Brian – Machine learning, Data Science, Big Data

To help musical artists and rights holders make their catalogs more "commercially viable" using AI as a tool rather than a replacement.

• Curated custom training datasets from multiple sources and built a hierarchical genre/subgenre classifier using AST (Audio Spectrogram Transformer) 768‑

dim embeddings ensembled with Librosa spectral features, achieving >90% genre and >85% subgenre accuracy with BCEWithLogitsLoss, minority class

oversampling, and per‑genre subgenre MLPs.

• Architected a 6‑stage async real‑time audio pipeline on Lambda + SQS with Apache Iceberg for dataset management, processing thousands of tracks, with

a shadow deployment framework enabling safe rollout of 5+ model versions to production without downtime.

• Fine‑tuned Stable Diffusion/Flux via LoRA adapters for theme‑specific album art generation and orchestrated a multi‑model generative media pipeline

(LTX‑Video, Real‑ESRGAN, Qwen3‑VL) with automated prompt generation, eliminating manual design work and generating hundreds of distribution‑ready

assets.

• Built end‑to‑end Metaflow‑based training pipelines with shared training/inference libraries to prevent train‑serve skew, automated class balancing, and

provisioned concurrency for cold‑start mitigation; making inference 4x faster and reducing AWS costs by over 5x.

• Provisioned full AWS infrastructure via Terraform (VPC, ECS/Fargate, RDS, S3, CloudFront, Lambda, SQS, Cognito), with Datadog observability (dashboards,

monitors, SLOs, synthetics) and 45+ CI/CD workflows across a microservice monorepo.

• Collaborated directly with founding leadership to align AI capabilities with short and long‑term product vision.

A multi-strategy hedge fund management firm that now focuses on delivering a financial platform for investors.

• Architected hybrid Natural Language Processing (NLP) + LLM pipelines to convert unstructured financial calls, news, and filings into structured, actionable investment signals, enabling clients to identify market-moving events faster and reducing time-to-insight for high-value investment decisions.
• Led fine-tuning of transformer and LLM models (FinBERT, BERT, GPT variants) for financial event extraction, sentiment analysis, and summarization, increasing signal accuracy and cutting manual analyst effort by 40%, directly improving operational efficiency and platform scalability.
• Integrated multi-source financial signals into production pipelines, implementing embeddings and ranking systems to surface high-impact events in real-time, accelerating client decision-making and enhancing platform value for enterprise investment teams.
• Developed advanced embeddings and clustering frameworks to identify recurring financial events and emerging market trends, enabling proactive alerts that empowered clients to capitalize on high-priority investment opportunities ahead of the market.

The world’s leading digital cross-device graph. It enables marketers to identify a brand customer or related household across multiple devices, unlocking critical use cases across programmatic targeting, media measurement, attribution, and personalization globally.

• • Built a terabyte‑scale cookieless household clustering pipeline using community detection and simulated annealing over hundreds of millions of device
• signals with GDPR‑compliant data retention, replacing legacy deterministic matching and increasing campaign ROI by 20–30%.
• • Trained 256‑dim device embeddings via PyTorch neural networks with GPU optimization and loss function experimentation, productionizing DS research
• models into a real‑time serving platform on a Ray cluster on Kubernetes hosting 1,000+ models, reducing targeting errors by 25%.
• • Led the Data & AI Platform initiative, building a custom DSL compiler adopted by 10+ engineers, enabling data scientists to write pipelines once and deploy
• to Kubeflow, Vertex AI, or Cloud Composer, with 10+ custom TFX components, a Vertex AI feature store, and MLflow for experiment tracking and model
• registry.
• • Architected hybrid NLP + LLM pipelines integrating 10+ consumer data sources into structured, actionable audience profiles, reducing time‑to‑insight for
• data enrichment products.
• • Led fine‑tuning of transformer and LLM models (BERT, GPT variants) for consumer data extraction, classification, and summarization, increasing signal
• accuracy and cutting manual data processing effort by 40%.
• • Integrated multi‑source consumer signals into production pipelines, implementing embeddings and ranking systems to surface high‑value audience seg‑
• ments in real‑time, accelerating client targeting decisions.
• • Developed advanced embeddings and clustering frameworks to identify recurring consumer behavior patterns and emerging audience segments, enabling
• proactive insights that empowered clients to reach high‑value audiences ahead of competitors.

Data & AI platform solutions for various IBM external clients across diverse industries for ensuring the scalability of their data and machine learning models.

• • Designed a Data & AI platform supporting terabyte‑scale data and ML pipelines with MLflow for experiment tracking, CI/CD (Jenkins), and Airflow orches‑
• tration on Kubernetes, boosting Data Science team productivity by 2x and reducing infrastructure cost by 40%.
• • Built an event‑driven real‑time ML serving platform on Kafka with KNN and Locality Sensitive Hashing, reducing inference latency for optimization used
• daily by users.
• • Productionized anomaly detection models by converting research code from Pandas to PySpark, optimizing SQL queries, and orchestrating training
• pipelines with Airflow on Kubernetes for daily processing of millions of payment records.
• • Developed TensorFlow text classification model achieving 92% accuracy for legal ruling prediction; engineered NLP features using n‑grams and custom
• scoring metrics, improving model accuracy by 11%.
• • Fine‑tuned computer vision models for object detection and image classification, delivering production‑ready solutions for enterprise clients.