Architecting the Future of Generative AI
Advanced consulting in Natural Language Processing, Computer Vision, and Large Language Models. Led by William Foland, PhD.
About
Engineering the optimistic future of AI. We partner with leaders in science and industry to build tools that expand human capability, turning cognitive labor into creative leverage for a better world.
We focus on high-impact sectors—Education, Healthcare, and Complex Engineering—delivering AI architectures that augment expertise and accelerate discovery, not distraction.
Capabilities
Agentic AI & Orchestration
Building autonomous systems that plan and execute. Custom MCP Servers and Claude Skills for complex workflows.
- LangGraph
- MCP Servers
- Claude Skills
- LangFuse
Interactive AI Experiences
Creating low-latency, immersive learning environments using real-time video avatars and responsive voice AI.
- Video Avatars
- Real-time Audio
- Low Latency
- HeyGen / D-ID
Generative AI & LLMs
Architecting advanced RAG accelerators, fine-tuning open source foundational models, and deploying production LLM services.
- RAG
- PEFT Fine-tuning
- Synthetic Data
- Model Eval
Computer Vision & Multimodal
Real-time video analytics and object detection. Integrating vision with language models for holistic understanding.
- YOLO
- OpenCV
- PyTorch
- Multimodal RAG
Full Stack & Cloud Architecture
End-to-end system design. From modern responsive frontends to scalable backends and cloud infrastructure.
- React / Vite
- FastAPI / Node
- AWS / Azure
- Docker / K8s
Domain Expertise
Deep experience in regulatory-heavy industries (Healthcare, Finance) and complex hardware systems.
- Clinical Support
- FinTech Analysis
- Semiconductors
Track Record
Gen AI Data Scientist
Harvard Business School, Cambridge, MA
- Solo-architected and built a real-time AI meeting simulator where a single LLM controlled four distinct video avatars simultaneously, coordinating speech, timing, and behavior to emulate a live multi-participant discussion.
- Engineered a versatile simulation architecture configurable for hundreds of scenarios, enabling rapid creation of new teaching, coaching, and role-play environments without modifying core logic.
- Designed a custom WebRTC + NextJS + FastAPI interface translating model outputs into deterministic multi-avatar actions under tight latency constraints.
- Integrated institutional systems into GPT-5.1 and Claude agents via custom MCP servers, enabling secure, real-time data access and controlled agent actions.
- Added automated evaluation using real-time synthetic dialog generation, leveraging Claude-agent dialogues to stress-test GPT-5.1 on multi-speaker coordination, reasoning quality, and timing stability.
- Delivered the platform to 900+ MBA students and demo’d it for deans, donors, and faculty as a flagship component of the school’s AI initiative.
- Owned the full lifecycle from prototype to production, including concurrency, evals, guardrails, and streaming reliability for multi-avatar operation.
Gen AI Consultant
IsoGrab.com, Houston, TX
- Architected and built a vision-LLM pipeline that ingests batches of complex isometric piping PDFs—oil & gas, water treatment, industrial facilities—and extracts high-accuracy structured data across widely varying drawing styles.
- Designed a multi-model system combining OCR, layout parsing, and LLM reasoning to recover bill of materials, elevations, insulation details, line identifiers, and component metadata with minimal manual correction.
- Implemented automated consistency checks and iterative refinement loops that drove error rates down to levels acceptable for safety-critical engineering workflows.
- Delivered a working 4-week proof of concept that reduced diagram-to-data conversion from weeks to hours, leading the client to approve a full production build.
- Engineered the pipeline for robustness across large document sets, non-standard symbol libraries, and heterogeneous drawing conventions.
Chief GenAI Data Scientist
BigRio LLC, Cambridge, MA
- Architected and built an LLM-driven engine embedded directly into clinical workflows (no UI), where GPT-4 analyzed patient data and generated draft care plans, reducing practitioner documentation time and improving consistency.
- Designed the ingestion and reasoning pipeline to handle heterogeneous clinical data—structured fields, free-text notes, historical records—while enforcing domain constraints and alignment with clinical guidelines.
- Implemented guardrails, validation rules, and safety filters to keep outputs reliable in a regulated healthcare setting.
- Delivered a backend service that became part of daily practitioner workflow, accelerating documentation and reducing administrative burden.
- Built an Azure-based RAG PoC for clinical decision support, delivered from concept to pilotable prototype in 8 weeks.
Founder and Chief Consultant
Itinitek Ltd, Golden, CO
- Enhanced open source foundational models for legal domain using PEFT fine-tuning and RAG.
- Engineered scalable AWS pipelines utilizing LLM models and SQL for financial options analysis.
Chief Scientist
Lilac Cloud Inc, Cupertino, CA
- Architected and designed vision-AI pipelines for edge computing, combining Dockerized GPU services with CUDA-accelerated model execution to support real-time video analytics in bandwidth- and latency-constrained environments.
- Built custom FFmpeg filters and Libav extensions to run inference inside the video pipeline, enabling object recognition, event detection, and metadata insertion without interrupting frame flow.
- Implemented GAN-based imperceptible video watermarking and other security features to protect high-value media across distributed streaming environments.
- Delivered low-latency frame processing systems for live sporting events and other real-time workloads, maintaining consistent throughput on varied edge hardware.
- Optimized GPU utilization, memory transfers, and batching strategies to achieve stable real-time performance.
Cofounder and Chief Scientist
Bolt Analytics Corp, Santa Clara, CA
- Architected time series anomaly detection using CNN, RNN, transformer, and gradient boost algorithms.
- Led team in developing automated diabetes detection from retina scans using TensorFlow.
Research Scientist
CU Computational Language and Education Research, Boulder, CO
- Developed recurrent NLP models for automated speech recognition and dialog analysis in K-12 STEM education.
- Built system to help teachers reflect on and improve instructional practices through AI-powered feedback.
Expert Technical Consultant
Dovel & Luner, LLP, Santa Monica, CA
- Served as an expert witness in semiconductor patent litigation.
- Contributed technical analysis leading to successful client outcomes.
Founder & Independent Developer
Itinitek Ltd, Golden, CO
- Architected, developed, and marketed six GPS, graphics, and skiing applications for iPhone (Objective-C).
- Worked concurrently with MS and PhD coursework at University of Colorado.
Senior Director, Optical Products
Marvell Semiconductor, Santa Clara, CA
- Led a 150-engineer division developing full System-on-Chip (SoC) solutions (ARM core, DSP, Read Channel).
- Managed complete product lifecycle from architecture through silicon to mass production.
- Drove multiple generations of mixed-signal optical storage controllers.
Research & Education
Education
PhD in Computer Science
University of Colorado, Boulder, CO (2017)
Dissertation: 'Natural Language Understanding: Deep Learning for Abstract Meaning Representation'
Emphasized state-of-the-art recurrent neural networks, utilizing hardware background and the latest research approaches for AI.
MS Computer Science
University of Colorado
BS Electrical Engineering
University of Colorado
Selected Publications
- Abstract Meaning Representation Parsing using LSTM Recurrent Neural Networks, ACL Conference, 2017, Vancouver
- CU-NLP at SemEval-2016 Task 8: AMR Parsing using LSTM-based Recurrent Neural Networks, NACL Conference, 2016, San Diego
- Dependency-Based Semantic Role Labeling using Convolutional Neural Networks, NACL Conference, 2015, Denver
Foundational Engineering
Deep roots in semiconductor design and complex system architecture provide a unique advantage in optimizing modern AI workloads.
Before pioneering GenAI solutions, I spent years architecting mixed-signal integrated circuits and read channels for data storage. This foundational experience in low-level signal processing, error correction, and hardware constraints directly informs my approach to building efficient, scalable, and robust AI systems today. Understanding the "metal" makes the software faster.
U.S. Patents
...and 12 other patents in signal processing, timing recovery, and sequence detection (19 Total).
Ready to Accelerate Your AI Strategy?
Available for consulting and development projects.