~/sameer-batra
← Portfolio

Sameer Batra

Data scientist and AI engineer

3+ years of AI/ML experience across industry and applied research.

Experience

George Washington University · LAiSER

Jan 2025 – Present

Research Specialist

  • Architected the LAiSER knowledge graph, unifying 18K+ skills across ESCO, O*NET, UK Skills and OSN with 11.6K links. Grounding an LLM in the graph improved faithfulness from 0.65 to 0.83.
  • Fine-tuned a 2.7B Phi-2 model (SkillPhi-R) with LoRA and 4-bit quantization for skill extraction. It won the GWU open-source award.
  • Mentored five interns: three on the knowledge graph, two on an NLP paper about phantom vacancies, now in review. Cut collection of 100K+ syllabi from 31 Texas universities to three days with an EC2 scraper.

FGS Global

Jan 2026 – May 2026

Data Science Fellow

  • Designed a tiered scoring system (DeBERTa-v3, XGBoost, GPT-4) to replace Perspective API, used daily by 10 analysts. It cut analyst review time by about 83%. Sending only the ambiguous 10–15% of cases to the costliest model cut inference cost by 85–90%.
  • Built a Neo4j knowledge graph mapping stakeholder influence across 50,000+ media records and ranked key audiences with centrality analysis.

George Washington University

Aug 2025 – Jan 2026

Graduate Instructional Assistant

  • Supported Introduction to Data Mining for 30+ graduate students: ran sessions, wrote quizzes, graded assignments and designed hands-on exercises.

Customs Declaration UK

Jan 2023 – Jan 2025

Junior Data Scientist

  • As sole technical owner, built and deployed a production RAG system on Azure with Qdrant vector search that turned UK customs regulations into guidance. About 10,000 people used it, and form completion time fell by about 60%.
  • Gave plain-English guidance across thousands of commodity codes by grounding the assistant in vector-indexed regulations.

Selected projects

ATLAS · Personal project · Oct 2026–Present. Voice agent and MCP server that has run my training jobs 24/7 across four GPU machines, saving ~50 hours of monitoring and scheduling. Crash to phone call in 32 seconds in a live test.

Harness vs Model · Independent experiment · 2026. Diagnosed a default model setting that moved relation F1 from 0.33 to 0.51 on 288 CoNLL04 test sentences.

Education

Master of Science in Data Science · George Washington University

Jan 2025 – Present

Bachelor of Science in Computer Science · SZABIST, Karachi

Aug 2020 – May 2024

Technical skills

Python · SQL · PyTorch / Hugging Face · LoRA / PEFT · RAG · Neo4j · Qdrant · MCP · Claude Code · Docker · GitHub Actions · AWS (EC2, Lambda, Bedrock) · Azure