DATA POINTS
8 research studies since 2019. 17 years in software quality and data engineering. Focus: validation, migration and security of financial data in the cloud
The first half of 2023 belongs to generative AI. Chat assistants have reached hundreds of millions of people, and nearly every large company now has an AI strategy, a task force or a pilot. Far fewer have a clear answer to a simpler question. Is the data that will feed those systems correct?
Sasank Aramadaka has been working on that question for most of his career. A senior architect specializing in data quality, testing and cloud migration in financial services, he has published eight research studies since 2019, the latest this month. None is about neural networks. All of them are about the conditions AI depends on: data that is complete, consistent, secure and verifiable.
His work offers a useful corrective to the current mood. Before organizations can trust what AI tells them, they have to trust what they feed it. Aramadaka’s research suggests how.
-
AI inherits every error upstream
Machine learning models learn from history. If a loan record was truncated in a migration, if a trade was dropped between tiers of a warehouse, or if a reference value was out of date when it was vended to downstream systems, the model learns that mistake and repeats it at scale. Generative systems that summarize or answer questions over enterprise data are no different. They can only be as accurate as the sources behind them.
Aramadaka has seen this first-hand. Earlier in his career he led quality engineering on an enterprise data lake in mortgage finance, a platform that fed Hive, Spark and HBase for analytics, Tableau for reporting and machine learning models for predictive analytics. The experience shaped a central idea in his research: assurance must be built into the pipeline, not bolted on at the end.
-
Proving data is right, at every layer
Much of his research deals with the mechanics of verification. One strand looks at the boundaries of systems, holding gateway services to API contracts and validating microservice architectures end to end in regulated environments. Another tackles the hard problem of comparing data across many different database engines with a single, source-agnostic framework.
He has also studied how to verify cloud warehouses tier by tier, from landing through staging to base tables, and how to couple batch schedulers with automated data comparison so every overnight run is checked automatically. A further study examined the accuracy and timeliness of financial instrument reference data, the shared facts that trading, risk and accounting systems all rely on.
Each of these is a building block for AI readiness. A model trained on a warehouse that has been verified at every layer starts from a far stronger position than one trained on data nobody has checked.
THE ARGUMENT
Data assurance is not overhead on the road to AI. It is the road.
-
Migration is where data is most at risk
Aramadaka’s two 2023 studies focus on the moment of greatest danger: moving data from old systems to new ones. Many financial firms are migrating in-house databases to managed cloud services and moving large warehouses from on-premises platforms to AWS. These moves are often justified as preparation for analytics and AI.
His March study addressed cutover planning, setting out how to validate a migration before, during and after the switch, and how to plan a clean rollback if something goes wrong. The study published this month turns to security, examining how capital markets data can be protected both in transit and at rest during an on-premises-to-AWS warehouse move.
Together they make a point many AI roadmaps overlook. A migration that loses records, or exposes sensitive data along the way, does not prepare an organization for AI. It sets it back.
-
Research drawn from the work
The studies track Aramadaka’s day-to-day work closely. As a senior architect he works across Amazon S3, Redshift, EMR, Lambda, Glue, Step Functions and CloudWatch, supports migrations from Netezza to the cloud and from in-house databases to Amazon RDS, and has implemented a data testing framework that compares Oracle, Netezza, Redshift, Postgres and S3. He runs Ab Initio workflows and data comparisons through Autosys, validates source-to-target loads in SQL, and integrates automation into CI/CD pipelines with Git and Jenkins.
His earlier roles in quality leadership built deep expertise in Selenium, Java, TestNG, JMeter, Postman, SoapUI, Prism, Kibana and Splunk, along with ETL testing on Informatica and Ab Initio. He is AWS Developer Associate, IBM, ISTQB and ITIL certified.
-
What the field should take from it
The lesson of Aramadaka’s eight studies is that trustworthy AI begins long before a model is chosen. It begins with contracts at system boundaries, verification at every layer of the warehouse, automated checks on every scheduled run, careful planning for every migration and security for data wherever it moves.
That message is especially important in regulated industries such as banking and mortgage finance, where an AI system that relies on bad data is not only inaccurate but potentially non-compliant. As regulators in the United States and Europe sharpen their focus on how AI systems are built and governed, the ability to prove where data came from and that it is correct will become a competitive advantage.
In a year dominated by talk of models, Sasank Aramadaka’s research is a reminder that the most important work in AI may be the work that happens before the model ever sees a single record.
