Data Scientist / ML Engineer— New York[ available full-time · january 2027 ]

Arjun Varma

I build forecasting, decision-support, and agentic systems — from data pipelines and model evaluation through deployed products.

Data science intern at Novo Nordisk, working on propensity scoring, next-best engagement, and LLM decision support. M.S. in Data Science at Columbia. Before that, three years at ZS building production ML and healthcare analytics.

EmailGitHubLinkedInResume
Arjun Varma
fig. 00 — the authorNYC

01/ SELECTED PROJECTS

Four projects, each with an interactive figure built from its real evaluation and architecture.

  1. 01selected project

    DEUCE Tennis Forecast

    A live ATP and WTA forecasting platform that combines surface-aware Elo, opponent-adjusted serve and return modeling, calibrated XGBoost, and tournament simulation.

    87,957
    matches in walk-forward evaluation
    0.1950
    ATP Brier score
    0.2017
    WTA Brier score

    Python · XGBoost · Elo · Markov models · Monte Carlo · Next.js · GitHub Actions · Firebase

    Live website GitHub Case study

    Maintained product · data refreshes hourly · model retrains daily

    forecast signals01 / 04interactive

    forecast inputs

    Four families of match evidence feed one calibrated forecast. Select a signal to see what it adds.

    walk-forward
    01
    selected signal
    Surface Elo

    A separate strength signal for each court surface keeps hard, clay, and grass performance from collapsing into one rating.

    modelXGBoost
    calibratedwin probability

    simulated: point → game → set → match → draw

    fig. 01 — Inspect how four match-evidence families enter a calibrated forecast and tournament simulation.
  2. 02selected project

    Airbnb Data Analyst Agent

    A reusable multi-agent analytics copilot that plans, writes, validates, charts, and narrates SQL analysis — every number cited back to source rows.

    5
    agents on a typed message bus
    ×3
    retry budget, exponential backoff
    100%
    numbers cited to source rows

    FastAPI · LangChain · DuckDB / Postgres / Snowflake · OpenAI function calling · matplotlib · pytest

    full trace →
    recorded agent trace02 / 04recorded run
    recorded failure → recoverystep 1/7
    planner0.82s

    Compare avg review_scores_rating for superhosts vs non-superhosts; check statistical significance before claiming a difference.

    fig. 02 — Replay a real failed query, database error, validator intervention, retry, and cited answer.
  3. 03selected project

    ClaimReady

    Planner-led workflow with four specialists that turns contracts, invoices, emails, and screenshots into structured NYC small-claims packets with multimodal extraction, legal retrieval, and typed handoffs.

    4
    specialist agents
    6
    legal-corpus documents
    4
    packet documents

    OpenAI Agents SDK · FastAPI · Next.js · GCP

    specialist pipeline03 / 04interactive

    specialist pipeline

    Four specialists build one shared case record. Select a stage to see what it adds.

    CaseFacts
    stage 01adds to CaseFacts

    Extractor

    Reads uploaded text and images, then converts the evidence into validated CaseFacts rather than passing free-form reasoning downstream.

    evidence / tool

    contracts · invoices · emails · screenshots

    fields added
    contractperformancebreachexhibits
    statement
    demand letter
    exhibit index
    filing guide

    NYC · unpaid-services contract claims · document generation only

    fig. 03 — Inspect four specialist stages around one shared CaseFacts record before packet assembly.
  4. 04selected project

    Biliary Tract Cancer Early Detection

    A production model that flags likely BTC patients ~45 days before claims data confirms them — scoring 250M patient-claims every month.

    ~45d
    earlier identification
    250M
    patient-claims scored / month
    PMSA '25
    methodology presented

    PySpark · XGBoost · SHAP · K-means / GMM · NLP clustering · MLflow

    Case study

    Private production system · public website and source are unavailable

    evaluation window04 / 04interactive

    evaluation window

    Drag the index date. The model sees history only up to 45 days before it, then is scored on the next 30.

    leakage safe
    available history
    45 days masked
    next 30 days
    index date
    01Trainonly information available live
    02Maskremove near-diagnosis leakage
    03Predictscore the following 30 days
    fig. 04 — Move the evaluation index while preserving the 45-day mask and following 30-day target.

02/ MORE SHIPPED WORK

  • Filing Intelligence RAG

    2026Retrieval

    Evidence-first research across 4,967 indexed passages from 128 filings, earnings decks, and call transcripts spanning 15 companies. Deterministic retrieval, bounded context, and page/line citations open directly in the source PDF; 72 automated checks cover retrieval, authentication, API contracts, and production builds.

    Python · FastAPI · ChromaDB · Next.js · GCP Cloud Run

    Live website GitHub Case study

  • GAFFER: Live World Cup Forecasting Platform

    2026Forecasting

    Forecasts the 2026 World Cup from ~49K international matches using Elo, a time-weighted Dixon-Coles goal model, squad-value blending, and 50,000 Monte Carlo runs. Walk-forward evaluation across 8,136 matches since 2018 reached 0.887 log loss versus a 1.05 baseline; frozen pre-match predictions are graded against Kalshi as the hourly pipeline refreshes.

    Python · pandas · NumPy · Next.js · GitHub Actions · Firebase

    Live website GitHub

  • ClassPulse

    2026LLM products

    Live classroom theme extraction — students answer via QR, an LLM clusters responses into themed cards every 10s over SSE, with a 5-model fallback chain.

    FastAPI · React · SSE · OpenRouter

    Live website GitHub

  • SeanceAI

    2025Agents

    Conversations with 60+ historical figures under era-appropriate knowledge boundaries; Dinner-Party mode runs 2–5 figure multi-agent dialogue.

    Flask · OpenRouter · SSE

    Live website GitHub

  • Citation Format Checker

    2026LLM products

    Narrow-scope chatbot that flags APA 7 / MLA 9 / Chicago 17 violations with rule-IDs and quoted evidence; three-method eval suite, 30+ test cases.

    Vertex AI · FastAPI · Cloud Run

    Live website GitHub

03/ EXPERIENCE

Novo Nordisk

Plainsboro, NJ · Jun 2026 — present

current

Data Science Intern — Commercial Data Science

Jun 2026 — present
  • Improved propensity scoring and next-best-engagement models, then built an LLM decision-support layer that turns behavioral, access, competitive, and adoption signals into explainable field guidance.
  • Integrated longitudinal claims with unmet-need, market-opportunity, and early-adoption signals to prioritize ~500 of 10,000+ providers nationwide.
  • Achieved 5.2x top-decile lift with an XGBoost propensity model using provider profile, prescribing, access, and engagement features; combined predictions with rule-based signals into explainable priority tiers and a nationwide planning dashboard.
  • Supporting pre-launch analytics for investigational denecimig (Mim8), informing HCP targeting and field readiness.

ZS Associates

Pune · Feb 2022 — Jun 2025

Advanced Data Science Associate Consultant

Feb 2025 — Jun 2025
  • Shipped the biliary tract cancer early-detection model — 250M patient-claims scored monthly, in production.
  • Reduced weekly decision cycles from days to minutes for 100+ stakeholders across a $10B portfolio by translating 5+ high-volume sources into a reusable Spark/SQL data warehouse, product KPIs, and self-service analyses.
  • Productionized models with MLflow tracking, feature and prediction drift monitoring, and CI validation to reduce silent failures.

Decision Analytics Associate Consultant

Jul 2024 — Jan 2025
  • Built and deployed positive-unlabeled learning models that recovered missing categorical labels in sparse medical transaction data, expanding analytical coverage from ~40% to 95% across segments and territories.
  • Raised first-pass analytical quality above 99% and saved ~50 hours per month by leading a five-member modernization of legacy business rules into validated, reusable logic.

Decision Analytics Associate

Feb 2022 — Jun 2024
  • Established audit-ready real-world evidence modeling and segmentation across $4B+ in products by designing cohort and entity features over millions of incomplete and miscoded records.
  • Engineered PySpark/SQL ETL across healthcare sources and defined patient inclusion and exclusion logic robust to missing and miscoded fields.

Promoted to Associate Consultant in 4 cycles (typical: 5). Expert Associate and Insight Illuminator awards.

Columbia University

New York · 2025 — present

Graduate Teaching Assistant

2025 — present
  • TA for Business Analytics II (Foundations of AI) and Hollywood & Big Data at Columbia Business School.

04/ ABOUT

I work on the part of machine learning that starts after the demo.

My work spans forecasting, experimentation, healthcare analytics, and agent workflows. Across them, the recurring problems are the same: reliable data, honest evaluation, useful interfaces, and enough monitoring to know when a system stops working.

Education

Columbia University

M.S. Data Science

Aug 2025 — Dec 2026

TA — Business Analytics II (Foundations of AI) · Hollywood & Big Data, Columbia Business School

Vellore Institute of Technology

B.Tech, Electronics & Communication Engineering

Jul 2018 — May 2022

Special Achiever Award · Merit Scholarship

Toolbox

Programming & ML · Modeling & Experimentation · Evaluation & Monitoring · Applied AI · Data & MLOps · Databases & Cloud

PythonSQLRpandasNumPyscikit-learnPyTorchXGBoostSHAPPropensity scoringA/B testingExperiment designHypothesis testingCausal inferenceRanking / recommendationProbabilistic forecastingMonte Carlo simulationWalk-forward validationBrier scoreMulticlass log lossRegression suitesFeature driftPrediction driftData quality validationLLMs / RAG
LangChainAgentic workflowsMulti-agent orchestrationLLM evaluationGuardrailsFastAPIREST APIsSpark / PySparkDatabricksMLflowProduction MLETL / ELTData warehousingDockerCI / CDGitHub ActionsPostgresSnowflakeDuckDBChromaDBAWS S3AWS EMRAWS AthenaAWS SageMakerGCP Cloud Run