Seasoned Music
Your Spotify listening, in context
Designed and built a privacy-first analyzer that maps Spotify listening across seasons and places, then turns those patterns into private playlists.
- Next.js
- TypeScript
- Spotify Web API
Open to software engineering opportunities
Hello, I'm
Valentine Ezikeoha
Software Engineer. Problem Solver
AI/ML & Full-Stack Engineering
Software engineer passionate about AI/ML and full-stack development, building technology that solves real problems and drives meaningful change.
Howard University CS graduate • New Jersey
About
Learning, growing, and contributing at scale.
I'm a Howard University Computer Science graduate who builds AI-powered products and dependable full-stack systems. I enjoy turning ambiguous problems into useful, accessible software.
My work spans production automation at Intuit, multilingual AI research, and products that connect modern interfaces with thoughtful backend systems.
- Education
- Howard University, B.S. Computer Science '26
- Experience
- Intuit + AI/ML research
- Focus
- AI products, full-stack systems, developer tooling
Selected work
Products I've built
AI, product engineering, and full-stack systems built around real user needs.
FinBuddy
Financial literacy, made personal
An authenticated budgeting product for creating budgets, understanding spending patterns, and tracking progress toward financial goals.
- Next.js
- Clerk
- Tailwind CSS
Experience
Where I've worked
- Sept 2025 – May 2026
Machine Learning Researcher
Howard University Research · Washington, D.C.
Speech-based Alzheimer's detection and severity assessment
Developed and evaluated a multilingual audio-and-text pipeline to predict Mini-Mental State Examination (MMSE) scores, a 0–30 measure of cognitive function, from spoken responses. The research explores speech as a signal for cognitive screening and monitoring.
- Combined five experimental cohorts from ADReSS, ADReSSo, ADReSS-M, and TAUKADIAL: approximately 659 labeled participants speaking English, Greek, and Mandarin, across different speaking tasks and recording conditions.
- Evaluated XLM-RoBERTa, ModernBERT, and mmBERT on automatically generated transcripts alongside AST, SSAST, and WavLM audio representations. The strongest standalone text model, mmBERT, reached approximately 4.19 RMSE; the strongest audio-only models reached 5.47–5.56.
- Compared weighted late, MLP, gated, uncertainty-aware, and embedding-level fusion. Uncertainty-aware fusion achieved approximately 4.04 RMSE, improving on standalone text by combining linguistic patterns with complementary speech acoustics.
- Used 5-fold speaker-level GroupKFold cross-validation to keep each participant's recordings within a single fold and prevent speaker leakage. Tuned label standardization, Smooth L1 loss, learning rates, audio augmentation, and regularization to improve robustness.
- Python
- PyTorch
- Hugging Face Transformers
- Torchaudio
- scikit-learn
- May 2025 – Aug 2025
Software Engineer Intern
Intuit · Mountain View, CA
- Developed scalable automation infrastructure for QuickBooks Online using TypeScript and Playwright, increasing test coverage across 40+ production user workflows.
- Engineered maintainable automation frameworks to validate complex role-based access control, reducing manual testing effort while improving software reliability.
- Collaborated with cross-functional engineering teams to translate customer feedback into production bug fixes and improved user experience.
- Jun 2024 – Dec 2024
Undergraduate Researcher
Howard University Research · Washington, D.C.
- Optimized neural text-to-speech pipelines by improving multilingual audio preprocessing and text segmentation, reducing synthesis artifacts in long-form speech.
- Improved speech model accuracy by 20% by preprocessing audio data and extracting high-fidelity speaker embeddings for multilingual synthesis.
Contact
Let's build something together.
I'm always open to new opportunities, collaborations, and a good conversation about technology.