Ojaswa
Yadav

Research framework for learning latent world models from unlabeled video — no rewards, no action labels, no pixel reconstruction. A frozen V-JEPA ViT-B/16 encoder extracts visual features; an Action Inferrer with a VQ codebook (512 discrete codes) discovers latent action structure purely from consecutive frames. Dual-level predictors handle frame-to-frame and segment-to-segment dynamics. ~50M trainable parameters. Trained on Something-Something-v2, generalization eval on Ego4D, robot manipulation eval on RoboMimic and BridgeData V2.

Silent-failure detection for distributed GPU training. NCCL only reports "rank unknown" when a job hangs — RankSentry instruments every rank via an LD_PRELOAD C shim, streaming collective timings, heartbeats, and GPU telemetry over gRPC to name the failing rank before the watchdog fires — with cause, evidence, and recovery path. Validated on Lambda Labs with injected faults.

CUDA-native LLM inference engine in C++/CUDA — no PyTorch on the hot path. Hand-written kernels for RMSNorm (fused residual+norm via warp-shuffle), RoPE (fused Q+K with precomputed tables), and flash-decode attention (warp-cooperative, online softmax, 0 bytes stack frame). Achieves 323 tok/s decode on an A40 at ~46% peak memory bandwidth. Loads models directly from safetensors; greedy decoding from cuBLAS GEMMs + custom elementwise + two-phase parallel argmax. Validated on Qwen2-0.5B, 14/14 kernel correctness tests passing. INT4/INT8 quantization and custom GEMM in progress.

Conversational AI for financial Q&A combining QLoRA fine-tuning of Qwen 2.5 (4-bit, consumer-GPU-friendly) on the Wealth Alpaca Dataset with three RAG strategies: naive, agentic, and fine-tuned. Vector store via ChromaDB + MiniLM embeddings. Benchmarked with RAGAS-style metrics — 84% answer relevance, 80% faithfulness.

Episodic MDP formulation for routing emails across five detectors spanning a 50,000× inference-time range. DQN achieves 99.54% accuracy while invoking the expensive LLM detector on only 6% of emails — cutting per-email latency from 3.29s to 208ms. Lagrangian dual mechanism enforces a 1.5% FPR hard constraint. Two-phase curriculum with exploration re-injection at phase transition.

Heterogeneous email graph with 5 node types, 7 edge types, dual-mode RAG retrieval, and three parallel role-specialized agents with a verdict-deferring orchestrator. Ablations across three datasets isolate per-agent contribution. Isotonic calibration on all outputs; honeypot filtering and train/test entity-overlap analysis.

CNN for classifying four skin diseases (acne, actinic keratosis, melanoma, psoriasis) built on EfficientNetV2 + Efficient Channel Attention (ECA). Achieves 84.70% accuracy with only 16M parameters — fewer than most prior deep learning approaches in the literature. Published in Biomedical Signal Processing and Control (Elsevier). Most-cited work in my portfolio.