Sakhinala Sanjay Bhargav

Challenge: Electromagnetic simulations for 6G antenna design required 500+ CST API calls per design iteration, each taking 15-20 minutes—making optimization prohibitively expensive and time-consuming for production environments.

• Designed hybrid SAC + LightGBM surrogate model architecture with uncertainty quantification
• Implemented custom Gymnasium environment with fabrication-feasible reward shaping
• Reduced CST simulation calls by 72% (420 calls vs ~1,500 CST Studio calls)

Outcome: Achieved 72% reduction in simulation costs, 40% fewer episodes to convergence, and uncertainty-aware predictions enabling safe real-world deployment. Research published in IEEE proceedings.

Challenge: Most RAG systems only support English, leaving 1.3 billion Indic language speakers without accessible document QA tools — and a single-pass retriever can't decide it needs another source or catch its own unsupported claims.

• Built hybrid retrieval on BAAI/bge-m3 dense vectors (1024d) fused with BM25 via Reciprocal Rank Fusion, reranked by bge-reranker-v2-m3, across 10+ Indian languages
• Rewrote the core (v2.0.0) into a LangGraph agentic pipeline — 6-node state machine, 6 tools (corpus retrieval, arXiv, Semantic Scholar/OpenAlex, web search, calculator, sandboxed Python), with a reflexion loop scoring faithfulness and completeness
• Added multi-provider LLM failover (v2.3.0) — Gemini + OpenRouter with per-model circuit breaking and a guaranteed Gemini backstop — plus multimodal figure/table indexing and cross-source contradiction detection

Outcome: Precision@5 of 1.00, Recall@5 of 0.917, and Citation Grounding of 0.938 on the retrieval layer, with an agentic pipeline that can plan, use tools, and critique its own answers — versioned to v2.3.0.

Challenge: Traditional messaging apps store messages on centralized servers, creating privacy vulnerabilities. Need for truly secure peer-to-peer communication with enterprise-grade encryption without server intermediaries storing plaintext.

• Engineered hybrid AES-256 + RSA-2048 cryptographic system for maximum security
• Implemented zero-knowledge architecture with client-side encryption and secure key exchange
• Built multi-threaded TCP socket server with SHA-256 integrity verification for real-time messaging

Outcome: CLI-based messenger with enterprise-grade end-to-end encryption, zero server knowledge of message content, and multi-user support—ensuring complete privacy for sensitive communications.

Challenge: Satellite imagery classification for land-use monitoring required interpretable models beyond black-box accuracy. Urban planners and environmental agencies needed transparent decision-making for policy guidance.

• Developed CNN architecture trained on EuroSAT dataset (10 land-use classes) using a 70/15/15 train/val/test split on a Tesla T4 GPU
• Integrated Grad-CAM explainability layer for visual feature attribution
• Implemented model interpretability pipeline for transparent predictions

Outcome: Achieved 89% test accuracy across 10 land-use categories with explainable AI visualizations showing which image regions influenced predictions, enabling trustworthy deployment for environmental monitoring.

Challenge: Network administrators lacked comprehensive tools to benchmark DNS resolver performance globally. Existing solutions didn't provide statistical rigor or actionable insights for infrastructure decisions.

• Built automated benchmarking framework testing latency across global DNS resolvers
• Implemented statistical analysis pipeline with percentile metrics and reliability scoring
• Developed CLI tool with automated reporting and network troubleshooting diagnostics

Outcome: Production-grade DNS performance analysis framework enabling network teams to make data-driven resolver selections, troubleshoot outages, and optimize query latency for enterprise infrastructure.