01 / 03
Personal Project
RabbitHole
An event-driven video platform where uploads trigger an AI-powered processing pipeline — transcoding, speech indexing, and semantic search across your entire library.
ECS Fargate · EventBridge · SQS · Lambda · ffmpeg · Claude Vision · AWS Transcribe · Terraform · OIDC CI/CD · React · FastAPI
- Chose ECS Fargate over Lambda for transcoding because ffmpeg workloads are CPU-bound and long-running. Lambda's 15-minute limit and cold starts would have killed the experience.
- Decoupled uploads from processing using EventBridge and SQS so the API layer never blocks. The fleet scales to zero between jobs and the user gets live WebSocket status throughout.
- Three AI services (Claude Vision, Transcribe, vector embeddings) had to work together without any one failure breaking the pipeline. Built each as an independent async step so partial failures degrade gracefully.
rabbithole.stephsimmons.dev
02 / 03
Co-Founded · Proprietary
QuicLens
A multi-tenant research analytics SaaS for neurodegenerative disease labs — replacing manual R scripts and Excel with structured, cloud-native plate analysis and audit-ready records.
AWS CDK · Lambda (Docker) · API Gateway · DynamoDB · S3 · CloudFront · WorkOS SSO · Anthropic API · React · TypeScript · FastAPI
- The hardest constraint wasn't technical, it was that the math had to be right. Built a 92-test scientific validation suite verifying every calculation against Sara's researcher-validated ground truth before a single user touched it.
- Multi-tenancy meant one misconfigured query could leak one lab's data to another. Enforced isolation at the partition key level in DynamoDB and at the IAM boundary, not just in application logic.
- Lambda's managed runtime silently failed to load ARM-incompatible scientific libraries with no error at deploy time, only at invocation. Switched to Docker container deployment to control the runtime environment and eliminate the dependency on Lambda's assumptions.
quiclens.stephsimmons.dev
03 / 03
AWS Academy Capstone
PureZen
A fully serverless AI spa booking platform with a conversational concierge that takes guests from service discovery to confirmed appointment, grounded in live availability so it physically cannot hallucinate an open slot.
AWS CDK · Lambda · API Gateway · DynamoDB · S3 · CloudFront · Anthropic Claude · Ollama · React · FastAPI
- PureZen started inside an AWS Academy sandbox with no GPU access and a capped budget, so inference ran on self-hosted Ollama on a CPU-only instance. Generation took 10 to 40 seconds, which is exactly what drove the regex-first routing design. Rebuilding outside those constraints meant replacing Ollama with Anthropic Claude API, which removed the bottleneck and made the fully serverless architecture possible.
- The admin assistant is built on Anthropic tool use in a structure inspired by the Model Context Protocol: a clean separation between the model, a registry of typed data operations, and the live database behind them. The model decides which data to fetch. The tools guarantee it is always real.
- The concierge never answers from model memory. Services resolve against the real catalog, availability comes from DynamoDB, and booking actions are deterministic state machines. Claude handles language understanding and phrasing. The regex-first router keeps it off the hot paths entirely for well-formed requests.
purezen.stephsimmons.dev
