Skip to content

ML Engineer / AI Solutions Architect

Krishna MihirTatavarthi

Building ML systems that ship
View projectsGet in touch

About

I build AI systems that survive contact with production — LLM agents, RAG pipelines, and the data infrastructure underneath them.

I'm finishing my MS in Computer Science at UMBC (GPA 3.93), where my trajectory has run from classical ML research into cloud-native AI engineering: shipping agentic workflows with LangGraph and LangChain, fine-tuning models with LoRA, and wiring it all into secure, role-based APIs.

Before the model there is always the data. I've built batch and streaming ETL/ELT pipelines on AWS, orchestrated with Airflow and dbt, and I care as much about observability and data quality as I do about model accuracy. My IEEE-published research on detecting machine-generated text sits right at the intersection I like most: rigorous ML with a real-world stake.

Krishna Mihir Tatavarthi
KM.T · PortraitHover for color
MS Computer Science
UMBC · GPA 3.93 · 2026
Research
IEEE-published, NLP & ML
Based in
Baltimore, MD
Focus
LLM systems · Data platforms · Cloud

Stack

  • LangChain
  • LangGraph
  • AI Agents
  • RAG
  • LLM APIs (OpenAI, Anthropic)
  • LLM Fine-Tuning (LoRA)
  • Hugging Face
  • PyTorch
  • scikit-learn
  • BERT
  • Pinecone
  • Prompt Engineering
  • PySpark
  • Apache Spark
  • Apache Airflow
  • dbt
  • BigQuery
  • Snowflake
  • PostgreSQL
  • FastAPI
  • ETL Pipelines
  • AWS Glue
  • AWS Athena
  • S3
  • AWS
  • Docker
  • Kubernetes
  • GitHub Actions
  • CI/CD
  • Supabase
  • Linux
  • Git
  • Python
  • SQL
  • TypeScript
  • JavaScript
  • C++
  • React
  • Next.js
  • Node.js

Projects

Real-time observability for data pipelines — catches bad data and failed runs before anyone downstream notices.

  • Built on a bronze → silver → gold medallion architecture with FastAPI, PostgreSQL, Airflow, and dbt, handling 10,000+ events per day at sub-200ms latency.
  • Automated data-quality checks intercept bad records before they propagate; a live Next.js dashboard streams pipeline health over WebSockets and flags anomalies as they happen.
  • Containerized with Docker and shipped through Kubernetes with GitHub Actions CI/CD, cutting failure detection from hours to under 10 seconds.
View on GitHub
10k+
events / day
<10s
failure detection
<200ms
event latency
Next.jsFastAPIPostgreSQLAirflowdbtDockerKubernetes

Experience

  1. Velociti Inc.

    Feb 2026 — May 2026
    Software Engineer Intern · Phoenix, AZ
    • Shipped 10+ production features for a customer-facing web app in React and TypeScript on Supabase and AWS CDK, improving reliability and user-facing performance by ~48%.
    • Built agentic AI workflows with LangGraph, LangChain, and RAG over Gemini and OpenAI-compatible LLMs, generating structured product-strategy artifacts from unstructured input via Supabase Edge Functions and PostgreSQL RPCs.
    • Enforced secure API design across the stack: authentication checks, input validation, and role-based access control.
    ReactTypeScriptLangGraphSupabaseAWS CDK

  2. Date Maroon

    Sep 2025 — Jan 2026
    Software Engineer Intern · Florida, USA
    • Built batch and streaming ETL/ELT pipelines on AWS (Athena, Glue, S3) with Python, SQL, and PySpark, turning raw data into analytics-ready tables stakeholders could trust.
    • Rebuilt the slowest SQL behind Mode Analytics dashboards — reports loaded 50% faster while doubling queryable history from 3 to 6 months.
    • Integrated LLM-powered SQL generation (OpenAI API) into pipeline development, cutting exploratory query development time by 60%.
    PythonPySparkAWSSQLMode AnalyticsOpenAI API

  3. UMBC

    Jan 2025 — May 2025
    Graduate Research Assistant · Baltimore, MD
    • Conducted pulsar candidate classification research on the HTRU2 radio telescope survey dataset in coordination with Dr. Milton Halem (formerly of NASA), building an end-to-end ML pipeline with SMOTE-based imbalance handling that lifted recall from 82.3% to 89.6% at 0.97 ROC-AUC.
    • Validated the trained models beyond the benchmark dataset in collaboration with astrophysicist Gnanesh (postdoctoral researcher), who stress-tested them on independent observational data — the models held up at 91% accuracy.
    Pythonscikit-learnPandasSMOTE

  4. Salesforce

    Apr 2023 — Jun 2023
    Virtual Salesforce Developer · Hyderabad, India
    • Built 3+ enterprise apps with Apex, REST APIs, and Lightning Web Components, automating manual workflows; earned Apex Specialist and Process Automation Specialist super badges.
    ApexREST APIsLWC

Contact

Open to ML engineering and AI architecture roles. If you're building something that needs models in production, let's talk.