Professional

My professional work has been hands-on AI/ML science: turning commercial or scientific ambiguity into algorithms, representations, model architectures, training objectives, evaluation protocols and usable AI systems.

The center of gravity is model and algorithm work: multimodal learning, deep learning, NLP, synthetic data, classification, anomaly detection, calibration, experimentation and the scientific judgment needed to decide whether a model is genuinely learning the right structure.

Choreograph

Choreograph is WPP's global data products and technology company, working across audience insights, planning, AI-based media optimization, predictive analytics, data enrichment and modeling for advertising. The projects I worked on were joint WPP-Google initiatives at the intersection of WPP's media/data business and Google's marketing and cloud technology. In practice this meant early-stage adtech product work around multimodal AI, synthetic data and unified customer understanding.

The AI/ML science work has focused on multimodal models that help form a unified customer view from heterogeneous evidence: surveys, video, text, imagery, behavioral signals, commercial records and large-scale internet data. The modeling work connects information extraction, representation learning, audience prediction, lookalike and propensity modeling, synthetic customer or product data and evaluation of whether the learned representation captures meaningful customer structure.

The scientific challenge is scale and heterogeneity. Customer signals are scattered across formats, platforms and levels of reliability, so the models have to extract useful structure from noisy multimodal data and turn it into representations that remain stable enough for downstream use. That makes architecture choices, sampling strategy, synthetic-data design, evaluation and error analysis central to the work.

Selected Contributions

  • Designed multimodal AI models across surveys, video, text, imagery, behavioral and commercial retail/media data.
  • Built extraction, representation-learning, audience-prediction, propensity-modeling and synthetic-data methods from problem framing through training and evaluation.
  • Developed modeling components and evaluation routines for customer-understanding systems at large scale.
  • Translated retailer and advertiser questions into model objectives, datasets, evaluation protocols and deployable ML capabilities.
  • Connected synthetic data and unified customer-view representations to planning or activation use cases where stability and interpretability mattered.

Methods and Tools

  • Multimodal representation learning
  • Synthetic data
  • Audience modeling
  • Classification and ranking
  • Experimentation
  • Python
  • Model evaluation

Yupana

Yupana was an early fintech startup building AI financial-management automation for B2B finance operations. The product targeted the repetitive operational work around documents, transactions, accounting evidence, exception handling and routing. I joined while the product and infrastructure were still being formed, so the role combined model development, NLP methods, classification algorithms, anomaly detection and applied ML delivery.

I worked across financial-document understanding, field extraction, entity and transaction normalization, transaction classification, anomaly detection, confidence scoring and human-review signals. The core ML questions were how to represent messy financial evidence, separate normal variation from genuine exceptions and calibrate predictions enough for operational use.

This shaped how I think about applied AI/ML science. Accuracy was only one part of the problem. The models had to handle missing fields, duplicate documents, inconsistent labels, customer-specific rules, low-confidence predictions and correction data. The technical work therefore mixed NLP, classification, anomaly detection, calibration, active correction loops and production constraints.

Selected Contributions

  • Designed ML models for financial document understanding, field extraction, transaction normalization, classification and anomaly detection.
  • Built confidence-scoring and correction-loop methods so uncertain predictions could be reviewed and fed back into the system.
  • Owned model code, evaluation logic, calibration checks and production debugging for core financial automation tasks.
  • Turned operational finance problems into NLP, classification and anomaly-detection tasks that could support daily customer workflows.
  • Balanced automation with reviewability so the models could improve from corrections and surface uncertainty clearly.

Methods and Tools

  • NLP
  • Document processing
  • Classification
  • Anomaly detection
  • Confidence scoring
  • Model calibration
  • Correction loops
  • Python

Technical Areas and Operating Principles

AI areas and concepts

  • Transformers and LLMs
  • AI agents
  • Multimodal learning
  • Representation learning
  • RAG and retrieval systems
  • Embeddings and vector search
  • Synthetic data
  • Biomedical AI
  • Robotics and human-machine interaction
  • Evaluation and benchmark reliability
  • Ranking and audience prediction
  • Classification and anomaly detection

Frameworks and tools

  • Python
  • SQL
  • PyTorch
  • TensorFlow/Keras
  • scikit-learn
  • Hugging Face
  • MLflow
  • LangGraph
  • OpenAI Agents SDK
  • AutoGen / AG2
  • LlamaIndex / CrewAI / PydanticAI
  • LangSmith / Langfuse / Phoenix
  • MCP and agent tool integrations
  • Docker/Kubernetes/Kafka
  • CI/CD and cloud platforms

Model evaluation and AI systems

  • Model serving and inference systems
  • Data representation and workflow automation
  • Experiment tracking and model registry
  • Feature stores
  • Model and agent monitoring
  • Trace-based agent evaluation
  • Prompt and workflow evaluation
  • Agent observability
  • Guardrails and policy checks
  • Tool permissions and authorization
  • Model calibration
  • Evaluation design

Operating Principles

  • Start with the product or scientific decision, then choose the model architecture.
  • Use the simplest model or workflow shape that can survive the real operating conditions.
  • Treat evaluation as part of the product from the beginning.
  • Design for correction loops: humans, labels, traces, monitoring and retraining all matter.
  • Move quickly in startups while keeping enough structure that future iterations can be trusted.
  • Translate between research, engineering and product so each perspective shapes the problem clearly.

Management Principles

  • Make ambiguous AI work legible to product, engineering and business stakeholders.
  • Give people enough context to make good decisions independently.
  • Remove blockers quickly and protect teams from unclear priorities.
  • Set standards through review, examples and clear expectations.
  • Create ownership so each person knows what they can decide and where alignment is needed.
  • Keep communication direct and respectful when work is ambiguous or under pressure.