AI Systems Portfolio
I build AI systems
Built for Compliance
Eight production AI systems across four regulated verticals: financial services, healthcare, education, and government. Every system ships with governance, compliance, and audit-ready documentation.
HIPAA · FedRAMP · FERPA · SR 11-7 / SR 26-2
Amazon Connect
Bedrock | Multi-Tenant
ConnectIQ — AI-Powered Call Center SaaS
B2B SaaS replacing traditional receptionists with AI voice agents. Post-call summarization, intelligent routing, escalation prediction, and sentiment analysis. Multi-tenant architecture with Stripe billing.
Amazon Connect
Lex
Bedrock
Lambda
DynamoDB
Cognito
Stripe
CDK
MCP Server
CI/CD
OpenAPI
Model Card
ADRs
ML Pattern
Voice-to-text with AI routing, sentiment analysis, escalation prediction
RAG Pattern
Agentic RAG — Lex intent + Bedrock routing with knowledge base retrieval
Architecture
Amazon Connect → Lex → Lambda → Bedrock → DynamoDB → Stripe billing
Business
Multi-tenant B2B SaaS with per-seat billing and enterprise features
Open Source
npm | TypeScript
@johnathan-horner/cdk-ai-constructs
Reusable AWS CDK construct library powering all eight AI systems. Eight production constructs including Bedrock agents, SageMaker serverless endpoints, multi-tenant auth, auditable storage, Stripe billing, and model cards.
CDK
TypeScript
Bedrock
SageMaker
Multi-tenant
npm
MCP Server
CI/CD
OpenAPI
Model Card
ADRs
Pattern
Reusable CDK constructs for AI infrastructure: agents, endpoints, auth, billing
Architecture
TypeScript CDK L3 constructs for common AI patterns
Impact
Powers all eight AI systems with consistent patterns and best practices
Healthcare
HIPAA
Medical Image Triage System
CNN classifier that triages medical images by confidence level, routing to the right physician queue with a full HIPAA-compliant audit trail. The model is 20% of the story. The production system around it is 80%.
TensorFlow
EfficientNetB0
SageMaker
Lambda
DynamoDB
SNS
Cognito
KMS
MCP Server
CI/CD
OpenAPI
Model Card
ADRs
ML Pattern
Computer Vision, CNN transfer learning with confidence-based triage routing
RAG Pattern
Multimodal RAG — vision model with confidence-based routing
Architecture
S3 ingest → Lambda → SageMaker endpoint → SNS routing → DynamoDB audit trail
Governance
No raw image retention post-classification, KMS encryption, Cognito RBAC, drift monitoring
Click to expand
Financial Services
SR 11-7 / SR 26-2
Transaction Anomaly Detection
Autoencoder flags suspicious transactions in real-time. When it flags one, a LangGraph investigation agent pulls customer history, checks merchant risk, and generates a natural language investigation summary. Not just a score, but an explanation.
PyTorch
LangGraph
LangChain
Bedrock Claude
Kinesis
ECS Fargate
SageMaker
DynamoDB
MCP Server
CI/CD
OpenAPI
Model Card
ADRs
ML Pattern
Autoencoder anomaly detection + agentic investigation with feature-level explainability
RAG Pattern
Agentic RAG — 4-node LangGraph investigation loop
Architecture
Kinesis → Lambda scoring → SQS → ECS Fargate (LangGraph) → SNS fraud alerts
Governance
SR 11-7 / SR 26-2 compliant: model documentation, validation, drift monitoring, feedback retraining loop
Click to expand
Government & Legal
FedRAMP
Legal Document Classification & Risk Scoring
A multi-agent system that classifies legal documents, extracts and scores high-risk clauses, cross-references regulations, and generates plain-English attorney briefings. Parallel agent execution cuts processing time in half.
PyTorch
DistilBERT
LangGraph
LangChain
Bedrock Claude
Textract
ECS Fargate
SageMaker
MCP Server
CI/CD
OpenAPI
Model Card
ADRs
ML Pattern
Agentic RAG: DistilBERT planner + 3 parallel LangGraph agents (fan-out/fan-in) + risk scoring reasoner
RAG Pattern
Corrective RAG — DistilBERT grader validates before LLM reasoning
Architecture
S3 → Textract → SageMaker → ECS (LangGraph fan-out/fan-in) → SQS attorney queues
Governance
FedRAMP-aware: KMS encryption, CloudTrail audit, Cognito RBAC, 7-year retention
Click to expand
Financial Services
MRM, 8 Guardrails
ShootItPicks
A live, revenue-generating NFL analytics SaaS with paying subscribers. 5-step Chain of Thought AI pipeline with MRM confidence calibration enforcing 8 guardrails before any pick reaches a user. This is a business.
Anthropic Claude
LangChain
Monte Carlo CVaR
8 Lambdas
API Gateway
Cognito
DynamoDB
S3/CloudFront
Stripe
MCP Server
ML Pattern
5-step Chain of Thought pipeline with MRM confidence calibration layer
RAG Pattern
Hybrid RAG — dense + sparse retrieval with 5-step Chain of Thought
Architecture
Fully serverless: 8 Lambdas, API Gateway, Cognito auth, DynamoDB, S3/CloudFront
Business
Revenue-generating SaaS with Stripe billing, tiered subscriptions, paying customers
Click to expand
Financial Services
SR 11-7 / SR 26-2
FinTech AI
Hedge fund intelligence platform with three LangGraph agents, Monte Carlo CVaR risk engine with fat-tail distributions, and a full SR 11-7 / SR 26-2 Model Risk Management framework: 6-model inventory, PSI drift detection, CRO report generator.
Anthropic Claude
LangGraph
Monte Carlo CVaR
Bedrock
Titan Embeddings
FAISS
Streamlit
ML Pattern
Agentic RAG: 3 LangGraph agents with shared state (market analysis → risk assessment → portfolio mgmt)
RAG Pattern
Agentic RAG — Planner/Risk/Recommender agents with tool use
Risk Engine
10,000 Monte Carlo simulations, fat-tail mixture distribution, 5 historical stress scenarios
Governance
SR 11-7 / SR 26-2: 6-model inventory, independent validation, PSI drift, immutable audit log
Click to expand
Education
FERPA
EduAI Connect
FERPA-compliant education AI that helps teachers proactively identify at-risk students. Multi-agent RAG on Amazon Bedrock with LangGraph orchestration and HyDE query transformation for improved retrieval quality.
Amazon Bedrock
Claude 3 Sonnet
Titan Embeddings
LangGraph
FAISS
HyDE
CDK
ML Pattern
Agentic RAG with HyDE query transformation and LLM-based re-ranking
RAG Pattern
Agentic RAG — multi-agent LangGraph with HyDE query transformation
Architecture
Bedrock agents, FAISS vector store, LangGraph StateGraph, S3, Lambda, API Gateway
Governance
FERPA: Bedrock Guardrails PII filtering, KMS CMK encryption, least-privilege IAM
Click to expand