Gautham
Manuru Prabhu

Four agents run in production today: service-request resolution, knowledge-base indexing, case follow-up, and proactive alerting. Between them they close 35% of incoming cases with no human in the loop, took 40% off mean time to resolution, and return roughly 10,000 engineer-hours a quarter to the supply chain organisation.

Under the Quality Transformation Program I built failure-analysis and predictive-maintenance agents for the manufacturing side. They root-caused 70% of failures that previously went undiagnosed, improved upstream defect detection and brought down repeat failure rates across global sites.

I also found coverage gaps in Paladin edge connectors across 10 production lines, and wired SplunkAI observability into legacy systems that had none. I now own the research-to-production path for these agents end to end: retrieval quality, evaluation, guardrails.

Promoted from SE1 to SE2 in 12 months, the second promotion on the same team inside 18.

The year was a rewrite. Legacy failure-analysis workflows became event-driven microservices on Kubernetes, built with FastAPI, Cassandra and Redis. I added anomaly detection over manufacturing telemetry, and SLA dashboards in Prometheus and Grafana so that an on-call engineer saw a breach before the customer did.

I converted from intern to SE1 in six months. I built microservices with circuit breakers and retry logic, and a release pipeline automated with Jenkins, Docker and Kubernetes that took deployments from six hours to two minutes.

For Cisco's intern case study competition I built an NLP pipeline that standardised supplier names across 50,000 records, and it placed third of more than 100 entries.

The problem is adaptation: models are trained once on a snapshot, and real graphs keep moving. I work on training Temporal Graph Neural Networks over dynamic graphs with millions of edges, under a fixed memory and compute budget, using PyTorch Geometric, DGL and CUDA across several GPUs.

Optical coherence tomography scans are noisy and the classes are badly unbalanced. I proposed an ensemble feature-fusion approach for anomaly detection, tuning preprocessing, augmentation and colour-space features against a multi-class benchmark, and contributed to a set of 3,000 annotated retinal scans.

We tested whether quantum models could beat classical baselines at detecting cardiovascular disease from ECG images. QSVC, Pegasos QSVC and a quanvolutional neural network reached 97% accuracy, 10 to 14 points above the classical baselines we trained ourselves. The work was funded by a MeitY and AWS grant, and took us to the top 16 of 1,600 teams at the QETCI hackathon.

Retinal vessel segmentation across 3,000 fundus images, combining CLAHE, Gaussian blur and edge detection. Because the images left Kasturba Medical College's network, we also had to encrypt them properly, so the pipeline carries Diffie–Hellman key exchange and AES-256. Published in Springer LNCS and presented at ATIS.

We modelled how vaccine misinformation travels, running BERT and XLNet over a corpus of 10,000 posts alongside a graph model of who repeats whom. Classification passed 90% F1 on the benchmark. The user-association signal did more work than the language model did.