Hi, I'm Yogeshvar.
I build things that ship.

Userspace VMM simulating OS demand paging on limited frames. Intercepts SIGSEGV via sigaction, extracts faulting addresses from ucontext_t, toggles mprotect(). FIFO and Third-Chance page replacement.

Uniprocessor scheduler with FCFS, preemptive SRTF, and 5-level MLFQ (5–25ms quanta). Thread-safe semaphores from scratch via pthread_mutex and condition variables fully passive-wait, zero spin-locks.

Buddy + Slab allocators on a custom heap power-of-two partitioning, recursive merge, kernel-style Slab Descriptor Tables. my_malloc / my_free with embedded metadata and user-aligned pointers.

End-to-end IoT sensor data processing and ML pipeline ingestion, cleaning, feature extraction, and predictive modeling on time-series data across multiple iterative development stages.

Full IST 597 Deep RL course MDPs, Bellman, REINFORCE, PPO, Actor-Critic, VDN multi-agent, DQN+ICM exploration, offline behavioral cloning. Unified runnable notebook covering all 3 homeworks and the final project.

Personalized news recommendation using Deep RL on MINDsmall. Compared MC, Q-Learning, SARSA, DQN, and DQN+ICM curiosity-driven exploration significantly improved cold-start recommendation performance.

6-layer Transformer encoder pre-trained on 3,883h of unlabeled IMU data with 15% masked patches, fine-tuned for forecasting + HAR. 25.5% MAE reduction, 72.6% accuracy across 6 activities, F1 0.91 on Sleep.

Binary and multi-class classification of cardiovascular disease severity replication of ML fusion approach. Compared Logistic Regression, Random Forest, Gradient Boosting, SVM, and ensemble fusion.

ML analysis of child welfare indicators across 33 Telangana districts using government ICDS data (2020–2021). Predicted wasting, stunting, and underweight from climate, infrastructure, and agricultural features.

Text classification pipeline for news article category prediction TF-IDF and embedding features, Logistic Regression, Naive Bayes, and deep classifier comparison with full evaluation metrics.

Regression ML predicting fish species life span from biological and ecological features size, habitat, diet. Random Forest and Gradient Boosting with RMSE/MAE/R² evaluation.

Q-learning agent trained to play Tic-Tac-Toe via epsilon-greedy self-play. Q-table convergence analysis, win-rate evaluation against random and greedy opponents.

One-Class SVM + CNN on 1K+ dermoscopy images. HOG feature extraction cut false positives 30%; tuned classifiers improved early-detection 20% on 500+ test images.

Three-model smoking detection: MobileNetV2 transfer learning baseline, EfficientNetB3 fine-tuned for accuracy, and River online learning for incremental streaming inference no full retraining.

Fine-tuned BERT and RoBERTa to detect logically flawed science questions. Up to 91% accuracy with error analysis on ambiguity and subtle logical failures across disciplines.

BERT classifier identifying which LLM generated a given text completion. Trained on multi-model completions with feature analysis on stylistic fingerprints across GPT, Claude, and open-source models.

Replication of Zero-Shot Composed Image Retrieval with Textual Inversion. Combines image + text modification query in CLIP embedding space via pseudo-word textual inversion no training pairs needed.

Modified CrossEncoder with novel scoring (logits + cosine similarity on [CLS]) on 96K query-doc pairs. Hard negative mining + combined loss → +0.26 nDCG@10 and 50%+ ranking precision lift.

Character chatbot fine-tuned on Friends TV show dialogue. Select Ross, Rachel, Monica, Chandler, or Joey and chat in their style Gradio UI with character portraits and themed background.

RAG chatbot serving 50+ students recursive PDF chunking, sentence-transformer embeddings, FAISS ANN retrieval. 35% answer accuracy lift via prompt and role-based tuning.

Predicting US flight delays by fusing 725K flight records with NOAA hourly weather. Cyclical sin/cos time encoding, traffic proxies, Random Forest model. 3.2× delay multiplier simulated in severe winter weather.

Statistical analysis of US gun death data with agglomerative hierarchical clustering. Trends by year, age, race, intent, and place dendrogram and demographic group interpretation.

Data science work from CHEARS internship EDA, feature engineering, and ML modeling on healthcare data supporting 40,000+ patient records with sub-second read latency under concurrent load.

Full-stack travel platform with admin dashboard, secure payments, and AI-driven trip recommendations. Real-time itinerary management, booking history, and customer inquiry support.

PHP-based tourism website with destination listings, travel package browsing, and hotel/tour booking form. Responsive design with image sliders and dynamic package pages.