AJAY SHARMA

• Built multi-agent systems to support real-time operational reporting for telecom towers, including metrics like energy consumption, uptime, diesel usage, and fault trends.
• Knowledge retrieval from internal repositories (e.g., SOPs, SLAs, safety checklists, compliance documents), enabling field engineers to query critical information on demand.
• Live web search integration for market and regulatory intelligence (e.g., vendor comparisons, environmental policy updates).
• Integrated APIs, such as the DuckDuckGo API for real-time fact-checking and source attribution.
• Enabled multimodal input processing: Text: fault logs, and alerts.
  Image: tower site condition photos via OpenCV.
  Audio: Technician reports using Whisper for speech-to-text.
• Implemented long-term memory and contextual caching, allowing agents to retain insights on recurring issues (e.g., power failures, equipment mismatches), vendor performance, and service history.
• Delivered structured Markdown and JSON outputs to integrate seamlessly into Indus Towers’ internal dashboards, including Power BI and custom admin panels.
• Developed real-time monitoring dashboards for agent health, task performance, and latency tracking, boosting transparency and operational trust in AI systems.

Text-to-SQL generation using GPT-3, LLaMA 3.0 (Q-LoRA), and RAG.

• Developed a Text-to-SQL pipeline leveraging the GPT-3 API to generate five diverse natural language questions per input text, enhancing dataset richness and semantic diversity.
• Fine-tuned LLaMA 3.0 model using Q-LoRA on the enriched dataset to efficiently map user queries to accurate SQL statements, while minimizing computational overhead. •
• Integrated Retrieval-Augmented Generation (RAG) in the input query stage to ground responses in relevant context and significantly reduce model hallucinations during SQL generation. Personalized AI Image Generation using SDXL and LoRA.

Personalized AI image generation using SDXL and LoRA.

• Built a personalized image generation pipeline using SDXL, fine-tuned with LoRA for subject-specific customization. Integrated Flux-dev and ComfyUI for modular, node-based image workflows, and real-time LoRA injection.
• Used quantized LLMs (LLaMA/Mistral) to auto-generate user-specific prompts for high-quality outputs. Applied PAFT to optimize training speed and reduce GPU usage without compromising output fidelity. Enabled scalable, multi-user image generation, with CLIP-based reranking for best result selection.

Interaction insights of customers (NLP using generative AI, ChatGPT, and Azure)

• Created a tool capable of extracting insights from text data in six different languages. Specifically focusing on chat conversations between customer service agents and customers.
• This tool categorized text data into four key categories: Agent, Process, Service, and Technology, and predicted customer loyalty, churn, journey stage, effort, and intent behind customer contacts using the ChatGPT API for advanced language processing.
• Used skills like NLP (Natural Language Processing), Azure, OpenAI, MySQL, Python, and Power BI.

Developed an age and gender detection system.

• Implemented YOLOv5 for accurate face detection and extraction.
• Trained a ResNet50 model for age regression and gender classification.

Social Media Caption Genre Classification

• Developed a BERT-based classification model to categorize social media captions into 21 genres using Hugging Face Transformers.
• Preprocessed data, performed label encoding, and fine-tuned a DistilBERT model, achieving high accuracy on the test data.
• Built an inference pipeline for real-time genre prediction and optimise model performance through hyperparameter tuning.