation gating to prevent catastrophic forgetting. Achieved 93.94% exact-match accuracy on CLI command generation. This project demonstrates first-principles understanding of transformer architecture and training dynamics — the kind of depth that informs how I design and debug production LLM systems.

End-to-end underwriting workspace that automates financial document processing, surfaces risks with complete provenance tracking, and generates audit-ready credit memos. Built for CDFIs where every lending decision requires documentation and regulatory compliance. The challenge: using LLMs for analysis while maintaining the auditability required in financial services. The solution: strict separation between probabilistic document extraction and deterministic decision logic, with every recommendation tied directly to source documents and calculations.

Policy and claims decision system designed for environments where "the model said so" isn't good enough. Uses LLMs to interpret unstructured documents, then enforces strict validation, confidence thresholds, and deterministic rule evaluation before any decision is made. Every output includes complete audit trails showing which documents informed which conclusions. A working proof-of-concept for the neuro-symbolic architecture that now underpins the underwriting platform — demonstrating the pattern at a smaller scale before productizing it.

Evaluation harness that treats AI agents as testable software artifacts. Supports task abstractions, async execution, composite judging, and trace-level logging to detect regressions across versions and workflows. Critical for production deployments where you need to know if your agent got worse after an update. Published as open-source packages on both npm and PyPI — because good developer experience matters.

Production fintech platform built while managing a $30M CDFI portfolio. Supports real-time financial analysis, metric computation, and compliance workflows with sub-second latency. Architected deterministic calculation pipelines with role-based access control and full audit logging. This became the deterministic financial core later integrated into the AI underwriting platform — proving the value of building reliable calculation engines before layering AI on top.

Research discovery platform that converts unstructured academic papers into a structured, queryable knowledge graph while enforcing provenance and confidence at every step. LLMs propose extractions, which are then validated deterministically before graph ingestion. Built to explore how to combine LLM flexibility with symbolic validation — every claim in the knowledge graph can be traced back to its source document with a confidence score.