Shashank Nampalli

Forward Deployed Data and AI Engineer · DTCC

Shashank Nampalli

Production data and machine learning platforms, and forward-deployed delivery. At Walmart I owned a multi-tenant PySpark recommendation runtime (30+ TB/day, 20+ strategies). At Verizon I led the revenue recognition migration from Teradata to BigQuery (115M+ customers, close reporting 12h → 4h). Currently Forward Deployed Data and AI Engineer at DTCC (Cybersecurity Defensive Engineering).

Case studiesResume
Shashank Nampalli, Forward Deployed Data and AI Engineer
Scroll

About

Focus

Multi-tenant machine learning platform engineering, warehouse migrations on the finance-critical path, and forward-deployed AI and automation with security teams. Personal projects include agent systems with SQL guards, human-in-the-loop gates, and metadata contracts.

Case studies

Production systems

Multi-tenant recommendation platform at Walmart. Revenue recognition migration at Verizon. Personal projects on agent failure-mode design.

Walmart

Multi-tenant recommendation platform

Walmart International · PySpark / Dataproc / Airflow

Walmart International markets needed distinct personalization strategies (buy-also-bought, similar-item, reorder, trending, deals, and others) without cloning the Spark/Dataproc stack per market. Forked pipelines would have multiplied Airflow operations, SLA drift, and onboarding cost.

I owned the shared multi-tenant recommendation runtime: config-driven strategy plugins, tenant isolation, and common PySpark batch paths into GCS / Cosmos DB serving layers. New strategies shipped as configuration on the shared platform instead of net-new jobs.

Tradeoff: stricter schema and plugin contracts up front. Result: 30+ TB/day across 20+ strategies; ~60% lower strategy onboarding effort; ~35% compute reduction via incremental/delta processing vs full recomputes.

Primary owner of the shared recommendation runtime

PySparkDataprocBigQueryAirflowCosmos DBGCS

Verizon

Revenue recognition: Teradata → BigQuery

Verizon · close-path analytics

Revenue recognition and revenue-close analytics for 115M+ customers ran on Teradata with ~12h reporting latency and brittle ETL that blocked iteration on the close path.

I led the migration workstream to BigQuery: dual-run extract-load-transform pipelines, parity validation gates, then cutover. We accepted a longer parallel period in exchange for auditable row/aggregate checks before decommissioning Teradata jobs.

Post-cutover close reporting latency: 12h → 4h at 400–600M records/day (~3 TB/day). Same finance-critical path, cloud-native warehouse.

Tech lead for the migration workstream

BigQueryTeradataPySparkSQL

Also shipped

  • Semantic similarity & vector search · Walmart

    SentenceTransformers + CLIP embedding pipelines, Spark batch indexing, Vespa approximate nearest neighbor retrieval for semantic neighbors in production recommendations.

  • LLM complementary items · Walmart

    Gemini query generation → retrieval → engagement rerank, with prompt/config management, batching, retries/fallbacks, and output validation. Qualitative lift only; no click-through rate percentage claimed.

  • Reconciliation & audit · Verizon

    Automated cross-layer reconciliation on the revenue recognition path: ~90% fewer defects, manual validation under one hour, lineage-backed audit trail.

Demos

Personal projects, including SQL safety and human-in-the-loop remediation agents.

Personal demo

RideQuery

Personal project · natural language to SQL / ETL

GitHubLive demo

Prompt-to-SQL demos usually omit the failure modes that matter: unconstrained joins, hallucinated schema, and non-SELECT execution. RideQuery uses LangGraph to route intent to a SQL path or an ETL/Pandas path, then applies read-only guards before SQLite execution.

Runtime schema discovery is acceptable at demo scale. In production I'd front the agent with governed metadata (table and column definitions, grain, metric contracts), retrieve only approved context, and limit discovery to structural detail inside that allowlist. Definitions need an owner; otherwise you amplify bad SQL.

Try it. Live demo: ask a rideshare question. Non-SELECT SQL is blocked before SQLite.

LangGraphLangChainDeepSeekSQLitePandasStreamlit

Personal demo

SOC triage agent

Personal project · human-in-the-loop remediation gates

GitHubLive demo

Unattended remediation (host isolation, credential revoke) without an audit trail is an unacceptable failure mode. This LangGraph graph runs classify → investigate → remediate with Pydantic-typed outputs and playbook retrieval (SQLite / Qdrant).

LangGraph interrupt() enforces human-in-the-loop approval on escalation flags and high severity. The language model cannot bypass the gate. Each run writes an audit log for replay.

Try it. Live console: classify → investigate → human approval → remediate.

LangGraphLangChainPydanticDeepSeekQdrantLlamaIndexStreamlit

Tools

PySparkBigQueryAirflowGCPVespaGeminiLangGraphPythonSQL

Career Timeline

Recent roles

Contact

Get in touch

I'm at DTCC. If you want to talk about the work, book time or email.

EmailResumeLinkedIn

svarma.de@gmail.com