Each month, over 42 million Americans depend on SNAP benefits to feed their families. The caseworkers who process these applications face a different challenge: outdated computer systems that can’t keep up with the volume of cases. Manual reviews pile up, and families often wait weeks for decisions while staff struggle through backlogs.
The numbers tell the story. Payment errors in SNAP programs cost billions annually, with states reviewing tens of thousands of cases each year to identify where benefits were calculated incorrectly. Legacy data systems can’t talk to each other. Income verification requires pulling information from multiple agencies. Cross-referencing household composition, work requirements, and financial records happens through manual processes that introduce mistakes. States processing hundreds of thousands of cases each month carry massive technical debt.
Missouri’s MEDES SNAP program ran into exactly this problem. The state needed infrastructure that could verify eligibility in real time, catch fraud before benefits went out, and produce audit reports for federal reviewers without staff having to build spreadsheets by hand. Fixing it required more than new software. The entire data pipeline needed to be rebuilt.
Building a three-tier data backbone
Srinubabu Kilaru joined the modernisation project as Senior Data Lead. His assignment was to design an architecture that could replace batch processing systems running on decades-old code. He built a three-tier framework: a Staging layer for raw data ingestion, a Delta Lake for version-controlled transformations, and a Data Warehouse layer optimised for reporting. The technical stack included Azure Data Factory for orchestration, Databricks with PySpark for distributed processing, Informatica for connecting legacy systems, and DBT for transformation logic.
“The old system was a black box,” Srinubabu Kilaru explains. “Data moved through undocumented scripts, and by the time analysts spotted an anomaly, the benefit had already gone out. We needed visibility at every step, not just at the end.”
The staging layer pulled data from case management systems, income verification APIs, and third-party employment databases without altering the raw records. Delta Lake applied versioned transformations while preserving lineage for audits and enabling rollback if problems appeared. The warehouse layer aggregated everything into semantic models ready for analytics and compliance reporting. This separation meant data engineers could refine transformations without breaking downstream dashboards.
He embedded Python-based anomaly detection models directly into the ETL pipelines. The models flagged unusual patterns: duplicate applications from the same address, income discrepancies between reported wages and employer records, and sudden changes in household size. These checks ran before the data reached the warehouse. Benefit issuance errors dropped by 35 per cent, and the volume of cases needing manual caseworker review fell sharply. Fraud patterns that used to take weeks to surface now trigger alerts within hours.
Dashboards that accelerate compliance
SNAP’s federal reporting requirements are strict. States track able-bodied adults without dependents (ABAWD) separately, monitor pending verifications, and submit demographic breakdowns to the Food and Nutrition Service on tight deadlines. Missouri’s old system made analysts pull data manually, reconcile discrepancies across multiple spreadsheets, and generate reports that often missed deadlines.
He built Power BI dashboards that displayed these metrics in near real time. One dashboard tracked ABAWD recipients approaching work-hour thresholds who needed outreach. Another showed pending verifications by the county, helping supervisors allocate caseworker resources where backlogs formed. A third provided demographic breakdowns (age, household size, disability status) updated nightly from the warehouse layer.
Report delivery times fell by 70 per cent. Program managers could answer federal inquiries within hours instead of days. Caseworkers accessed self-service analytics to check case status without waiting for IT support. The dashboards also exposed inefficiencies that had been invisible. One county consistently showed higher pending verification rates, which prompted a workflow review that uncovered a staffing gap.
Governance and continuous deployment
Making the system scalable required more than pipelines and dashboards. He implemented the Unity Catalogue to centralise data governance, establishing role-based access controls so only authorised users could view personally identifiable information. Audit logs tracked every query and transformation, creating a trail that satisfied both state auditors and federal reviewers.
CI/CD workflows using Git and Azure DevOps automated deployment. Data engineers committed transformation logic to version control, which triggered automated tests before changes reached production. This cut deployment cycles by 50 per cent and reduced the risk of untested code breaking live pipelines. Data refresh cycles improved by 60 per cent, and overall system availability increased by more than 4 per cent as maintenance downtime became rare.
“Audit readiness was non-negotiable,” Srinubabu Kilaru notes. “Every transformation had to be traceable, every access logged. That discipline also made troubleshooting faster because we could pinpoint exactly where a data issue originated.”
The modernisation provided a foundation for future AI initiatives. The clean, versioned data in Delta Lake can now feed machine learning models for predictive case management, forecasting which households might need additional support before a crisis or identifying eligibility changes that could be processed automatically. The semantic models make it easier to prototype new analytics without starting from scratch each time.
A template for other states
Missouri’s project reflects a broader shift in public benefits administration. Research from Code for America shows states are investing in data modernisation to reduce errors, speed processing, and improve outcomes. The challenge is consistent across jurisdictions: ageing systems, fragmented data, and compliance requirements that eat up resources better spent on service delivery.
The techniques Srinubabu Kilaru applies, including layered architecture, embedded anomaly detection, and centralised governance, are not proprietary. They draw from cloud-native practices that have worked in finance and healthcare. What makes Missouri’s project notable is its execution under real constraints: tight budgets, legacy integrations, and high stakes where benefit decisions affect families immediately.
The results demonstrate what becomes possible when states treat data infrastructure as a priority. Faster processing means families receive benefits when they need them. Fewer errors translate to fewer improper payments and reduced fraud. Systems built for audits let staff focus on casework instead of paperwork. A scalable foundation prepares programs to adopt AI and automation as those technologies mature.
Other states facing the limits of outdated systems can look to Missouri’s MEDES modernisation as a practical example. The work involves complex technical challenges, but the payoff is direct: shorter delays, fewer mistakes, and a benefits system that actually serves the people who depend on it.
