AI Case Studies
Four AI systems our engineers have designed and delivered.
From a production machine-learning system for the U.S. Courts to GenAI on Amazon Bedrock. Outcomes are described qualitatively; we publish metrics only with customer approval.
One Method Behind Every Case
Each system below was built with the same loop, and each case shows where a person stays in it.
- PlanA person approves the plan
- ActScoped, isolated work
- ObserveTests and logs, inside your boundary
- EvaluateA person reviews the diff
Federal Civilian · Production AI/ML
AI-Driven Threat Detection for the U.S. Courts
U.S. Courts – Ninth Circuit, via Bent Ear Solutions
The Challenge
Protecting federal judges means keeping up with a fast-growing volume of online content, including social media and dark-web sources, that manual threat assessment could not keep pace with.
What We Built
- A SaaS application integrated with commercial threat-assessment feeds through APIs for continuous data ingestion
- A machine-learning sentiment-analysis model that classifies content as a potential threat or non-threat
- An analyst application to review classified results, confirm actual threats and export them for coordination with the U.S. Marshals and other agencies
- A feedback loop: analysts reclassify mistakes, and the model is retrained on their corrections
How the Loop Stays Reviewable
Human analysts stay in the loop. The model flags; people decide, and their decisions make the next model better.
Outcome
Automated first-pass triage of large data volumes, with analysts focused on the items that matter.
Technology
Machine learning, sentiment analysis, API integration, SaaS, continuous retraining
Navy · GenAI Prototype · 2024
AI-Enabled Proposal Evaluation on Amazon Bedrock
NAWCTSD Prize Challenge, semi-finalist (rapid prototype)
The Challenge
Technical evaluations of sole-source proposals were manual, slow and inconsistent, with little use of historical data.
What We Built
- A workspace-based evaluation app where uploaded proposals are parsed and summarized by models on Amazon Bedrock
- Retrieval-augmented generation (RAG) over past proposals and contract documents, using PostgreSQL with pgvector
- Answers to Technical Evaluation Checklist questions with citations back to the source documents
- Serverless processing with AWS Lambda and Step Functions, designed for AWS GovCloud (US)
How the Loop Stays Reviewable
Every AI answer cites its sources, so evaluators can check the work rather than trust it.
Outcome
A working prototype that reached the semi-finalist round, and the start of the Bedrock engineering that led to Topsoil.
Technology
Amazon Bedrock, RAG, pgvector, AWS Lambda, Step Functions, AWS GovCloud (US), Angular, Java / Spring Boot
DoW-Funded · GenAI Training
GenAI Roleplay and Performance Analytics for Interpersonal-Skills Training
I2ST (Immersive Interpersonal Skills Trainer), led by Design Interactive
The Challenge
Equal Opportunity professionals need realistic practice for sensitive workplace conversations, at scale and with consistent feedback.
What We Built
- Persona-driven roleplay using Azure OpenAI Service and LangChain, with configurable attributes such as hesitant or combative
- Post-scenario evaluation: LLM + RAG with pgvector scores transcripts against a rubric of key performance attributes and generates specific feedback
- Containerized services on Azure with CI/CD, automated testing and infrastructure as code
How the Loop Stays Reviewable
Rubric-based evaluation turns each AI conversation into structured, reviewable feedback rather than an opaque score.
Outcome
Phase I MVP delivered, with engineers embedded side by side with Design Interactive.
Technology
Azure OpenAI, LangChain, RAG, pgvector, containers, CI/CD, IaC
Commercial · GenAI Engineering
GenAI-Assisted Engineering Software for Detecting Gas Loss
Dominion Energy
The Challenge
Field engineers needed faster, more reliable fluid-dynamics calculations to determine pipeline gas loss.
What We Built
- A field-ready SaaS portal for engineers
- Implementations of the Colebrook-White, IGT-Improved and AGA fully turbulent flow equations
- Generative AI integrated into the calculation workflow to improve the speed and accuracy of analysis
How the Loop Stays Reviewable
Established engineering equations provide the ground truth; AI accelerates the work around them.
Outcome
Engineers can access data and run gas-loss calculations on site, through the portal.
Technology
Generative AI, scientific computing, SaaS, cloud
Have an AI Idea That Has to Stay Inside Your Boundary?
We will help you scope it, build it, and keep it reviewable.