gitesh.poudel

My journey

From catching fraud patterns to shipping AI agents

8+ years working with data, the last 5 building production ML, LLM and data systems in Python. Each step added a layer: risk analytics taught me to read data for signal, statistics taught me to measure it, machine learning taught me to model it, and LLM agents let teams talk to it.

years of professional experience
8+
years of professional experience
enterprise deals priced
40+
enterprise deals priced
revenue-per-user lift in pricing experiments
10%+
revenue-per-user lift in pricing experiments
Tableau seats cut by a Claude Skill
15
Tableau seats cut by a Claude Skill

  1. 2026–now · Bay Area, CA

    NowBuilding in public

    AI Engineer: agents, RAG, evals and MCP

    Independent projects

    Building measured AI systems in the open: every project ships with an evaluation that says when it works and when it doesn't.

    • EDGAR RAG Analyst: cited agent over SEC filings; 97.5% on a held-out eval, hybrid search, a number-grounding guard and a CI eval gate
    • Experimentation Copilot: an MCP server and agent for A/B test readouts; 98% correct decisions vs 73% without tools
    • Alpha Research Engine: autonomous quant research with strict in-sample / out-of-sample separation
    If you can't measure it, you can't claim it. Evals come first.
    RAGMCPEvalsHybrid searchOpenTelemetryCI eval gates
    EDGAR RAG Analyst →Experimentation Copilot →Alpha Research Engine →

  2. 2022–2026 · San Francisco, CA

    WorkPricing, experiments and the first agents

    Data Scientist, Revenue Analytics (Pricing)

    Extend

    Owned pricing analytics: experiments, revenue simulation and enterprise deal sizing. Then built LLM agents that let non-technical teams query live revenue data in plain English.

    • Shipped chat-based LLM agents with function calling that let sales and other non-technical teams query live Snowflake revenue data in natural language
    • Built a Claude Skill that converts Tableau workbooks into Streamlit-in-Snowflake apps: 8 dashboards migrated, adopted by other teams, 15 Tableau seats cut
    • Built the A/B testing pipeline end to end (Segment → dbt → Snowflake, LaunchDarkly flags) and ran 10 pricing experiments, with 10%+ lift in normalized revenue per user in several
    • Replaced a unit-economics forecast with a Monte Carlo revenue simulator, packaged as a reusable Python engine for forecasting and deal sizing
    • Primary pricing data scientist on 40+ custom and enterprise deals over 4 years
    The highest-leverage AI work is grounding: curated data, metric logic and clear instructions matter more than the model.
    LLM agentsFunction callingClaude APIdbtA/B testingMonte CarloStreamlit

  3. 2021–2022 · San Francisco, CA

    WorkTeaching models to read catalogs

    Data Scientist, Machine Learning

    Extend

    At a fintech offering product and shipping protection, I built the NLP models that decide which products in a new merchant's catalog can be protected, and by which plan.

    • Warrantability classifier on product titles, descriptions and taxonomy using BERT embeddings (Hugging Face Transformers)
    • Warranty-mapping model with multi-class taxonomy prediction, paired with a rules layer into a hybrid ML + rules system for production
    • Solved early data scarcity by assembling a training corpus of millions of product records, with a Snowflake feature store
    In production, a model plus well-chosen rules beats a cleverer model alone.
    NLPBERTHugging Facescikit-learnSnowflakeFeature stores

  4. 2020–2021 · Boston, MA

    EducationLearning to measure

    MPS, Analytics (Statistical Modeling)

    Northeastern University

    Moved to Boston to formalize the statistics behind the work: modeling, inference and experimentation.

    Intuition finds the question; statistics decides whether the answer is real.
    Statistical modelingPythonR

  5. 2016–2019 · Bangalore, India

    WorkReading data for risk

    Sr. Risk Control Analyst / Transaction Risk Investigator

    Amazon Development Center

    Investigating suspicious transactions and merchants under KYC/AML rules, and feeding the patterns I found back to the ML team that ran auto-decisioning.

    • Monitored transaction data with key risk indicators, pattern recognition and anomaly detection
    • Escalated emerging risk patterns to the ML modeling team to improve auto-decisioning accuracy
    • Ran AML investigations and enhanced due diligence with SQL on Amazon Redshift
    • Built an Excel/VBA tracker on ETL pipelines that cut reporting effort by 3 hours a week
    • Trained and mentored 15 new analysts to production quality
    Models are only as good as the patterns people notice and feed back to them, which pulled me toward building the models myself.
    SQLRedshiftAnomaly detectionRisk analyticsVBA

  6. 2012–2016 · Bangalore, India

    EducationSignals and systems

    B.Tech, Electronics & Communication Engineering

    Visvesvaraya Technological University

    An engineering degree in signals, circuits and systems: the first place I learned to think about inputs, noise and outputs.

    Every system has noise; the job is to separate it from the signal.
    ElectronicsCommunication engineering

What's next

The next chapter is in the projects.

Each one ships with an evaluation that says when it works and when it doesn't. Have a look, or get in touch.

See projectsGet in touch