William Foland, PhD
I turn complex human workflows into deterministic AI engines — from architecture through production delivery. PhD in neural language understanding, 19 US patents, 20+ years shipping.
AI systems architect with a PhD in neural language understanding and 20+ years delivering production software across software, silicon, and applied machine learning.
Itinitek builds real-time multimodal platforms, vision analytics, and domain-specific LLM systems that streamline complex human workflows. From Harvard's AI teaching simulations to healthcare and industrial automation, we clarify the problem, design the architecture, and close the execution gap quickly — with deterministic checks, not brittle automation.
Production systems and the architecture behind them. Each card is a look at the real work — vision pipelines, clinical extractors, and real-time orchestration engineered to run reliably under load.
A vision-LLM pipeline that ingests a stack of flat 2D isometric drawings and reconstructs them into a single validated 3D system model. Perception (vision/OCR) is separated from reasoning (LLM) so each stage tunes independently, with geometric consistency checks driving error rates to safety-critical levels — then you chat with your system in plain English: "find all butterfly valves in the 2-inch line." Weeks → hours.
An LLM engine embedded directly into clinical workflows — with an efficient UI for occasional
human-guided correction — that reads freeform patient notes and grounds every finding to standard
ontologies (ICD-10, SNOMED, HPO, LOINC, RxNorm). Exact negation
handling ("denies history of" → [NEGATED]) keeps it reliable in a regulated
setting. High-quality output supports sophisticated downstream workflows.
A real-time meeting simulator where a single LLM controlled four distinct video avatars at once — coordinating speech, timing, and behavior under tight latency over WebRTC + Next.js + FastAPI. Speech-driven scenarios allow active student participation in the conversation. Delivered to 900+ MBA students; a closed-loop eval system stress-tested coordination with synthetic dialogue.
A BAA-secured LLM agent sits outside the VPC; an MCP broker inside the private subnet routes each call to controlled MCP servers — augmenting the agent's capabilities within secure execution boundaries instead of open-ended access. Every MCP server runs secured inside the VPC and exposes a typed schema that whitelists exactly which operations it permits — so the agent gets approved, validated actions only, never direct access to the data or systems behind them.
Three ways to work together as a fractional GenAI scientist & architect — from strategic oversight to dropping a proven engine into your cloud.
Retainer-based
Architecture roadmapping, multi-model cost/latency routing analysis, vendor vetting, evaluation strategy, and security governance for agent action spaces. Async-first strategic oversight plus scheduled executive syncs.
Project-based
Direct ownership of engineering teams — designing planner-executor-verifier self-correcting loops and deterministic code-as-judge constraint checks. Active, project-bound engineering leadership from prototype to production.
Accelerator integration
Licensing, optimization, and deployment of a proven engine — the Spatial-Semantic or Clinical Extractor core — directly into your cloud environment. A focused 90-day execution sprint that bypasses standard R&D timelines.
Multi-agent orchestration for production workflows, including planner-executor-verifier loops, MCP tool use, and stateful versus stateless reasoning strategies.
Closed-loop evaluation systems, failure-mode analysis, and iterative tuning for AI products that need to work reliably under real constraints.
RAG pipelines, fine-tuning, prompt engineering, and model routing for domain-specific systems that improve expert workflows instead of adding brittle automation.
Low-latency voice, video, and avatar systems built for human interaction, teaching simulations, and other applications where timing and responsiveness matter.
Vision pipelines for industrial and edge workloads, combining detection, OCR, parsing, and multimodal reasoning to recover usable structure from complex visual inputs.
End-to-end system design spanning model serving, cloud infrastructure, and application delivery across prototypes, pilots, and production systems.
University of Colorado, Boulder, CO (December 2017)
Dissertation: Natural Language Understanding: Deep Learning for Abstract Meaning Representation
Deep learning models similar to those that power ChatGPT and modern LLMs.
University of Colorado, Boulder
University of Colorado, Boulder
Before pioneering GenAI solutions, I spent years architecting mixed-signal integrated circuits and read channels for data storage. That foundation in low-level signal processing, error correction, and hardware constraints directly informs how I build efficient, scalable, robust AI systems today.
19 U.S. patents spanning signal processing, read channel design, mixed-signal integrated circuits, and optical recording systems. Full patent list available on request.
Available for fractional GenAI leadership, architecture, and engine deployment.