Hi, I'm Vishal Kumar Sarvade, a Technical Architect at Grid Dynamics with 12+ years in data engineering. Today I own a 1 TB Amazon Redshift data mart that about 63 operating companies report from, and I lead the ten engineers who keep it running.
I also:
Real work for a global direct-selling business and, earlier, a healthcare technology company. Client names stay private; the numbers are real.
A critical aggregation workload took about 12 hours. I rebuilt it on Redshift and reworked how it executes. It now finishes in 30 minutes.
Write jobs on Redshift Serverless slowed from about 10 seconds to 10 to 15 minutes. I traced it to base capacity being set equal to the maximum, and recommended RA3 for writes with a Serverless workgroup for reporting.
Consolidation saved about $5,000 a month. On a 3-year Redshift reservation, I found the cost model had never included the biggest lever, a move to Graviton-based RG nodes. I advised against signing either option on the table.
A four-cluster Redshift estate, one producer and three consumers, stayed in step through sync jobs. I moved it from DC2 to RA3 and set up data sharing. The sync jobs are gone and consumers read live data.
Replication from Aurora MySQL ran on AWS DMS, with maintenance pauses, full reloads after failures and binlog workarounds. I replaced it with Zero-ETL, which ended all three. The same pipeline now carries the move off AS400, routed country by country.
I built the capacity and pricing plan for moving 110 Informatica Cloud jobs to ETL Works, presented the options to the client and won the project. The move runs from September 2026 to March 2027, in two phases and one job at a time, so nothing breaks on the way over.
A static, S3-hosted metadata portal built from the business's data dictionary, with quality gates in the build. A Python tool on the Looker SDK that traces every dashboard to the column. A job metadata framework that tracks every pipeline across Informatica, ETL Works and Zero-ETL.
I moved government IDs, bank account details, retail and customer service records from CSV files and MySQL into Aurora MySQL. Every batch was checked for accuracy, and rejected rows went back to the customer as CSV.
For a healthcare technology company, I led the move of member, provider, claims, authorization and eligibility records off legacy systems with SQL Server stored procedures and SSIS, validating every business rule and running the gap analysis with the client.
I architected this data mart and lead the team that runs it. Every dashboard and report the business produces reads from it. Simplified, with client names removed.
Standards I setNaming, keys, required audit columns, distribution style, documentation and access grants. Every database change traces to a Jira ticket and a Git commit.
Who I work withThe client's CTO, business analysts, and the application, production support and legacy teams I coordinate releases with.
How I run the teamSprint planning, Scrum calls and architecture reviews with a ten-person offshore team, turning business requirements into delivered Jira tickets.
I don't start with the model. I start with the question people keep asking, the business rule behind it and the data that answers it. Once I understand the root cause, I automate the part that repeats.
I built an LLM-based NL-to-SQL agent over the data mart and its Looker tables. It can only read. It blocks PII, forces the filters every query needs, and pins down what "last quarter" means instead of guessing.
The chat front ends add a trust layer: a confidence score, a plain account of how the question was read, and the SQL it ran, right there on screen.
SELECT country, COUNT(*) AS orders FROM sales.orders WHERE status = 'COMPLETE' AND order_date BETWEEN '2026-07-01' AND '2026-09-30' GROUP BY country ORDER BY orders DESC;
One answers the team's recurring FAQs; the other does first-pass data analysis.
BuiltTook report requirements, built the data mart tables and semantic models, generated Power BI and HTML reports, validated them against the database and presented them to the customer.
DeliveredPipeline health in one view. Built with Claude Code.
BuiltPuts the business data dictionary in front of the team in one searchable place. Built with Codex.
BuiltStructures delivery estimates, validation effort, risks and deployment steps. Built with Codex.
BuiltSix dashboards, three of them built with Copilot, plus Microsoft Fabric set up. Presented to CTO and VP audiences.
Proven outFabric data agents so Microsoft 365 Copilot answers from Power BI semantic models.
In progressA short daily note on how companies are putting AI agents to work, plus longer posts on Redshift and data architecture. Every source is linked so you can check the numbers.
I've worked on the same global data platform since 2018, first as an engineer building its pipelines and now as the architect who owns it.
I own the data architecture for a global direct-selling business: the Redshift data mart, its ingestion, its standards and its AI layer. I lead a ten-person offshore team and work directly with the client's CTO and business teams.
Turned business requirements into solution designs and delivery plans. Reviewed the team's work, cleared blockers and coordinated production fixes. Worked with application teams on design and release support so data integration kept pace with platform changes, and with client teams on options and trade-offs, in plain terms.
Led a small team that built end-to-end pipelines from PostgreSQL and AS400 into Amazon Redshift with Informatica Cloud and ETL Works.
Grew from Associate Software Engineer to Associate Technical Lead. Led a healthcare data migration from 2014 to 2018, then joined the e-commerce data platform I still work on today.
The architecture comes with the TCO comparison, the FTE math and the cost model.
If I can't trace a figure back to its source, it stays out of the deck.
Mismatched reports, failed pipelines and slow queries get investigated before anyone proposes a fix.
My architecture diagrams are made so an executive can follow the lineage.
A structured comparison beats an endless vendor evaluation.
I won the ETL Works migration by presenting the options to the client directly.
A short, sourced note each day on how companies are putting AI agents into production, plus longer posts on Redshift and data architecture.
I started in 2014 in Hyderabad, writing SQL Server stored procedures for a healthcare migration. Members, providers, claims, eligibility. It taught me early that a migration is really a long list of business rules, and every one of them has to be proven.
In 2018 I moved onto an e-commerce data platform for a global direct-selling company, and I'm still on it. I've worked on it across two employers, first as an engineer building pipelines out of AS400 and PostgreSQL, and today as the architect who owns it. It's a 1 TB Redshift data mart with more than 200 tables, and about 63 operating companies report from it.
A lot of my best work is plain engineering. A job that took 12 hours now takes 30 minutes. Four clusters share data instead of syncing it. A replication pipeline no longer needs babysitting. A 3-year reservation didn't get signed because the cost model left out the option that mattered most.
More recently I've been putting AI to work on the same platform: an NL-to-SQL agent that shows its SQL, agents that handle the questions my team gets every week, and Copilot running on governed semantic models.
I lead a team of ten. I care about standards, root causes and numbers you can trace back to their source.
The fastest way to reach me is email or LinkedIn.