WhatsAppEnroll

Portfolio

Things I have built and shipped

The same systems I teach, running in production. Client names are left out; results are as measured on the project.

Build something like this for youLearn to build it yourself

Client and production work

Client delivery · US database company

Agentic AIOps platform for incident triage

Multi-agent system that reads production alerts, pulls the right runbooks, suggests a fix and waits for human approval before acting.

Problem. On-call engineers lost time on noisy alerts and hunting through runbooks and post-mortems during incidents.

What I built. LangGraph agents (Planner, Validator, Forecasting, Notification, Supervisor) with agentic RAG over runbooks and post-mortems, human approval gates on every high-impact action, and guardrails plus tracing on every endpoint.

  • About 60% lower MTTR on noisy incidents
  • Llama 3 and Mistral fine-tuned with QLoRA on 2M+ incident records
  • Grounded answers with citations instead of guesses
  • LangGraph
  • Anthropic MCP
  • GPT-4o
  • Pinecone
  • vLLM
  • EKS
  • LangSmith
  • NeMo Guardrails
  • FastAPI

Client delivery · FinTech company

Customer support RAG chatbot with human escalation

Production chatbot answering KYC, billing and policy questions for 80,000+ active users, handing over to a human when it is not sure.

Problem. Support agents spent most of their day on repeat tier-1 questions about KYC, billing and policy.

What I built. Hybrid retrieval with citations, PII redaction and confidence-based escalation to human agents over WhatsApp and Slack.

  • 68% first-contact resolution on tier-1 queries
  • 41% lower average agent handling time
  • 80,000+ active users
  • GPT-4o
  • Azure OpenAI
  • LangChain
  • Pinecone
  • Azure AI Search
  • FastAPI
  • WhatsApp
  • Slack

Client delivery · UK payments processor

Real-time fraud detection MLOps platform

Champion vs challenger MLOps platform for fraud detection, forecasting and customer lifetime value.

Problem. New models needed a safe, automatic way into production without silent decay or risky manual releases.

What I built. MLflow and Kubeflow Pipelines on GKE: a new model must beat the live one on precision, recall, F1 and AUC-ROC before it is promoted, with drift-triggered retraining and canary rollouts through Argo CD.

  • 18% fewer fraudulent chargebacks year over year
  • About 22% better analyst precision
  • About 99.9% deployment success rate
  • MLflow
  • Kubeflow
  • GKE
  • Vertex AI
  • BigQuery
  • Scala-Spark
  • Evidently AI
  • Argo CD
  • Terraform

Client delivery · Legal and finance teams

Document intelligence for contracts and compliance

Pipeline that pulls key clauses out of scanned and digital contracts, scores risk and writes a short brief for reviewers.

Problem. Reviewing one contract by hand took about 90 minutes.

What I built. Azure Document Intelligence and Textract for extraction, Claude for clause analysis, Qdrant for retrieval, and an evaluation harness on a 200-contract golden set with LLM-as-judge scoring.

  • Contract review down from about 90 to about 12 minutes
  • 200-contract golden dataset for hallucination tracking
  • Azure Document Intelligence
  • Claude
  • LlamaIndex
  • Qdrant
  • AWS Textract
  • Streamlit

Client delivery · Retail pilot

Multi-agent inventory and supply chain assistant

Four agents covering inventory planning, reordering and supplier coordination, with human approval at every high-cost step.

Problem. Stock-outs and over-ordering were both eating margin.

What I built. LangGraph agents on GPT-4o, Claude and Bedrock with pgvector memory and approval gates.

  • 35% fewer stock-outs in the pilot quarter
  • 22% less over-ordering
  • LangGraph
  • GPT-4o
  • Claude
  • AWS Bedrock
  • pgvector
  • FastAPI

Products and open source

Open source · Free desktop app

Orphanaut: find and delete leftover cloud resources

Desktop app for Windows and macOS that finds leftover AWS, Azure and Google Cloud resources, including AI services, before they run up a bill.

Problem. Labs and experiments leave SageMaker domains, Bedrock resources, Azure OpenAI deployments and Vertex AI endpoints running, and the bill keeps growing.

What I built. Scans each cloud, shows a map of what every resource is connected to, estimates monthly cost, and deletes in the right order after you confirm.

  • AWS, Azure and Google Cloud in one app
  • AI services like SageMaker, Bedrock, Azure OpenAI and Vertex AI
  • MIT licensed, auto-updating
  • Python
  • AWS
  • Azure
  • Google Cloud
DownloadSource code