Tiffany
Degbotse
I design and deploy end-to-end ML solutions, from data pipelines and modeling to evaluation and productionization. I care about clarity, reproducibility, and measurable impact.
Machine Learning Engineer
Specializing in
NLP · CV · MLOps
About Me
Bridging Research & Production
I'm a Machine Learning Engineer passionate about NLP, computer vision, and building AI systems that have real-world impact. My work spans speech recognition, retrieval-augmented generation, multimodal AI, and computer vision, with a particular focus on evaluating and improving the reliability of machine learning systems. Across research labs, industry roles, and personal projects, I've designed evaluation frameworks for LLM agents, computer vision models, recommendation systems, and speech recognition pipelines.
I enjoy bridging research and production, turning innovative ideas into reliable, usable systems. Outside of work, I love to host interviews, podcasts, and speak with people making waves in their entrepreneurial journeys.
10+
Projects
NLP · CV · MLOps
Focus Areas
Duke University
Institution
Background
My Journey
Bold Path Fellow
Selected as 1 of 5 inaugural Bold Path Fellows, a highly competitive fellowship providing mentorship, funding, professional development, and hands-on experience advancing a venture from concept to implementation.
Air Force Research Laboratory
AFRL Sensors Directorate (SDIP Intern): Conducting research on confidence calibration for Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems to improve the reliability and trustworthiness of AI-driven target classification.
Spotlight Research Presenter, Duke Conference on Society-Centered AI
Selected to deliver a spotlight research talk at the Duke Conference on Society-Centered AI after my abstract, EchoML: Conversational Explainability via Retrieval over Model Reasoning Traces, ranked among the top 10% of submissions.
Duke Trust Lab Research Lead (Aira)
Led development of AÏRA (AI-Augmented Interdisciplinary Research), an OpenAI-funded research collaboration platform that transforms interdisciplinary research discussions into structured collaboration plans, captures key decisions, synthesizes interdisciplinary knowledge, and enables actionable research outcomes.
BMW Group Machine Learning Engineer Co-op (Data-Driven Dealer Allocation)
Built predictive analytics solutions for vehicle inventory management, leveraging machine learning and feature engineering to improve dealer allocation and forecasting accuracy.
Winner, BMW Robotics Track (Duke AI Hackathon 2025)
Developed Nomad as part of a multidisciplinary team, a multimodal robotic system combining hardware, computer vision, and conversational AI to identify landmarks and provide real-time contextual insights, winning the BMW Robotics Track at the Duke AI Hackathon 2025.
Duke Brinson Lab In collaboration with Eaton Corporation
Led the evaluation of large language models for materials science information extraction, analyzing model performance on identifying materials, sample characteristics, and experimentally measured properties.
Duke Master's In Artificial Intelligence and product innovation
Pursued a Master of Engineering in Artificial Intelligence and Product Innovation at Duke University.
Bachelor's Degree In Computer Science (Magna Cum Laude)
Received a Bachelor of Science in Computer Science from Ashesi University, graduating with Magna Cum Laude honors.
Best Computer Science Thesis Award
Recipient of Best Computer Science Thesis Award for developing an AI-driven video-to-video style transfer system that enables arbitrary style application without retraining while maintaining temporal coherence and audio fidelity.
Portfolio
Flagship Projects
Mood Art
An AI-powered image transformation platform that detects a user's emotional mood from an uploaded photo and applies a complementary artistic style using neural style transfer. Evaluated and compared rule-based, classical machine learning, and deep learning approaches for mood recognition and style generation, analyzing tradeoffs in accuracy, visual quality, and inference performance.
Fitness Agent
Designed and implemented evaluation frameworks for an LLM-powered fitness coaching agent, benchmarking tool accuracy, reasoning quality, trend analysis, goal tracking, and personalized recommendation generation through both deterministic testing and end-to-end agent assessments.
VibeTrack
Designed and conducted evaluations of mood-based music recommendation models trained on the MTG-Jamendo dataset, benchmarking rule-based, classical machine learning, and deep learning approaches for mood classification and song retrieval. Analyzed recommendation quality and model performance to quantify the advantages of deep learning-based music understanding.
Capstone Thesis: AI-Enhanced Video Editing
Award-winning undergraduate thesis evaluating arbitrary video-to-video style transfer methods for visual quality and temporal consistency. Developed a pipeline combining neural style transfer, RAFT optical flow, video stabilization, and adaptive frame blending, then designed evaluation protocols using SSIM and Temporal Warping Error (TWE) to quantify style fidelity and temporal coherence across generated videos.
Gesture Globe Travel Agent
Solely owned the evaluation and benchmarking of a gesture-controlled vacation-planning system, developing a comprehensive framework to compare deep learning and traditional computer vision approaches. Designed metrics for perception accuracy, itinerary relevance, notes adherence, task completion, response time, and overall itinerary quality using both automated scoring and LLM-based evaluation.
Galaxy Morphology AI Classifier
Conducted a comparative evaluation of naive, classical, and deep learning methods for galaxy morphology classification. Analyzed robustness, label uncertainty, calibration, and augmentation strategies using the Galaxy Zoo dataset, showing improved generalization and reliability through confidence-aware labeling and data-centric modeling.
MediTriage‑LLM: Fine‑tuned Triage Model
Solely responsible for evaluating a QLoRA fine-tuned Mistral-7B-Instruct model for medical triage information extraction. Designed and implemented a custom evaluation framework for structured JSON outputs, benchmarking performance across department classification, symptom extraction, sentiment analysis, and urgency prediction. Validated improvements over the baseline model on 760 test cases, demonstrating a substantial increase in overall extraction accuracy.
Akan–Sindhi Automatic Speech Recognition (ASR)
Conducted a comparative evaluation of baseline, classical, and transformer-based speech recognition systems for Akan and Sindhi. Analyzed the impact of feature extraction, language modeling, tokenizer selection, and Whisper fine-tuning on transcription quality in low-resource language settings.
Time Series Forecasting
Production-ready temperature forecasting model for RDU Airport using Random Forest and engineered temporal features (lags, rolling stats), achieving 8% RMSE improvement over baseline across 336 hourly forecasts.
Expertise
Skills & Tools
ML/DL
- Supervised/Unsupervised, XGBoost, Random Forests
- CNNs, Transformers, Sequence Models
- Evaluation: ROC/AUC, F1, MSE, SHAP
NLP/LLMs
- RAG, vector search, embeddings
- Fine-tuning, LoRA, prompt engineering
- Speech: ASR/diarization
MLOps
- CI/CD, experiment tracking
- Monitoring, data/version control
- Serving: FastAPI, Streamlit
Data/Systems
- SQL/NoSQL, Pandas/PySpark
- APIs, scraping, pipeline orchestration
- GPU training & optimization
Get in Touch
Let's Work Together
Open to full-time offers, internships, and research collaborations. The fastest way to reach me is via email.