Geddy Dukes
ML and Sales Engineer at Danti, building geospatial intelligence products on grounded LLM workflows. I specialize in systems where hallucinations aren't acceptable and every decision needs an audit trail—from training a 67M-parameter model from scratch to shipping production underwriting platforms for $30M portfolios.
What I Bring to Your Team
Most ML engineers can train models. Few understand how to deploy them where errors cost money or violate regulations.
Ship production systems fast
Built a complete financial analysis platform and AI underwriting system while simultaneously managing a $30M loan portfolio. Comfortable across the full stack—PyTorch training loops to React frontends to PostgreSQL schemas. I prototype rapidly and ship incrementally.
Domain expertise in finance
Spent years as a credit analyst underwriting deals up to $3M and managing compliance workflows across ~$100M in programs. I understand lending requirements, risk assessment, and regulatory constraints—not just the ML. This means I can build AI systems that actually work in regulated environments without months of onboarding.
Safety-first, neuro-symbolic architecture
Every production system I've built separates probabilistic extraction from deterministic decisions using neuro-symbolic design patterns. Complete audit trails, provenance tracking, and confidence thresholds at every layer. I know when to use LLMs and when to use rules—and I build the infrastructure to enforce that boundary.
Featured Projects
TinyLLM – Language Model from Scratch
67M parameter transformer trained entirely on consumer hardware — proving deep understanding of architecture, not just API calls.
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.
Key capabilities:
- Full transformer implementation (RoPE, RMSNorm, SwiGLU)
- Continual learning with replay buffers
- Evaluation gating to prevent catastrophic forgetting
- Trained on consumer GPU
- 93.94% exact-match accuracy on CLI commands
Provenance – AI-Native Credit Decision Platform
AI underwriting platform designed to reduce loan analysis from weeks to days while maintaining full regulatory compliance for CDFIs.
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.
Key capabilities:
- Agent-orchestrated workflow with deterministic + LLM hybrid analysis
- Three-tier OCR pipeline for financial documents
- Automated risk surfacing with full source provenance
- End-to-end auditability for regulatory compliance
- Applicant portals + analyst dashboards
AgentFT – Agent Evaluation Framework
Testing infrastructure that makes AI agent reliability measurable, repeatable, and debuggable. Everyone is building agents — almost no one knows how to test them.
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.
Key capabilities:
- Async evaluation engine for agent tasks
- Regression detection across versions
- Trace-level observability
- Composite judge patterns
- Published on npm and PyPI
Professional Experience
Sales Engineer
Danti
- Sales Engineer supporting both ML and general engineering
ML Engineer Intern → converted to full-time
Danti
- Conceived and built Danti's Monitoring and Daily Brief products, automating recurring geospatial intelligence workflows from data collection and geospatial analysis through synthesis and delivery
- Identified a commercial real estate site-intelligence opportunity and built it from concept through MVP
- Built AI-driven investigation workflows spanning grounded LLM analysis, automated report generation, scheduled orchestration, synthesized insights, and persistent geospatial work products
- Designed personalization and data architecture for context-aware intelligence, resolving preferences across user, role, and organization levels while integrating property, demographic, hazard, imagery, and POI data
- Identified and resolved scaling, schema, and reliability risks across the stack, hardening geospatial processing, access controls, provider fallbacks, map behavior, and production AI workflows
Intern to full-time in seven months, after conceiving and shipping products from concept to MVP.
Credit Analyst & Catalytic Capital Program Manager
Community Vision Capital & Consulting
- Built production financial spreading platform (TypeScript/Next.js/PostgreSQL/Supabase) with RBAC, audit logs, and real-time calculation pipelines — used by underwriting teams to analyze loan applications across millions in volume
- Managed $30M catalytic capital program and personally underwrote loan deals up to $3M, gaining deep hands-on experience in lending compliance, risk assessment, and regulatory requirements
- Encoded complex financial rules as deterministic backend logic, creating reliable automated workflows that became the foundation for later AI underwriting systems
- Integrated lending logic directly into internal tooling, reducing manual analysis overhead and standardizing underwriting processes
Finance → software builder: moved complex domain rules into auditable, production workflows.
Catalytic Capital Program Analyst
Community Vision
- Improved program operational efficiency ~20% through data modeling and financial analysis optimization
- Coordinated budgeting and performance tracking across ~$100M in programs
- Managed AHP and AHEAD compliance cycles, building deep expertise in regulatory requirements and audit processes
Finance Project Manager
Dorjil Company
- Built Python/SQL automation tools that secured $400K+ in operating subsidies for affordable housing portfolios
- Led multi-year audit that recovered $2M in overdue operating subsidies
- Automated financial data pipelines, achieving ~25% efficiency improvements in reporting and compliance workflows
Full-Stack Software Engineer (Contract)
Blindly
- Implemented ID parsing and verification system (Node.js/React/Redux) for identity authentication workflows
- Shipped production authentication features in fast-paced, cross-functional startup environment
Core Skills
Engineering
AI Systems
Delivery & Leadership
Let's work together
If your current AI strategy is "just use a bigger prompt," let's talk about how to build a system that actually survives an audit. I'm currently building at Danti and always glad to talk about AI systems that ship to production and handle real consequences, especially in regulated industries: fintech, healthcare, insurance, legal.
Based in SF Bay Area · Open to in-person and remote
geddydukes@gmail.com · 707-799-1271