Yash Ravindra
Charde
Software engineer with 3+ years shipping production backend systems, currently pursuing an M.S. in Computer Science at Arizona State University (4.00 GPA). I've built fan-out event pipelines that carry security-critical business events, cut billing runs from a full day to hours, and moved 48,000+ clients onto a leaner delivery model. Lately I've been building distributed ML systems on AWS and co-authoring research on latent reasoning in LLMs.
Grounded in production, curious about research.
I spent the first three years of my career at ADP, working on RUN — payroll and HR software used by tens of thousands of small businesses. I moved from maintaining the messaging backbone that keeps security tools in sync with every business event, to owning client-facing features like billing consolidation and year-end compliance tracking.
That work taught me to care about the boring parts of engineering that matter most in production: retry logic, at-least-once delivery, and knowing exactly what happens when something fails at 2am.
I'm now doing an M.S. in Computer Science at Arizona State University, splitting my time between coursework in data systems and machine learning, and independent projects in distributed and federated systems on AWS. I'm also co-authoring a paper on latent reasoning in large language models.
Education
Three years at ADP, on the systems behind RUN payroll.
Senior Member Technical
Apr 2024 – Jun 2025- Built a consolidated invoices page in RUN with React.js, Node.js, and REST APIs — a responsive, mobile-friendly UI with consistent behavior across browsers and devices.
- Shipped an end-to-end Year-End Readiness Tracker (React.js + ASP.NET Web API) giving clients a single dashboard for W-2 progress, deadlines, and required actions.
- Optimized year-end billing and production workflows — cut billing processing time by 3 hours and brought W-2 production down to 6 hours; migrated 48,000+ clients to Local Delivery for $180K in annual savings.
Member Technical
Jul 2021 – Mar 2024- Architected an event-driven integration platform for GSO to receive, persist, and asynchronously fan out every RUN business event to downstream security consumers over REST, IBM MQ, and Kafka using shared integration contracts.
- Maintained the core processing components — MQReader (consumer & dispatcher) and BatchRetry (retry & reprocessing) — keeping delivery at-least-once and resilient across distributed messaging workflows.
- Built RED fraud alert email notifications for security-sensitive events, and led production support, incident triage, and root-cause analysis on high-priority event pipelines.
- Modernized the shipping details page, integrating delivery preferences into an MVC architecture to cut manual follow-ups and simplify the workflow.
Distributed systems, applied to ML.
Federated Learning System on AWS Cloud and Edge
- Architected an end-to-end federated learning pipeline on 10 distributed EC2 workers with an event-driven Lambda aggregator triggered via S3, training a LeNet-5 CNN on MNIST to 97.69% accuracy with consistent round-over-round convergence and no polling overhead.
- Extended the system to the edge with AWS IoT Greengrass, deploying custom MQTT-triggered components with scoped permissions for secure model uploads, sustaining 97.39% accuracy in the edge deployment.
Scalable Cloud-Native Face Recognition Pipeline on AWS
- Designed a multi-tier elastic pipeline — a stateless EC2 web tier, SQS-decoupled PyTorch inference workers, and S3 + DynamoDB for storage — sustaining sub-second average latency across 100 concurrent requests, scaling queues cleanly from 0 to peak and back.
- Wrote a custom autoscaling controller in Python that scales the application tier from 0 to 15 EC2 instances based on SQS queue depth, with DynamoDB result caching for a ~3x warm-cache speedup.
Currently under review.
Robust Latent Reasoning in Large Language Models
Co-authored with A. V. Malarkkan, M. R. Choudhury, U. Byahut, V. Gupta, and Dr. Yanjie Fu, under whose supervision the work was carried out.
Studies how large language models reason within their latent space, and how that reasoning can be made more robust.
Under review — NeurIPS 2026 & ICML 2026The stack I build with.
Languages
Frameworks & Libraries
Cloud & Tools
Practices & Platforms
A couple of nods along the way.
Team of the Quarter
For leading the Billing and Invoices initiative at ADP Private Limited.
"Win as One" Award
For contributions to the Year-End Health Tracker initiative at ADP Private Limited.