Shubham Luharuka

Designed and deployed an AI-driven APIs Triaging Agent, leveraging Playwright for Postman collection reverse-engineering. Built Docker-isolated cURL validation pipelines for multi-tenant reliability and integrated an LLM layer to analyze API traffic, enabling automated detection of client-specific threat patterns. Engineered an event similarity pipeline using Qdrant vector database and leveraging Dense and Sparse embeddings for sentence similarity detection. Used cross-encoder model for result re-ranking for enhanced accuracy. Led a team of four, providing technical mentorship and driving deliverables while fostering professional growth.

Built a production-grade Nexus AI Command Center using a RAG architecture, integrating NeMo Guardrails for query-intent validation, LLM-based query rephrasing, hybrid retrieval with Qdrant (vector) + Elasticsearch (BM25), and transformer-based re-ranking for high-precision context selection. Designed dynamic result rendering and delivered a secure, scalable NLQ system enabling natural-language access to complex client databases. Migrated legacy models from OpenAI API to open-source fine-tuned LLMs using Quantization, PEFT/LoRA, GPTQ, AWQ, and BitsandBytes; deployed at scale, cutting costs by 65% and serving 10M requests/month with 400480 tokens/sec throughput. Converted a generic 7B LLM into a cybersecurity-focused model by performing large-scale unsupervised domain pre-training, followed by SFT on tasks like log triage and threat classification. Applied model pruning, quantization, and LoRA/DoRA-based optimization to reduce latency and improve inference efficiency while maintaining high domain accuracy. Developed a cyber crime classification model for the Indian Cybercrime Coordination Centre, including Text classification, OCR, Translation, and Language Detection. Deployed on a bare-metal Kubernetes cluster with GPU access, scaled using KEDA, and monitored using Grafana. Built a PII and fraud-detection module leveraging LLaMA (RLHF) for intent/PII detection, RAG for evidence grounding, and vector similarity search raising fake customer-care detection accuracy from 63% to 98%. Designed and implemented a video classification model to identify Beautified and AR-applied videos. Incorporated scene detection to perform scene-based splitting and applied frame segmentation for identifying multiple subjects within a single frame (e.g., video conferencing). These enhancements significantly improved deepfake detection by reducing false positives.

Built a performance monitoring dashboard to track classification model APIs across environments, capturing latency, throughput, errors, and accuracy, enabling real-time validation, reducing regressions, and ensuring reliable deployments. Built and deployed a multilingual BERT-based pipeline (15 languages) for cyber threat classification across social media platforms; Dockerized and scaled on Kubernetes, improving accuracy from 74% to 96% by reducing false positives.

Built a unified monitoring and visualization system with real-time API dashboards and integrated memory/time profiling, enabling faster debugging and environment validation boosting employee efficiency by 25% and reducing Spindle-X runtime from 2.5 to 2.1 minutes.