NAGRAJ GAONKAR.

Software Engineering Intern at American Express, May '26 - July '26. Optimizing latency and token cost with LLM query caching; exploring latent-concept query pivoting for smarter agentic routing.

California wildfire forecaster over a 580 km x 560 km VIIRS/GridMET grid. Rebuilt NASA FIRMS points into daily fire maps, cleaned 1:641 imbalance with density masks and spatio-temporal features, then stacked PatchTST + DLA + tree meta-classifier for PR-AUC 0.8765 and Monte Carlo t+7 risk.

CMake-based C++ CPU ray tracer for donut/toroid scenes. Uses MPI row distribution plus OpenMP threading across 40 CPUs, dropping a 1600 x 900 render from 19.64h to 0.81h.

CS6370 NLP search engine over 1400+ Cranfield docs. Added LSA, K-Means grouping, custom preprocessing, cached SVD, and beat the VSM baseline by +13.3% recall, +9.3% mAP, and 11% response time.

Minimal production-oriented macOS background agent in C++17. Phase 1 nails lifecycle, POSIX shutdown, thread-safe logs, launchd integration, and timed execution safety.

Multithreaded Formula-1 simulation engine. A producer thread emits 20 ms telemetry snapshots into a bounded ring buffer; a consumer renders a live ANSI leaderboard with tyres, pits, laps, and speed.

Added constexpr support for X86 shuffle intrinsics in Clang, including bytecode interpreter and constant-expression evaluator paths across SSE, AVX2, and AVX512 variants.

InterIIT quant stack for masked exchanges: regime-aware intraday trading, trend following, mean reversion, no-lookahead constraints, parallelized backtests, and cross-dataset generalization.

Adobe InterIIT vision work that became a real/fake image pipeline: CNN + ViT ensembles, adversarial checks with FGSM/PGD, 15,000+ synthetic images, GradCAM, CLIP, and VLM explanations.