Token-efficient AI routing platform balancing cost, latency, and quality.

• Multi-tier orchestration combining deterministic solvers, local LLMs, and cloud models with confidence-based routing.
• Real-time telemetry on cost, latency, and quality trade-offs.

Eligibility recommendation platform for government schemes.

• Semantic search and vector retrieval with RAG for contextual recommendations.
• Hybrid structured/semantic retrieval via ChromaDB + MongoDB.

Full-stack translation platform for 15+ languages.

• Asynchronous FastAPI REST APIs for translation at scale.
• Automatic language detection across 15+ languages.

Short-term wind power forecasting with AI-driven accuracy models.

• Compares LSTM, transformer, and hybrid models for wind power prediction horizons.
• Published in Springer Electrical Engineering with 10+ scholarly citations.

Multilingual AI voice/text engagement platform with real-time RAG.

• RAG pipeline using pgvector and LangGraph for context-aware responses.
• Multilingual speech-to-text, calling, and WhatsApp workflows with async lead scoring.

Industrial knowledge intelligence engine turning fragmented engineering evidence into a connected operating memory.

• Graph RAG fusing Neo4j knowledge graphs with ChromaDB vector search for grounded, source-cited answers.
• Automated ingestion pipeline extracts equipment tags, entities, and syncs metadata across Firestore, ChromaDB, and Neo4j Aura.

Multi-agent decision intelligence platform for urban health risk management.

• Cooperating AI agents ingest urban signals and run GPU-accelerated DBSCAN/regression forecasts to triage hazards.
• LangGraph-orchestrated decision and reflection agents route risky calls to a human-in-the-loop approval queue.

Detection and prevention system for ARP spoofing attacks in LAN environments.

• Monitors ARP traffic to identify IP-to-MAC inconsistencies indicative of spoofing.
• Implements mitigation to block man-in-the-middle attacks originating from poisoned ARP caches.

Regularized regression solvers implemented via the Alternating Direction Method of Multipliers.

• Implements Lasso (L1) and Ridge (L2) regression by formulating each as an ADMM consensus optimization problem.
• Compares convergence behavior and coefficient shrinkage across regularization strengths.

Regularization framework for recovering signals from noisy, ill-posed inverse problems, with LLM-assisted parameter tuning.

• Implements and compares Pseudoinverse, Tikhonov, TSVD, NSIT, and FNSIT reconstruction methods against ill-conditioned forward operators.
• Integrates the Gemini API to dynamically suggest regularization parameters from residual and noise diagnostics.

ML pipeline for predicting customer churn from imbalanced tabular data.

• End-to-end workflow from EDA and preprocessing through ADASYN resampling for class imbalance.
• Benchmarks Logistic Regression, Random Forest, and Gradient Boosting models for churn classification.