~/sameer-batra

Sameer Batra · Data scientist and AI engineer

I build AI systems people depend on.

3+ years of AI/ML experience

Industry & applied research · Washington, DC

I've built a RAG assistant for UK customs declarations, a threat-scoring system for analysts at FGS Global, and open-source skill-extraction tools at George Washington University.

sameer@portfolio: ~
$ ▋

Experience

Roles, projects and the skills behind them.

Pick any piece of work or skill to see what it connects to.

Full experience →
  • PERSONAL PROJECT

    ATLAS →

    Voice agent that has run my ML training 24/7 since October 2026, through an MCP server I built from scratch. In a live test it phoned me 32 seconds after a crash.

    Agents & MCPKnowledge graphsRAGSecurityEvaluationCloud
  • RESEARCH SPECIALIST · GWU

    LAiSER →

    Knowledge graph of 18K+ skills that lifted LLM faithfulness from 0.65 to 0.83. Mentored five interns.

    Knowledge graphsAgents & MCPFine-tuningEvaluationCloudTeaching & mentoring
  • DATA SCIENCE FELLOW

    FGS Global →

    Tiered threat-scoring system used daily by 10 analysts. It cut inference cost by 85–90% by sending only ambiguous cases to the costliest model.

    NLP & modellingKnowledge graphsEvaluation
  • JUNIOR DATA SCIENTIST

    Customs Declaration UK →

    Sole technical owner of a production RAG system on Azure, used by ~10,000 people while I worked there. It cut form completion time by about 60%.

    RAGCloud
  • INDEPENDENT EXPERIMENT

    Harness vs Model →

    One default setting moved relation F1 from 0.33 to 0.51 and failures from 11% to 0%.

    Knowledge graphsNLP & modellingEvaluation
  • CAPSTONE RESEARCH

    FactAgent-Bench

    Capstone benchmark, in progress: multi-agent orchestration against RL fine-tuning for factual consistency.

    Agents & MCPFine-tuningEvaluationCloud
  • RESEARCH EXPERIMENT

    Context-length extension →

    With Linear PI fine-tuning, Phi-2 perplexity on 8K-token inputs fell from 286 to 15 (lower is better).

    Fine-tuningEvaluation
  • GRADUATE INSTRUCTIONAL ASSISTANT · GWU

    GWU teaching →

    Graduate instructional assistant for Introduction to Data Mining, 30+ graduate students.

    Teaching & mentoring
  1. George Washington University · LAiSER

    Research Specialist

    Jan 2025 – Present

    Open-source skill-extraction tools, published on PyPI and used by students and university professionals. I built the knowledge graph, the fine-tuned model and the agentic system, and mentored five interns.

    Knowledge graphsFine-tuningMCPMentoring
    Read the case study →

    0.65 → 0.83

    LLM faithfulness

    Shipped on PyPI · 17 releases

  2. FGS Global

    Data Science Fellow

    Jan 2026 – May 2026

    A tiered threat-scoring system that 10 analysts used every day. Only the ambiguous 10–15% of cases went to the costliest model.

    NLPEvaluationCost optimisation
    Read the role details →

    85% lower

    inference cost

    Used daily by 10 analysts

  3. George Washington University

    Graduate Instructional Assistant

    Aug 2025 – Jan 2026

    Supported Introduction to Data Mining for 30+ graduate students.

    TeachingData mining
    Read the role details →

    30+

    graduate students

    Teaching and project coaching

  4. Customs Declaration UK

    Junior Data Scientist

    Jan 2023 – Jan 2025

    A production RAG assistant on Azure that turned UK customs regulations into plain-English guidance. About 10,000 people used it while I was its sole technical owner.

    RAGAzureQdrantOwnership
    Read the role details →

    ~60% less

    form-completion time

    ~10,000 users · sole technical owner

In daily use

The agent that runs my ML training.

ATLAS

Personal project · Oct 2026 – present

A voice AI agent that has run my ML training around the clock since October 2026. It monitors and queues jobs across four cloud GPU machines, catches errors before a run starts, diagnoses crashes and phones me with the cause. When I say yes on the call, it starts the next job, stops a machine or prepares a fix.

24/7
in use since October 2026, across four cloud GPU machines
~50 h
of monitoring and scheduling work it has taken off me
32 s
from crash detected to phone ringing (11 s for a known crash)
90 / 90
differently worded requests routed to the right tool in evals
AgentsMCPSecurityEvals
Read the case study and watch a crash replay →

Master of Science in Data Science · George Washington University · Jan 2025 – Present

Research and education →