Dr. Jeffrey Otoibhi

• Boosted LLM output quality by 20% while cutting prompt length by 88% through a clinical evaluation of 200+ user dialogues.
• Engineered an agentic data pipeline for real-time conversation analysis, enabling automated health insights and dynamic visualization.
• Led the development of an automated "LLM-as-a-judge" framework to evaluate safety, triage accuracy, empathy, and cultural alignment.
• Optimized system architecture to enhance functionality and healthcare assistant reliability.
• Integrated clinical data governance with AI engineering to ensure safe, high-quality, and validated triage protocols.

• Designed and deployed a clinical AI system integrated with EHR via FHIR, featuring context-aware history and agentic appointment orchestration for a live healthcare environment.
• Developed automated evaluation frameworks (including LLM-as-a-judge) to mitigate hallucinations and ensure factual accuracy, prioritizing clinical safety at scale.

Lagos, Nigeria

• Responsible for designing and evaluating an AI-powered dermatology diagnostic system tailored to analyze skin images and handle workflows, providing:
• Diagnosis of dermatological conditions and possible home remedies.
• Scheduling appointments and handling payment systems

Lagos

• Led the development of Nigeria's first decoder-only foundational language model supporting multiple Nigerian languages (Train loss- 2.5, Val loss-2.86).
• Outperformed GPT-4 by 48% and LLaMA2-13B by 259% on average across Nigerian language benchmarks for downstream tasks.

Lagos

• Designed, developed and deployed advanced AI chatbot systems seamlessly integrating them into the business applications for 4 clients.
• Employed strategic prompt engineering approaches improving AI responses (correctness and hallucinations) by 25%.
• Reduced AI application latency by 15% through optimized features.
• Provided technical support to over 20 client teams with a 95% resolution rate.

• Distributed LLM Pre-training (Bittensor): Continuous pre-training of GPT-NeoX-20B and GPT-J-6B on the Bittensor dataset using DeepSpeed ZeRO-3 across multi-node GPU clusters. Achieved a final cross-entropy loss of 1.89 (6.62 PPL), boosting TAO token earnings by 15%.
• RAG-Powered Customer Support Chatbot: Designed and deployed high-throughput RAG systems using VectorDB and GPT4All, improving response accuracy by 30% and resolving 10% of inquiries in real-time.
• AI-Driven Traffic & Accident Event Analysis Engine: Engineered a complex vision-language system (VideoViT, CNN-LSTM, Gemini 2.5 Pro, GPT-4o) for accident classification and narrative generation, improving baseline accuracy by over 20%.
• Multi-Modal Clinical Depression Detection (DAIC-WOZ): Prediction of clinical depression using speech and text. Accuracy: 82%, Precision: 0.69, Recall: 0.70, F1: 0.63.
• SignSynth: Speech-to-Sign Language Translator: Built a multi-modal pipeline using Transformers for pose generation and U-Nets for sign language image synthesis (SSIM: 0.9478, MAE: 0.04).
• Adversarial Network Packet Generation: Developed a novel LSTM-based model to generate 350,000 unique adversarial network packets (IP + TCP) for proactive security vulnerability testing.
• Automatic Post-Edit (APE) Translator with Real-time Adaptation: Designed a language translator with real-time adaptation, achieving BLEU-4: 0.43, TER: 0.66 for domain-specific text.
• BERT-Based Appliance Repair Assistant: Created and integrated a BERT-based system via a scalable REST API for appliance repair recommendations, reducing human support needs and improving first-time fix rates.
• Neuronal Response Prediction (Macaque V4): Designed a CNN Core + attention readout model for predicting neuronal responses to excitable images. Correlation score: 0.27.

Remote

• Leading the development of an AI-powered mental health support platform.
• Designing and implementing advanced dialogue systems for therapeutic interactions.
• Developing monitoring and progress tracking features for mental health support.