PS Resume

Generative AI Engineer at Pumo Technovation, Coimbatore

Praveena S

I build RAG assistants, LangGraph agents and MCP servers that answer data centre operations questions with sources you can check, then ship them with FastAPI, Docker and AWS.

  • RAG
  • LangChain
  • LangGraph
  • MCP
  • AI Agents
  • Hybrid Search
  • RAGAS
  • Langfuse
  • Python
  • FastAPI
  • Qdrant
  • Pinecone
  • ChromaDB
  • PostgreSQL
  • Docker
  • AWS
  • GPT-4
  • Gemini
  • Claude

Three systems that work together

A RAG assistant and an agent, each useful on its own, plus an MCP server that lets any MCP client call both and query the operations database safely. Code is on GitHub.

DC-Ops Copilot

Production RAG assistant

Answers data centre operations questions from runbooks, SOPs and incident reports, and cites the source for every answer.

  • Hybrid search (BM25 + vector) with cross-encoder reranking and metadata filtering
  • Refusal guardrail for out-of-scope questions
  • RAGAS evaluation on a 30-question golden set, comparing chunk size and top-k configurations
  • Semantic caching for repeated questions, with Langfuse tracing for per-query latency and cost
  • Deployed in Docker on AWS EC2, with source documents in Amazon S3
  • LangChain
  • Qdrant
  • FastAPI
  • Streamlit
  • RAGAS
  • Langfuse
  • Docker
  • AWS
View code on GitHub

EnergyGuard

Energy bill intelligence agent

A LangGraph agent that reads electricity bill PDFs, flags cost and usage anomalies, and forecasts spend.

  • Structured extraction of bill fields from PDFs into Pydantic models
  • Anomaly detection on cost and usage, with spend forecasting in ARIMA
  • Human-in-the-loop approval for flagged bills, with checkpointing so interrupted runs resume safely
  • Field-level accuracy measured against ground truth from synthetic bills
  • LangGraph
  • Pydantic
  • ARIMA
  • FastAPI
  • Streamlit
View code on GitHub

DC-Ops MCP Server

AI tools for data centre analytics

Exposes a PostgreSQL operations database, the RAG assistant and the bill-analysis agent as tools any MCP client can call.

  • Read-only SQL guardrails: a read-only database role, SELECT-only validation and row limits
  • Text-to-SQL accuracy evaluated on a 40-question test set
  • Microservices: RAG API, agent API, MCP server and PostgreSQL run in containers with Docker Compose
  • MCP
  • PostgreSQL
  • Python
  • Docker Compose
View code on GitHub
How the three systems connect An MCP client calls the DC-Ops MCP server. The server exposes three tools: ask_runbooks runs the DC-Ops Copilot RAG pipeline (hybrid search over runbooks in S3, reranking, a scope guard, then an answer with sources); analyze_bill runs the EnergyGuard agent (bill PDF extraction, anomaly checks, human approval, ARIMA forecast); query_ops_db turns a plain-English question into SQL, checks it is SELECT-only with row limits, and runs it on PostgreSQL with a read-only role. MCP client Claude, IDEs DC-Ops MCP server Docker Compose DC-Ops Copilot (RAG API) ask_runbooks Hybrid search runbooks in S3 Rerank cross-encoder Scope guard refuses off-topic Answer with sources EnergyGuard (agent API) analyze_bill Extract bill PDF fields Anomalies cost and usage Approval by a human Forecast ARIMA Ops database (text-to-SQL) query_ops_db Question plain English Text to SQL LLM SQL guard SELECT-only PostgreSQL read-only role
AI step Service Guardrail or human check Data

Swipe the map sideways to see every step.

Earlier projects

AI Search Engine

Python, LLM APIs, SQL

A ChatGPT-style search engine that retrieves information and writes concise answer summaries with LLM APIs. It shows relevant images next to results and stores queries in SQL for search history and analytics.

Portfolio Optimization

Python, machine learning

Efficient Frontier analysis to maximise returns while managing risk, plus Hierarchical Clustering and PCA for data-driven allocations that improved diversification.

View code on GitHub

How I build AI systems

The habits behind every project above, from the first prototype to the deployed service.

  • Measure before tuning

    Golden sets decide what changes: RAGAS on 30 questions for retrieval, 40 questions for text-to-SQL, and field-level accuracy for extraction.

  • Guard the edges

    Out-of-scope refusals, read-only SQL with SELECT-only validation and row limits, and human approval before flagged results go through.

  • Trace every query

    Langfuse traces show latency and cost per query, and semantic caching stops repeat questions from costing twice.

  • Ship in containers

    FastAPI services in Docker and Docker Compose on AWS EC2 and S3, with Git, GitHub Actions and Jenkins for CI/CD.

Ask my resume

A small RAG demo you can try: BM25 search over 28 chunks of my resume, running in your browser. It answers only from what it retrieves, shows its sources, and refuses questions my resume doesn't cover.

  1. 1TokenizeWaiting for a question
  2. 2Score with BM25Ranks every chunk
  3. 3Boost titlesFavours title matches
  4. 4Answer or refuseNeeds a score of 1.6
Pick a question above or type your own. Answers come only from the chunks the search retrieves.

Experience

From analytics and forecasting into LLM engineering, newest role first.

  1. Jun 2026 – PresentCoimbatore

    Generative AI Engineer

    Pumo Technovation India Pvt Ltd

    • Built RAG pipelines with LangChain that give accurate, context-grounded answers from internal enterprise documents in a data centre environment.
    • Implemented vector search with ChromaDB and Pinecone, covering document ingestion, chunking and embedding generation.
    • Tuned chunk size, overlap and top-k to improve answer relevance and retrieval accuracy.
    • Applied prompt engineering to improve response relevance and reduce hallucinations.
  2. Dec 2025 – Jun 2026Coimbatore

    Data Analyst Trainer

    Qtree Technologies

    • Built hands-on data analytics, machine learning and Generative AI examples in Python and SQL, including LLM-based applications.
    • Worked across data analysis, ML model building and LLM application development using modern AI tools.
  3. May 2024 – Jul 2025Chennai

    Data Analyst

    Arcadia (Urjanet Energy Solutions Pvt Ltd)

    • Built AI-enhanced analytics, including real-time chatbot improvements with GPT-4 that generated customer engagement insights.
    • Developed ARIMA and trend forecasting models to support business planning.
    • Turned raw data into Power BI and Tableau dashboards and reports with SQL and Excel for faster decisions.
    • Automated repetitive data workflows with Jenkins, reducing manual effort and reporting errors.
    • Collaborated with cross-functional and international teams on strategic initiatives.
  4. Sep 2023 – Dec 2023Coimbatore

    Data Science Intern

    Aapin Solution Limited

    • Developed automated data-cleansing scripts, reducing processing time by 20% and improving data accuracy.
    • Applied data analysis and machine learning in Python to real business problems.
    • Performed exploratory analysis and model building with Python, SQL and Tableau in project teams.
  5. Mar 2023 – Sep 2023Coimbatore

    AI Engineer

    Vegas Project Management Service

    • Enhanced chatbots with GPT-4 and NLP, including a ChatGPT-style assistant grounded in a custom company knowledge base.
    • Designed and built a Flask web app for AI-driven user interactions, increasing user engagement by 25%.
    • Delivered sentiment analysis and predictive analytics projects.

Skills

Grouped by how I use them: the AI layer I design, the engineering that runs it, and the analytics background I came from.

AI layer

Models, agents, retrieval and the checks around them.

Generative AI & LLMs

  • Large Language Models
  • OpenAI GPT-4
  • Google Gemini
  • Anthropic Claude
  • Prompt Engineering
  • Structured Outputs
  • NLP

Agentic AI

  • LangChain
  • LangGraph
  • AI Agents
  • Agent Orchestration
  • Model Context Protocol (MCP)
  • Tool Calling
  • Human-in-the-Loop

RAG & vector databases

  • Retrieval-Augmented Generation
  • ChromaDB
  • Pinecone
  • Qdrant
  • Embeddings (Hugging Face)
  • Semantic Search
  • Hybrid Search
  • Reranking

Evaluation & observability

  • LLM Evaluation
  • RAGAS
  • Langfuse
  • Guardrails

Engineering

The APIs, databases and pipelines that put models into service.

Programming & APIs

  • Python
  • SQL
  • FastAPI
  • Flask
  • Streamlit
  • REST APIs

Databases & deployment

  • MySQL
  • PostgreSQL
  • Docker
  • AWS (EC2, S3)
  • Git
  • GitHub
  • GitHub Actions (CI/CD)
  • Jenkins

Analytics

Where I started, and still how I check whether a model helps the business.

Data & analytics

  • Pandas
  • NumPy
  • Time Series Forecasting (ARIMA)
  • Power BI
  • Tableau
  • Excel

Education

An MBA in Business Analytics, so I connect model choices to business outcomes, plus certifications in data analytics, machine learning and LLM applications.

Degrees

  • MBA, Business Analytics & FinanceKV Institute of Management and Information Studies2021 – 2023
  • B.Com, Computer ApplicationsSNMV College of Arts and Science2018 – 2021

Certifications

  • Postgraduate Program in Data Analytics with Machine LearningImarticus Learning2023
  • Generative AI & LLMs: Building ChatGPT with a Custom Knowledge BaseVegas Project Management Service
  • SQL in Data ScienceGreat Learning
  • Data Visualization with Power BIGreat Learning

Have a RAG or agent problem? Let's talk.

I'm happy to walk through any of these systems, from retrieval tuning to guardrails and deployment.