LN / PORTFOLIO / 2026 COLLEGE PARK, MD
000
Open to Data & AI Engineering roles College Park, MD —

Lakshminarayana Natarajan

Data & AI Engineer

— you can call me Lax

  • Data Engineering
  • Analytics & Warehousing
  • Data Analysis / ML
  • Applied AI
  • Agentic AI

I turn messy, late-arriving, real-world data into correct, auditable systems — and I ship AI products that reason, cite their evidence, and know when to act, ask, or escalate.

See the work lax.lakshminarayana@gmail.com
Scroll
0 yrs experience M.S. Information Systems · UMD AWS Data Engineer – Associate
01 About

I build the whole stack — warehouses and pipelines, the analytics on top of them, the models that score them, and the agentic AI products that act on them. Five years of production data work taught me the part that matters: a number is only worth what its audit trail can prove.

Experience 0 years SQL Developer → ALIP Application Lead at TCS, on the AIG Life & Retirement account.
Education M.S. Information Systems University of Maryland, Robert H. Smith School of Business — Dec 2025. B.Tech ECE, 2019.
Certification AWS Data Engineer Associate, 2026. S3, Glue, Athena, Redshift — plus Databricks and Snowflake exposure.
Based in College Park, MD Open to Data Engineering, Analytics Engineering and Applied / Agentic AI roles.
02 Products, not demos

Two I designed, built and shipped end-to-end. Both are being hardened into SaaS — real tenants, real production use.

P/01 Productizing → SaaS 2025 · OpenAI Build Week

ReplayOps

Operational Reasoning Compiler

  • Agentic AI
  • Applied AI
  • Product

Turn live work into memory your team can trust — capture someone doing a task, compile it into an evidence-linked SOP, operational memory and cited process Q&A, then safely replay it.

The problem

Operations experts rely on judgment that never reaches an SOP — exception handling, rule precedence, approval boundaries, and when not to guess. Happy-path automation is unacceptable in support, finance, claims, onboarding and back-office work.

  • Capture screen frames, live narration, and semantic browser actions with page context.
  • Compile with GPT-5.6 structured extraction — decisions, exceptions, follow-ups and retrieval chunks, each tied to a stable evidence ID.
  • Trust pass — a deterministic score for evidence coverage, grounding and replay-readiness.
  • Evidence-grounded Q&A that cites the original transcript, action or page — and abstains when the answer isn't in the session.
  • Safe DOM replay — confidence-and-margin gating, verifies real post-action change, pauses on protected or manual steps, re-approves on external change.
  • Enterprise MCP server (stdio) exposing allowlisted ServiceNow, SharePoint, Outlook, Power Automate and Confluence behind a fail-closed action-authorization gate.
Stack
  • Next.js
  • TypeScript monorepo
  • GPT-5.6
  • gpt-4o-transcribe
  • Zod Structured Outputs
  • IndexedDB
  • Microsoft Graph
  • MCP
replayops · session #4182 REC
Capture Compile Trust Replay
app.internal / vendor-intake VendorNorthwind Ltd CategoryLogistics Submit for review

“…and if the category is logistics it needs a second reviewer, so I never submit it straight through.”

EV-01 frame EV-02 narration EV-03 dom-action EV-04 page-context
SOP · compiled from session #4182 1.Open vendor-intakenav a[href="/vendor-intake"]EV-01 2.Fill vendor nameinput#vendor-nameEV-03 ◆Decision — category is logistics?EV-02 !Exception — add second reviewer before submitEV-02 3.Submit for reviewbutton#submit-reviewEV-04
0 Trust
Evidence coverage94% Grounding89% Replay-ready76%
app.internal / vendor-intake agent VendorNorthwind Ltd CategoryLogistics Submit for review
✓resolve input#vendor-name0.96 ✓type + verify DOM changeok ◆match 0.61 · margin 0.09below gate ‖protected controlPAUSED

Confidence-and-margin gated. It would rather ask than guess.

P/02 Productizing → SaaS 2025

R.Y.W

“The Resume You Want”

  • Agentic AI
  • Applied AI
  • Developer tooling

A desktop app that fuses your saved profile with a pasted job description and produces a tailored, one-page, ATS-clean resume — driven by an LLM and your own MCP build server.

How it works

Paste a JD → Claude tailors the content (Anthropic API, model swappable, default claude-opus-5) → the resume-tailor MCP server, spawned as a sidecar, builds the final one-page .docx/.pdf with structure, auto-fit, and no-fabrication rules enforced regardless of which model ran.

  • Deterministic build — layout and one-page fit live in the MCP server, not the prompt.
  • No fabrication — the builder only emits facts present in the stored profile.
  • Model-agnostic — swap the model without touching the document contract.
  • Local-first — Tauri's Rust core, your profile stays on your machine.
Stack
  • Tauri (Rust)
  • Node orchestrator
  • Anthropic API
  • resume-tailor MCP sidecar
R.Y.W · tailor claude-opus-5
Job description

We need a data engineer fluent in PySpark and Delta Lake to own medallion pipelines, enforce data quality, and partner with analytics on dimensional models.

MCP sidecar
resume-tailor
toolbuild_resume fitone_page fabricatefalse
Output
Lakshminarayana Natarajan Data Engineer 1 page
ATS clean One page No fabrication
03 Data Engineering

Pipelines that are provably correct, not just green.

E/01 · 2025–2026

Payments Settlement Lakehouse

  • Data Engineering
  • Data Quality
  • Auditability

A Delta Lake lakehouse that resolves late-arriving, out-of-order payment lifecycle events into correct, auditable merchant settlement figures — proven to the cent against ground truth.

A payment is a lifecycle: authorized → captured → settled → refunded → chargeback. Events arrive late and out of order — a chargeback can land 60 days later. Naive append-only pipelines double-count, overpay, and let bad records corrupt totals. That is real financial loss plus an audit failure.

Naive append-only $0.04 settlement error vs ground truth
vs
This lakehouse $0.00 settlement error vs ground truth

60,000 transaction lifecycles · 190,446 raw events. Naive would pay merchants $14.45M against a true $12.47M owed — off by roughly $2M.

MetricNaive append-onlyThis lakehouse
Settlement $ error vs ground truth$1,988,116.04$0.00
Duplicate / spurious settled rows5,8730
Late reversals correctly applied0 of 2,4672,467 / 2,467 · 100%
Bad records reaching settlement1970 · 1,200 quarantined
Bronze190,446append-only, all-string
Silver60,000idempotent MERGE, 1 row / txn 1,200 quarantined ↘
Gold3,666merchant × settle_date
OPTIMIZE + Z-ORDER · 1.6M rows 0 → 0 active files −99% file count
Single-merchant lookup
before1,409 ms after112 ms
12.6× faster

Z-ORDER clusters by merchant_id, so file skipping — not more compute — does the work. Same query, same cluster, 1/12th the wall clock.

Senior signals
  • Quarantine, never silent drops
  • Idempotent backfill
  • Point-in-time audit via Delta time travel
  • Run-metrics observability table
Stack
  • PySpark 3.5.3
  • Delta Lake 3.2.1
  • Medallion
  • Idempotent MERGE
  • Java 17
  • Local → MinIO → Databricks, same code
E/02 · 2025–2026

Supply Chain Late-Delivery Intelligence Platform

  • Data Engineering
  • ML
  • BI

End-to-end platform ingesting ~180K supply-chain order events a day, modelling them in a MySQL star schema, orchestrating with Airflow, scoring late-delivery risk with ML, and serving Tableau and Power BI.

SCD Type 220 YAML DQ rules Partitioned facts5 marts Auto-retraining ML DAG70% coverage gate
  • SCD Type 2 on customer and supplier dimensions.
  • 20 YAML-driven data-quality rules — critical failures block writes.
  • Random Forest late-delivery risk classifier with an auto-retraining DAG.
  • 40+ pytest tests, 70% coverage enforced; CI on every push (ruff + black + pytest + DAG validation).
  • Parameterized backfill DAG.
Stack
  • MySQL
  • Python
  • Airflow 2.9
  • scikit-learn
  • Docker Compose
  • Tableau
  • Power BI
  • GitHub Actions
View on GitHub ↗
E/03 · 2025

Banking Customer-Analytics Lakehouse

  • Data Engineering
  • Analytics Engineering

Production-style medallion pipeline turning daily raw CSV drops into curated banking customer-payment marts for Power BI. 

source: raw/payments/*.csv
target: silver.payments
transform: normalize_channel
dq:
  - not_null: [customer_id]
  - unique: [payment_id]
  - range: {amount: [0, 1e6]}
job · silver.payments read 274,118 dq 3 / 3 pass merge scd2 10,000 profiles write delta ✓
  • Config-driven PySpark framework — each job is a YAML spec: source, target, transform, DQ.
  • Runs identically on Databricks or locally; Delta or Parquet via env var, so it stays air-gap friendly.
  • Answers paid totals over 30/90/365 days, payment-gap and churn signals, channel trends (ACH/wire/card/check), and SCD2 profile history.
  • Synthetic data generator — 10k customers, ~274k payments — plus a pytest suite.
Stack
  • PySpark
  • Delta Lake
  • Medallion
  • SCD2
  • Synapse / Power BI
  • Databricks
04 Capstone
Deloitte-sponsored · University of Maryland · Feb – Dec 2025

Veteran Sentiment Analysis

  • Applied AI
  • Data Engineering
  • NLP
  • RAG
  • Dashboard

A Deloitte-sponsored graduate capstone measuring what veterans actually say — an end-to-end Python ETL and NLP pipeline that turns unstructured veteran feedback into themes, sentiment and a dashboard the people it describes can read.

  • End-to-end Python ETL producing 20,003 curated records across three datasets.
  • Normalized PostgreSQL model with documented schemas and data-quality tests.
  • Sentiment and theme extraction — spaCy, KeyBERT, UMAP, HDBSCAN, scikit-learn.
  • A final veteran sentiment dashboard — sentiment mix, theme ranking and trend over time, built for a non-technical audience.
  • RAG Q&A copilot on OpenAI + FastAPI returning grounded, source-cited answers.
Companion site ↗
veteran-sentiment · deloitte FastAPI
Ingest Sentiment Themes Dashboard Ask
survey responses8,412 support transcripts7,190 open feedback4,401
0curated veteran records
not_null ✓unique pk ✓ fk integrity ✓row counts ✓
Sentiment classification · scikit-learn
Positive Neutral Negative
spaCyKeyBERT lemmatisedPII stripped
benefits & claimshealthcare access transition & careersunclustered
Veteran Sentiment Dashboard FY 2025 · all regions
20,003responses 46%positive 3sources 12themes
Sentiment over time
Top themes Claims wait time Appointment access Career transition Housing support
?
chunk 04 · resp_2187chunk 11 · resp_0774 chunk 02 · resp_4310

Claims wait time1 and appointment access2 dominate, with career transition3 a distant third. Every claim is traced to an indexed response — nothing outside the corpus is asserted.

05 Experience

Tata Consultancy Services

Client: AIG (Life & Retirement) · Chennai, India · Jan 2020 – Aug 2024

Promoted SQL Developer → ALIP Application Lead. Oracle SQL/PL-SQL, SQL Server T-SQL, IBM DataStage.

  • Four and a half years on one insurance platform — long enough to own the run book, not just the tickets.
  • The vendor transition is the piece I'm proudest of: a support function moved between companies with zero disruption.
alip · nightly cycle Oracle PL/SQL
Batch Tune Handover Controls
policy billing transaction
5M+transactions processed daily
procedurespackages functionstriggers
Cycle time · 200+ users
before4h 00m after15m
execution plansindexing partitioning
Runtimes down 30–40%
Production support transition · outgoing vendor
01Knowledge transfer 02Shadow 03Reverse-shadow 04Steady state
Zero disruption · runbooks owned end-to-end
Reconciliation & incident control
✓null checkspass ✓primary-key uniquenesspass ✓duplicate scanpass ✓row countspass ✓balance reconciliationpass
RCA + triage in ServiceNow · manual effort down ~25%
Jan 2022 – Aug 2024

ALIP — Accenture Life Insurance Platform

Application Lead

  • Processed 5M+ daily transactions across policy, billing and transaction streams.
  • Advanced PL/SQL — procedures, packages, functions, triggers.
  • Owned the end-to-end transition of production support from an outgoing vendor — KT, shadow and reverse-shadow, runbooks, steady state, with zero disruption.
  • Automated a 4-hour cycle down to 15 minutes for 200+ users.
  • Tuned runtimes 30–40% via execution plans, indexing and partitioning.
  • SIT/UAT and release validation; mentored 3 developers; ran code reviews.
0daily transactions 4h → 15mcycle time 30–40%runtime reduction
Sep 2020 – Dec 2021

ACC — Monitoring & ETL Reporting

SQL Developer

  • Data quality, validation and reconciliation — null / PK / duplicate checks, row counts, balances.
  • Root-cause analysis and incident triage in ServiceNow.
  • Documented data flows and lineage; cut manual effort ~25%.
Jan 2020 – Sep 2020

NextGen eStart — Submission Platform

SQL Developer

  • Cross-stream SQL extracts and controlled DML deployments.
  • ~80% manual-effort reduction.
06 Toolkit
Languages & Query—

Five years of production SQL underneath everything else — the reason the pipelines above are auditable rather than merely green.

  • SQL — Oracle
  • SQL Server
  • PostgreSQL
  • MySQL
  • PL/SQL
  • T-SQL
  • Python — pandas, NumPy, scikit-learn
  • TypeScript
  • Rust (via Tauri)
  • Bash / Korn shell
07

Let's build something correct

Open to Data Engineering, Analytics Engineering and Applied / Agentic AI roles. Fastest way to reach me is email.

Phone+1 240 467 4056Call ↗ LinkedIn/in/lakshminarayananatarajanOpen ↗ GitHub/iamlax-byteOpen ↗ RésuméPDF · Data & AI EngineerOpen ↗