Neuroscale AI operates as an affiliated entity of Intellectual Point, collaborating to advance AI-driven decision-making solutions.

AI Agents for Decision Making: DeepSpeed, PyTorch, vLLM, LangChain, PGVector, RAG, Multimodal

• Developed Arbi, an AI copilot using AI Agents to evaluate RFPs, RFIs, resumes, and customer call transcripts.

• Extracted data from PDFs and images using Qwen-2.5-VL, handling complex structures and interpreting diagrams.

• Fine-tuned QwQ/Qwen-32B with PyTorch, DeepSpeed for memory-efficient distributed multi-GPU training,

achieving 10x speedup, and LoRA for PEFT, enhancing reasoning capabilities via knowledge distillation.

• Applied Quantization Aware Training to models, reducing size and compute by 4 times with minimal loss.

• Leveraged vLLM for optimized inference, continuous batching, and low-latency high-throughput deployment.

Emotional support Chatbot: NLP, Hugging Face, PyTorch, LORA, SFT, and Kafka

• Built emotional assistance chatbot by fine-tuning MISTRAL-7B on ExTES dataset using Low Ranked Adaptation (LORA) and Supervised Fine Tuning (SFT) with Hugging Face and PyTorch, reducing trainable parameters by 10,000 times and GPU memory requirement by 300%.

• Utilized Bag of Head Nouns and Personal Space Graph to increase the context length, while preserving the content.

• Utilized GPT-4 to evaluate responses, achieving Spearman and Pearson coefficient of 0.71 and 0.84 against human scores. Deployed on humanoid robot, reducing transmission time by 30% using Kafka.

Author Attribution using Embedding Fusion: ML, PyTorch, LLMs, CNN, LSTM, and Transformers

• Leveraged embeddings from GPT, FLAN, Llama and RoBERTa to get the stylometric of writing between AI and Human generated texts using CNN, BiLSTM and Transformer architecture.

• Performed embedding fusion of the above embeddings to classify achieving a minimum of 96% accuracy and 93% MCC on Deepfake Text Detection dataset.

Efficient Data Extraction and Retrieval: NLP, Selenium, LangChain, docker and pinecone.

• Scraped details of around 100K staff members from 2000+ educational institutions with SELENIUM and preprocessed into a structured format using prompt techniques with GPT-3.5 turbo from OpenAI, resulting around 300% boost in efficiency for targeting clients.

• Powered up storage and retrieval pipeline by leveraging LANGCHAIN, integrating PINECONE, an embedding model (instructor-base) to retrieve details, and deployed on DOCKER, contributing to streamlined operations.

Web-based Query Resolution Chatbot: GPT, HTML, CSS, Bootstrap, FastAPI

• Created a chatbot to answer queries about the company using GPT-3.5 model and integrated it into the company’s website using HTML, CSS, Bootstrap and FastAPI.

Data Extraction, Search and Retrieval Pipeline: NLP, ML, Haystack, Tabula, Elasticsearch, Azure

• Extracted data from pdf using TABULA and preprocessed it into sentences and text files followed by indexing.

• Developed a searching mechanism using HAYSTACK to retrieve the answer from embeddings attributing source.

• Implemented ELASTICSEARCH as a search engine for indexing and retrieving increasing search efficiency by 60%.

• Collaborated with 4+ clients from the construction industry to improve search results and ease of usage.

• Azure Blob was used for storage and application was deployed using Azure VM.