AI/ML Engineer building production-oriented machine learning systems.

ChurnOps

Problem

Churn models only create value if they survive production — training must be tracked, predictions served reliably, and data drift caught before decisions silently degrade.

Engineering

• •Train Random Forest and Logistic Regression on Telco churn data (80/20 stratified split)
• •Track parameters and metrics per run in the churn-prediction MLflow experiment
• •Register versioned artifacts — model, preprocessor, reference distributions — synced from Hugging Face Hub on startup
• •Serve single and batch predictions through FastAPI: POST /predict, POST /predict_batch, GET /health, GET /metrics
• •Monitor PSI and Jensen-Shannon drift over a sliding window (default 500) of recent predictions from the SQLite log

Evidence

• •19-field Pydantic request validation with typed prediction, health, and metrics responses
• •Drift status thresholds (OK / WARNING / ALERT) with Slack webhook alerts on ALERT
• •Streamlit dashboard with Plotly gauge chart and live drift monitoring sidebar
• •Docker Compose deployment (API + MLflow server) with a persistent prediction-log volume
• •GitHub Actions CI: format check, pytest suite, and Docker build verification
• •Structured JSON logging to stdout

RAGNar

Problem

Generic chatbots guess — teams need answers grounded in their own documents, so uploads must become cited, verifiable responses instead of plausible-sounding text.

Engineering

• •Ingest PDFs and text files with automatic chunking (POST /ingest)
• •Embed chunks with OpenAI text-embedding-3-small into a persistent ChromaDB store
• •Retrieve top-k chunks above a similarity threshold for each question (POST /ask)
• •Generate grounded answers with citations via gpt-4o-mini at temperature 0, returning answer + sources + grounded flag
• •Evaluate automatically with the eval/ suite for RAG quality

Evidence

• •Retrieval recall@5: 97.06% and grounding accuracy: 100.00% on the automated evaluation baseline
• •Answer quality of 4.88 with 0 unparseable responses in the same baseline
• •Methodology in the eval/ suite — reproducible with make eval; see the evaluation baseline in the repository README
• •Streamlit UI for document upload and Q&A over a FastAPI + ChromaDB backend
• •Docker Compose multi-service deployment with CI/CD via GitHub Actions

Banks ETL

Problem

A coursework ETL only convinces when it ships like production software — installable, logged, tested, and containerized instead of a one-off script.

Engineering

• •Extract — scrape the archived Wikipedia table of largest banks by market cap
• •Transform — convert USD market cap to GBP/EUR/INR via a live rates API with CSV fallback
• •Load — write CSV and SQLite (Largest_banks table, replaced each run)
• •Query — three SQL queries answering London/GBP, Berlin/EUR, and New Delhi/INR rankings

Evidence

• •Installable CLI: banks-etl full run plus banks-etl-cli run / query --city subcommands
• •Live Frankfurter exchange-rate API with graceful fallback to a cached CSV
• •Structured logging to logs/code_log.txt and stdout
• •Hermetic pytest suite: 24 tests with fixtures and mocked HTTP, flake8 clean
• •GitHub Actions CI: lint, test, and coverage on push and pull request
• •Dockerized with named volumes persisting the database and logs
• •Modular typed package with single-source config in banks_etl/config.py