Geddy Dukes
AI/ML systems engineer who builds production ML infrastructure from first principles. Implemented a 67M-parameter language model from scratch, multimodal self-supervised learning systems (JEPA/CPC), and neuro-symbolic AI frameworks. Published open-source agent runtime on npm. Deep experience in full-stack development, LLM orchestration, and production deployment. Domain expertise in financial systems, real estate, and regulated environments.
Featured Projects
Deep Learning Research
TinyLLM - Language Model from Scratch
live67M parameter transformer with continual learning system
Implemented a 67M parameter transformer from scratch, featuring a custom continual learning system to prevent catastrophic forgetting.
- 67M Parameters
- Continual Learning
- Transformer Architecture
World Model - Multimodal Self-Supervised Learning
liveJEPA + CPC + CLIP implementation with episodic memory
Vision–audio contrastive learning with episodic memory; nightly training/validation and rollout evaluation to test temporal dynamics and embedding coherence.
- JEPA + CPC + CLIP
- Episodic Memory
- Self-Supervised
AI Infrastructure & Frameworks
Daedelos SDK - Neuro-Symbolic AI Platform
deployingEnterprise AI governance with audit trails and hybrid reasoning
Hybrid multi-agent framework that composes deterministic symbolic rules with LLM agents for explainable, traceable decisions in regulated contexts.
- Neuro-Symbolic
- Audit Trails
- Hybrid Reasoning
Feather Agent - Open-Source Agent Runtime
liveProvider-agnostic agent framework (published on npm)
Open-source runtime for stateful agents, tool execution, model routing, and transparent streaming plans with config-driven state/behavior/lifecycle.
- Provider-Agnostic
- npm Package
- Agent Runtime
agentbench – Open-Source Agent Evaluation Framework
liveFramework for benchmarking LLM and tool-using agents
Framework for benchmarking LLM and tool-using agents. Supports async execution, retries, rate limits, composite judges, JSONL artifacts, traces, and HTML reports.
- Agent Benchmarking
- Open Source
- Python Package
- Report Generation
Full Stack Applications
Audio Routing System - Hybrid ML + Rules
liveIntelligent audio classification and dynamic stem routing
Ingests multi-track audio events (speech/SFX/music), classifies via rule + ML, and dynamically routes stems through a virtual post-production graph.
- Hybrid ML + Rules
- Dynamic Routing
- Audio Classification
Financial Analysis Platform
liveReal-time underwriting metrics with RBAC + audit
Underwriting workflow platform processing millions in loan volume; 15+ metrics with sub-second latency, RBAC, audit logs, and secure APIs.
- Sub-second calc engine
- Compliance-ready
- Used in production
Professional Experience
Credit Analyst & Program Manager, Catalytic Capital
Community Vision Capital & Consulting
- Built a financial spreading platform (TypeScript/Next.js/PostgreSQL/Supabase) with RBAC, audit logs, and real-time pipelines used by underwriting teams.
- Encoded complex financial rules as reliable backend logic across automated workflows.
- Managed $30M program and underwrote deals up to $3M; integrated lending logic into internal tooling.
Finance → software builder: moved complex rules into auditable, real workflows.
Catalytic Capital Program Analyst
Community Vision
- Improved program efficiency ~20% via data models and analysis.
- Coordinated budgeting/performance across ~$100M in programs; managed AHP/AHEAD compliance cycles.
Finance Project Manager
Dorjil Company
- Built Python/SQL tools securing $400K+ in subsidies; automated data pipelines for ~25% efficiency.
- Led multi-year audit recovering $2M in overdue operating subsidies.
Full-Stack Software Engineer
Blindly (contract)
- Implemented ID parsing/verification (Node/React/Redux) and shipped authentication flows in a cross-functional team.
Core Skills
Engineering
AI Systems
Delivery & Leadership
Get In Touch
Open to in person roles in the SF Bay Area and remote opportunities