Production AI systems for markets, sports, and the enterprise. Three divisions, two operators, one company. Led by Robert D. Scott II (CEO) and Mark Conway (Director of AI).
Recent and prior engagements
What we build
End-to-end AI system design — architecture, model selection, production hardening, and cost modeling. The engineering practice that ties everything below together.
Bedrock · Vertex AI · Azure AI
From document ingestion through vector stores (OpenSearch, pgvector, FAISS) to retrieval-augmented chat with history and citations.
OpenSearch · pgvector · FAISS
LangGraph orchestration, tool use, multi-step reasoning, evaluation harnesses with Patronus AI or custom judges.
LangGraph · Patronus · Bedrock
OpenAI vision, image-to-structured-annotation flows, hybrid image and text search.
OpenAI Vision · Hybrid retrieval
PortKey gateways, LangFuse observability, model routing across Bedrock, OpenAI, Anthropic, and OSS.
PortKey · LangFuse
Recommender systems, demand forecasting, record linkage, semantic search with learning-to-rank.
XGBoost · Nixtla · AutoML
Recent engagements
Visual AI for in-store compliance
Visual AI models that measure in-store compliance across the chain — store openings, pylon placement, end-of-season resets. Adjacent work in Databricks: GenAI for costume fitting, plus ML sales and shipment forecasting.
Databricks · OpenAI Vision · Postgres
Four Generative AI Products. Shipped.
Advised the C-suite on AI strategy and shipped four products over the engagement: Report Pro, Consumer Preferences, Menu Trends, and a chatbot. Built a Bedrock Agent pipeline that turned thousands of PowerPoint, PDF, and Excel reports into a searchable Knowledge Base — OpenSearch vectors, Foundation Models for parsing, Claude and GPT-4 for generation. Added hybrid image-and-text search via OpenAI vision, and engineered an agentic React + TypeScript chat with LangGraph orchestration and LangFuse observability.
Bedrock · OpenSearch · Claude · GPT-4 · LangGraph · LangFuse
Scalable agentic healthcare system on GCP
Designed a scalable agentic healthcare system on GCP with LangChain and LangGraph. Deployed with GitHub Actions and Workflows. Load-tested with Kubernetes and Locust, then tuned Firestore throughput under peak traffic.
GCP · LangGraph · GitHub Actions · Kubernetes · Firestore
RAG pipeline for a consumer GenAI chatbot
Sentence-transformer embeddings, Postgres pgvector on AWS RDS, LangChain chat with conversation memory. Custom Patronus AI evaluators graded competing LLMs in a GitHub Actions regression suite. Kotlin microservices on EKS with Helm, Argo, PortKey AI gateway.
pgvector · LangChain · Patronus · PortKey · EKS
27% recommendation lift at 3 ms inference latency
Semantic search on T5 + SentenceTransformers + FAISS, with XGBoost learning-to-rank. Neural Collaborative Filtering on Azure Databricks scaled 20× via Petastorm and TorchDistributor. Hourly demand forecast across 2,000 stores on Nixtla.
T5 · FAISS · XGBoost · Databricks · Nixtla
Recommenders that funded a multi-year program
Built AzureML personalization and recommender models for an AI-driven shopping platform. The architecture blueprint was instrumental in securing several years of project funding. Implemented affinity and behavioral scoring on Azure Databricks using synthetic data, calculating Shapley values for visual feature-importance explanations. Wrote extensive model-validation Jupyter notebooks for the ensembles.
AzureML · Databricks · Shapley · Jupyter
Record linkage that unlocked $75M in contracts
Created synthetic, anonymized datasets from merchant transactions — work that spurred $75M in potential contract deals. Eliminated manual record matching with automated PySpark linkage over name, address, and merchant fields using tokenizers, N-grams, and Locality Sensitive Hashing. Generated demographic predictions with a Spark random forest classifier over consumer transactional features, extracting distribution features with vector assemblers, bucketizers, and custom UDFs.
PySpark · LSH · Random Forest · Synthetic Data
$2.5M absence reduction across six auto plants
Led a small team that cut unplanned absences at six North American auto plants — saving $2.5M in employment costs using gradient boosting and time-series models over event and weather data. Built a Palantir Foundry pipeline that streamed vehicle sensor data (SQDF, Witech, Vstat, Data Logger), compressed it with Dynamic Time Warping, and chained LSTMs with Chi-Square analysis to predict warranty repairs. Presented a non-parametric Monte Carlo production-loss simulation to the CTO, contrasted against a parametric negative-binomial fit.
LSTM · Palantir Foundry · Gradient Boosting · Monte Carlo
A dozen ML models for Sam's Club on Hadoop
Led onshore and offshore teams to deliver over a dozen production models to Sam's Club on their Hadoop platform — demand forecasting, K-means customer segmentation, multinomial propensity scoring, churn/renewal/attrition (libsvm), basket analysis (R arules), X-13 ARIMA seasonality, and a dynamic-bidding offer-assignment algorithm. Ran extensive A/B experiments to measure uplift and statistical power on member campaigns. A greedy algorithm built in Apache Spark gave a 50× speedup on offer assignment.
Hadoop · Spark · libsvm · ARIMA
How an engagement works
A written one-pager with proposed scope, milestones, and a fixed-fee or T&M option.
A working slice of the system, checked into your repo, running against your data.
CI/CD, evals, observability, cost controls, handoff docs.
Tuning, model updates, and incident response on a steady cadence.
Not a fit if
Tell us what you're building. We'll tell you whether we can help within a day.
