Curriculum Vitae

• Trained a ConvLSTM encoder-decoder in PyTorch on WRF weather-model output fused with ground-sensor records of 163,000+ lightning strikes, comparing three architectures
• Built a React and Leaflet dashboard with state-level filtering
• Submitted as a research prototype to ISRO's internal research library

• One of a team of five on daily financial time-series reporting; I built ETL stages and the Power BI and Tableau dashboards
• Rebuilt the reporting twice after asking the five non-technical colleagues who used it what they actually read
• The team's forecasting pipeline reached 84% directional accuracy

• Shipped four Java web applications (JSP, Tomcat) to paying clients, taking requirements from them directly
• Support contact after handover: agreed fixes and timelines with clients and closed the loop

• Two merged contributions: OpenTelemetry metrics middleware in Go, and table-driven unit tests for the critical-path helper
• Through the project's GitHub Actions CI and maintainer review

• Parallel programming on ARCHER2 (UK National Supercomputer) using MPI, OpenMP, and CUDA
• Dissertation: early detection of memory leaks in containerised environments, supervised by Ben Carpenter. Submitted August 2026, results pending

• First Class with Distinction, CGPA 8.57/10

Streaming Jailbreak Detection

Change-point detectors over first-token confidence, entropy and Gini read from the model's own logits. After auditing my first run (mislabelled benign set, stream-length leak, in-sample threshold): 0.74 AUROC (95% CI 0.61–0.84), about 3% false alarms on held-out benign conversations, 0.3% overhead.

MSc Dissertation: Memory Leak Detection

eBPF and container-stats collectors into TimescaleDB across four cloud providers; 703,000 records over 755 runs; ten detectors compared at one 5% false-alarm budget, where six lost most of their recall; gradient-boosted primary detector at 98.6% recall and 4.9% false alarms; CI release gate.

KVScope: KV-Cache Memory Measurement

Two self-authored technical reports (Zenodo, DOIs) measuring KV-cache memory on H100 and H200 with PyTorch forward hooks correlated against NVML; density varies about 13x across models.