I turn complex expert workflows into reliable AI products.

Chris Van Buren — CVBAMBIGUOUS → AMBITIOUS

Proof in production

Owned the outcome

Founded and scaled Grantease to $100K ARR, then led AI engineering and product delivery at Aligrade, translating expert research workflows into production systems.

Embedded with experts

Biostatisticians, life-sciences research teams, NIH-funded researchers, and scientific grant writers.

Translated work into product

Turned those workflows into retrieval and reuse tools, content review, literature search, citations, and reviewable document editing.

Engineered for production

Built multi-agent orchestration, hybrid retrieval and reranking, structured validation, resilient model routing, editing APIs, and end-to-end evaluations.

Production AI agents for biostatistics study planning

I joined after Aligrade acquired Grantease's technology and led customer discovery, solution design, and production delivery for an AI study-planning product. I owned architecture, roadmap, hiring, and engineering execution while leading the AI engineering and product team.

Discovery → product

Interviewed biostatisticians to map how experts searched prior studies and reused approved methods, definitions, and code. Translated expert research workflows into tools for evidence retrieval, content reuse, and review for adherence, clarity, and internal consistency.

Engineering

Architected a Pydantic AI supervisor/subagent system with structured validation and editing APIs; built hybrid BM25 and semantic retrieval with parallel LLM reranking for large document collections; and made the production app agent-operable through a CLI for end-to-end evaluations, failure reproduction, and faster debugging.

Production deployments with design partners
Led AI engineering and product team

AI grant writing platform for research labs

I co-founded, built, and operated an AI grant-writing SaaS from zero to $100K ARR within three months. Researchers at 30+ universities produced 1,000+ NIH proposals before the technology was acquired by Aligrade.

Discovery → product

Their feedback informed literature search and automated citations; users reported working on as many as 10 proposals at once versus three under their previous manual process.

Engineering

Built the Next.js platform end to end: AI writing agent, collaborative editor, Supabase data layer, WorkOS enterprise authentication, and DOCX/PDF ingestion. Engineered fallbacks, structured-output recovery, and model tiering across five providers, plus citation data parsing, cleaning, and Crossref enrichment.

$100K ARR within 3 months of launch
1,000+ NIH proposals produced
Adopted at 30+ universities
60% faster proposal preparation

Enterprise AI solutions and production ML

I worked across customer-facing enterprise solution delivery and production ML for Lenovo's public PC support operations, helping win three Fortune 500 customers and handing completed systems to production engineering teams.

Discovery → product

Owned technical discovery with enterprise customers, turning their domain and infrastructure constraints into fine-tuned LLM/RAG software, evaluation plans, and hardware requirements for on-premises deployment.

Engineering

Developed an XGBoost model and SHAP explainability layer trained on tens of thousands of historical support cases, improving held-out three-class accuracy from 65% to 95%. The scheduled production workflow scored every survey-eligible open support case across North America and Europe. I also built webhook monitoring and maintained localized responses and branching logic for a public-facing support chatbot.

Helped win 3 Fortune 500 customers
Production ML deployed across North America + Europe

Public engineering artifact

Editor MCP Public repository · Demo-ready

A collaborative document workspace for people and their AI agents.

I built a TypeScript MCP server and Tiptap/Yjs editor where agents make atomic, reviewable document changes instead of pasting text into chat. People can inspect, navigate to, accept, reject, or rewrite every proposed edit.

View source and demo

01
Versioned semantic tools with revision guards and atomic edit groups
02
Live shared state through Tiptap, ProseMirror, Yjs, and Hocuspocus
03
Authenticated Streamable HTTP with OAuth and fail-closed validation

More building

Proto-Town Hardware Incubator · 2024

Built the place where other founders could build.

I helped launch a hardware-focused startup ranch, secured its first startup tenants, and built the prototyping workshop.

Research & writing

First author · TPCTC 2024

Evaluation Considerations of Synthetic Natural Language Datasets for Question Answering Applications

Designed and ran an experimental pipeline that generated synthetic question-answer datasets with four open-source LLM configurations, fine-tuned the same base model on each, and tested which dataset signals predicted downstream QA performance.

Found that semantic similarity can indicate domain diversity and that short chain-of-thought answers may flag lower-quality training data.

Springer Author copy

Co-author · TPCTC 2023

Benchmarking Large Language Models: Opportunities and Challenges

Co-authored a Lenovo and AMD review of how to compare LLM systems across training and inference—not only on accuracy and throughput, but also cost, energy, bias, trust, and sustainability.

Published by Springer in Lecture Notes in Computer Science; the chapter has been cited in subsequent benchmarking work.

Springer Author copy

Lenovo Press · Three-part engineering series

Making LLMs Work for Enterprise

I co-authored a practical series for enterprise teams building domain-specific RAG systems, from model selection and evaluation through synthetic dataset creation, LoRA fine-tuning, hardware sizing, latency, and production tradeoffs. I also developed a containerized framework that customers could use to implement the approach.

Model training experiments

Graduate AI coursework · 2024

Transformers under noisy sequence data

Designed a card-shuffle simulator and trained encoder-only PyTorch transformers to test how model depth and increasing randomness affect ace-sequence prediction. Deeper models learned subtle structure in low-noise sequences; performance degraded as shuffle complexity increased.

PyTorch
Transformers
Synthetic data

Read the experiment

Graduate AI coursework · 2024

Imitation learning in a real-time game

Collected gameplay trajectories and trained PyTorch policies with behavioral cloning and DAgger for two-agent control. The strongest independent-policy pair scored 18 goals across 32 evaluation games and held the same scoring rate against unseen agents.

PyTorch
Imitation learning
DAgger

Read the experiment

Background

Education

University of Texas at Austin

M.S. in Artificial Intelligence · Part-time · 2024–Present

Duke University

B.S.E. in Electrical & Computer Engineering · Certificate in Decision Science · 2017–2021

Technical toolkit

Languages: Python, TypeScript, SQL
AI & ML: PyTorch, Hugging Face Transformers, Pydantic AI, MCP, Langfuse, scikit-learn, XGBoost
Web & APIs: FastAPI, Node.js, Next.js, React
Data & infrastructure: AWS, Docker, Redis, PostgreSQL, pgvector, Supabase

Beyond the work

I am from Philadelphia, PA. I studied at Duke University. I currently live in Denver, CO.

I like playing and designing strategy board games, running, and tennis. My most useless skill is Etch A Sketch.

Contact

pcvbjr@gmail.com Message me on LinkedIn