Skip to content
Sai Eswar Gujjala

Now: Building agentic AI workflows for clinical data standards at ZS.

> history --work

Experience

My roles, newest first. My path from school to today is on the Journey page.

  1. Jan 2026 – present

    Data Scientist

    ZS · Bengaluru, India

    Builds agentic AI tools for clinical data standards (CDISC SDTM and ADaM), with human review and full audit trails in a regulated environment.

    Responsibilities

    • Own end to end the AI-assisted SDTM mapping workflow in MINT+: validation and generation agents on AWS Bedrock AgentCore, a FastAPI backend, and mapping documents with audit history in DocumentDB.
    • Took it to production on Amazon EKS (Kubernetes, Helm), with GitLab CI/CD and governed releases across Dev, Test, BVE and Production, plus a LangSmith evaluation pipeline for controlled prompt experiments.
    • Designed and built OAK, a retrieve-then-verify error-resolution agent. Known errors are matched exactly against a rulebook first, an LLM verifies candidates only when needed, and fixes come only from the curated knowledge base.
    • Built an agentic double-programming workflow that independently re-creates the analysis results behind TLG outputs in R and reconciles them with production, inside guardrails.

    Outcomes

    • Data programmers start from AI suggestions instead of writing every SDTM mapping and annotation from scratch, and mappings are more consistent across studies.
    • Known ROAK runtime errors are resolved immediately by deterministic matching, which reduces escalations to the core ROAK team.
    • Built for scale: studies with up to 5,000 mappings, and up to 50 error logs per OAK request on 10 parallel workers.

    Related projects

  2. Jul 2024 – Jan 2026

    Data Scientist

    Zetwerk · Bengaluru, India

    End-to-end owner of AI and ML systems for manufacturing and procurement: multi-agent platforms, a supply-chain knowledge graph, computer vision and predictive models.

    Responsibilities

    • Designed, built and deployed ForgeOps, an event-driven multi-agent platform for manufacturing exceptions.
    • Built a supply-chain knowledge graph from 5,000+ documents with a PydanticAI multi-agent pipeline, powering supplier matching, risk detection and cross-business-unit discovery.
    • Built a procurement copilot that turns requests into vendor recommendation reports.
    • Built a multi-modal pipeline for supplier audit recordings.
    • Built a regression pipeline that turns raw battery sensor streams into 200+ degradation indicators, so the EV division can estimate warranty risk 12 months earlier.

    Outcomes

    • Exceptions were spotted in 1.8 hours instead of 11.4.
    • On-time delivery rose from 86.3% to 90.5%.
    • Premium freight and expediting cost fell by 11%.
    • Supplier matching time was cut by 85%, from about 2 days to under 4 hours.
    • Procurement costs fell by about 15% across 3 business units, by consolidating volume with 15 shared suppliers.
    • The procurement cycle took under 30 minutes instead of about 3 days.
    • The EV battery models saved ₹1.4 crore from projected replacement budgets.

    Related projects

  3. Jun 2023 – Jul 2023

    Data Science and Business Analyst Intern

    The Sparks Foundation · Remote

    Analytics internship: exploratory analysis and ML models on retail, finance and IoT data.

    Responsibilities

    • Framed business questions for multi-domain datasets.
    • Engineered features.
    • Built regression and classification models in Python (pandas, scikit-learn).
    • Presented dashboards and A/B insights.

    Outcomes

    • The dashboards and A/B insights were adopted as examples in the foundation's learning repository.