Hi, I'm Mehdi.

A FinOps watchdog that found a $21/month zombie subscription on day one

Serverless FinOps monitoring for AWS: daily cost digests, idle-resource detection, tag enforcement, and anomaly alerts, all running for under $2/month.

• Automated daily cost and waste monitoring (spend digests, idle-resource detection, tag-compliance checks via AWS Config, and cost-anomaly detection) across four event-driven Python Lambdas on EventBridge Scheduler and SNS, pushing findings to Slack and Telegram in real time
• Persisted findings to DynamoDB and surfaced them in a React dashboard served through an API Gateway HTTP API and S3/CloudFront
• Provisioned the whole stack as code with Terraform (remote S3 state, DynamoDB locking) and shipped it via GitHub Actions CI/CD using OIDC federation, with zero static AWS credentials

A daily market brief that writes itself, for a few dollars a month

A serverless AI market brief on AWS. Pulls real price and news data, summarizes it with Claude, and runs hands-free on a daily schedule, every piece of it provisioned as code.

• Automated a daily AI market brief that pulls price and news data via yfinance and summarizes it with Claude, running hands-free on an EventBridge daily schedule
• Guarded the LLM output with automated pytest checks (non-empty, no refusals) enforced in a GitHub Actions pipeline alongside ruff linting and a terraform plan gate
• Deployed as an ARM64 container image on ECR with zero static AWS credentials via GitHub OIDC, with API keys in a single Secrets Manager secret behind a least-privilege IAM policy

An AI receptionist that answers the phone and books the meeting

A full-stack AI website on a serverless AWS backend: a React site with an AI chatbot and a Vapi voice receptionist that answer questions, book appointments, and capture leads.

• Built and deployed the site and serverless backend end to end, across three Lambdas behind API Gateway, with infrastructure as code in AWS SAM
• Cut chatbot latency and Bedrock spend by front-running a Trie-based intent classifier ahead of Claude Haiku 4.5 streaming responses, so common questions never reach the model
• Hardened it with a WAFv2 web ACL, HMAC-verified webhooks, least-privilege IAM, and Secrets Manager, over a DynamoDB data layer with Global Secondary Indexes, tested through a GitHub Actions pytest pipeline

Ask your documents anything, then measure whether it answered well

Retrieval-augmented generation over policy documents, with hybrid retrieval and cross-encoder reranking, scored against a real evaluation set rather than vibes.

• Built a RAG pipeline with hybrid retrieval (dense embeddings and BM25) and cross-encoder reranking, measured with RAGAS at 1.00 faithfulness and 1.00 context precision on a 10-question evaluation set
• Engineered the ingestion and chunking pipeline with unit tests and a GitHub Actions CI pipeline, documenting the key retrieval design tradeoffs
• Served it through a FastAPI backend with a Streamlit UI, containerized with Docker and deployed on Hugging Face Spaces

A marketing site with a chatbot that costs 90% less to run

Live marketing site for an AI agency with an embedded chatbot (Buzz) built on Claude. Streaming responses over SSE, prompt caching, per-IP rate limiting, and a full security header policy.

• Prompt caching cut inference cost roughly 90%
• Hardened with HSTS, CSP, and Permissions-Policy headers
• Cal.com booking flow and a custom WebGL shader hero