Abdul Haseeb Khan

• Collaborated with faculty on cutting-edge AI and Neural Networks projects, focusing on current problems with AI such as limiting hallucinations, detection of AI-generated text/images/art/video, and AI in Emergency Medicine in Italy.
• Developed a lightweight CNN-based architecture for AI-generated text detection, enhancing cross-domain generalization via refined stylometric feature extraction and lifting accuracy from 97% to 99% through optimized convolutional depth, regularization, and feature normalization that mitigated overfitting and domain bias.
• Architected a multi-agent decision-support framework integrating fine-tuned LLMs to cross-validate diagnostic outputs, mitigating hallucinations and boosting prediction reliability for critical conditions like Sepsis.

• Designed Clinical GenAI, a Retrieval-Augmented Generation system for EMR analysis: PDF-based clinical data ingestion, vectorized semantic retrieval with Milvus, FastAPI/Uvicorn backend, and GCP deployment, with Gemini-powered generation for accurate, low-hallucination, clinician-focused responses.
• Designed WorkspaceBot AI, a personal multi-agent automation system on n8n that reduced manual workspace effort by 15%, leveraging DeepSeek-R1 and Gemma-3n with Sheets, Calendar, Tasks, and Gmail integrations, advanced prompting, HILT, guardrails, and lightweight memory.

• Led full-stack development with Node.js and TypeScript, building scalable services and growing the user base from 24 to 7,000+ through close collaboration with product and marketing.
• Owned end-to-end component-driven UI architecture, API design, and production deployment, driving adoption by 500+ users and increasing sign-ups by 15%.
• Managed and secured API servers, optimized webhook integrations for faster data exchange, and improved Docker-based deployment workflows for scalability and performance.

Clinical GenAI — Retrieval-Augmented Generation for EMR Analysis

• Built a HIPAA-compliant PDF ingestion pipeline using PyPDF2 to extract patient health records, with token chunking and overlap to preserve clinical context, plus standardized metadata for optimized retrieval.
• Deployed a dockerized Milvus vector database with text-embedding-004, converting unstructured EHR text into 768-dimension vectors for sub-second semantic similarity searches across clinical data and medical literature.
• Combined semantic search (cosine similarity > 0.82) with Gemini 2.5 Flash and prompt engineering to reduce hallucinations and deliver clinician-focused responses.

NEULIF v2 — Lightweight CNN for AI-Generated Text Detection via Stylometric Features

• Enhanced CNN-based architecture for AI-generated text detection, integrating refined stylometric feature extraction to improve generalization across human and LLM-generated content.
• Lifted accuracy from 97% to 99% by redesigning the CNN, addressing overfitting through optimized convolutional depth, data regularization, and feature normalization.

WorkspaceBot AI — Multi-Agent Automation Workflow on n8n

• Self-hosted multi-agent automation that reduced manual workspace effort by 15%, orchestrating agent workflows for email categorization, drafting replies, sending mail, calendar event creation/updates, reminders, and task verification.
• Self-hosted LLM stack with DeepSeek-R1-0528 and Gemma-3n using few-shot prompting, attention for deterministic behavior, and a HILT validation layer for reliability.
• Integrated Sheets, Calendar, Tasks, and Gmail with guardrails and lightweight memory for reliable task execution.

StoxPrediction — Equity Market Analysis & Prediction Engine

• Built a dynamic web app using LSTMs (Keras + TensorFlow) for real-time bond-equity stock price prediction, improving market trend forecasting accuracy.
• Implemented real-time data ingestion and preprocessing pipelines (normalization, sliding-window feature generation) and an intuitive UI on a Django backend supporting continuous LSTM-based forecasting.