Cristian – Python, PySpark, Databricks

UpToDate Expert AI, a clinical decision-support platform from Wolters Kluwer that answers complex medical questions for clinicians using trusted UpToDate evidence, clinical reasoning and expert validation. Answers must be grounded, cited and safe, because clinicians rely on them in patient care. The platform later expanded to approximately 2,500 U.S. hospitals and health systems.

• Designed and built the multi-agent orchestration with LangGraph, using Phi-3 Mini as a lightweight orchestrator for intent classification, task decomposition and tool routing across seven specialised agents (Analyzer, Builder, RAG, Medication, Evaluator, Critic, Summarizer), escalating only complex reasoning to larger models on Amazon Bedrock;
• Built hybrid clinical evidence retrieval in PostgreSQL with pgvector, combining BM25, dense vector search, clinical metadata filters and Reciprocal Rank Fusion, and extended it with Neo4j GraphRAG over diseases, medications, contraindications and guidelines to answer multi-hop clinical questions;
• Fine-tuned a Llama 3.1 clinical reasoning model with LoRA on Amazon SageMaker using clinician-reviewed questions, evidence and reference answers, and integrated the adapter into production agent workflows;
• Built MCP tools for dosing, contraindication and drug-interaction services, so agents call authoritative clinical sources instead of relying on model knowledge;
• Designed a multi-layer safeguard combining Bedrock Guardrails, a fine-tuned DeBERTa-v3 classifier and the Evaluator and Critic agents to block, regenerate or escalate unsafe, PHI-exposing or unsupported answers for clinical review;
• Built multimodal evidence ingestion with Amazon Textract, Comprehend Medical and pydicom, turning scanned reports, free text and DICOM metadata into structured entities and retrieval-ready documents;
• Set up automated RAGAS evaluation of retrieval relevance, faithfulness and citation coverage, and deployed all services on Amazon ECS with CloudWatch monitoring of latency, token usage and safeguard triggers.

TrialLens, Clinical ink’s analytics platform for decentralised clinical trials, which study teams use to monitor connected medical devices, patient adherence and digital biomarkers. Glucose monitors, wearables and phones send data in different formats and at different frequencies, often late or duplicated. Researchers need to know which readings they can trust before those readings feed trial analytics

• Designed and built the real-time medical-device streaming platform behind TrialLens. Azure Event Hubs ingested telemetry from CGMs, wearables and mobile devices into Databricks Auto Loader and Spark Structured Streaming, and Azure Service Bus handled retries and dead-letter processing between services;
• Designed the Bronze/Silver/Gold layers on Delta Lake (ADLS Gen2): an immutable bronze layer with the raw payload for replay, silver for schema, range, duplicate and late-event checks, and gold for participant summaries and windowed biomarker features, with Unity Catalog lineage and Databricks Workflows orchestration;
• Introduced quality flags for late uploads, sensor gaps, duplicates, clock drift and implausible readings instead of silently dropping records, so researchers could still review questionable data and no unreliable reading reached analytics unflagged;
• Built Python/Pydantic validation and PySpark/SciPy feature pipelines that check device IDs, timestamps, units and ranges and derive glucose, heart-rate, activity and sleep features for digital-biomarker and anomaly-detection models;
• Linked data from multiple devices through a study-level participant ID and a device-to-participant mapping, keeping the original device ID and timestamp for traceability;
• Set up Azure Data Explorer as the low-latency layer for recent telemetry, letting study teams investigate device freshness, sync failures and missing-data windows without scanning full history;
• Built conversational AI agents with Semantic Kernel and Azure OpenAI, so study teams can ask plain-language questions about participant compliance, device failures and site performance and get answers from validated data through governed tools.

An automated scoring system at ETS, a major nonprofit testing organisation, that rates spoken English responses on pronunciation, fluency, vocabulary, grammar, prosody and discourse. Its scores have to closely match those of human raters, unusable responses have to be filtered out, and high test volumes have to be scored fast.

• Built the end-to-end training and production inference framework for automated English scoring, helping scale production scoring to 12,000+ responses per hour;
• Built reusable Python training pipelines (scikit-learn, pandas, NumPy) for preprocessing, feature selection, tuning and model comparison, and benchmarked regression and ensemble models against human scores on 219K+ held-out responses;
• Designed the runtime inference framework on Apache Storm and Java for ingestion, feature extraction, filtering, model inference and score aggregation, integrating the Python models into high-throughput services;
• Developed quality-control models that filter out low-quality audio, unreliable ASR output, non-English and off-topic responses before scoring;
• Built a hybrid ML/LLM scoring architecture with a confidence-weighted score-fusion layer that routes uncertain responses to fallback or additional review;
• Containerised the Python and Java services with Docker and deployed the distributed platform on AWS (EC2, S3, ECR, CloudWatch).

A customer churn analytics programme for T-Mobile, delivered through SAS, that finds subscribers at risk of leaving so marketing can target them with retention campaigns. It combines usage, billing, subscription, device, CRM and campaign data across T-Mobile’s subscriber base.

• Built and deployed an XGBoost churn prediction model on Amazon SageMaker that produced regular customer-level churn risk scores for T-Mobile’s retention and engagement campaigns;
• Built the customer data pipeline: Apache Kafka ingested call records, SMS, billing, subscription, device and campaign data into Amazon S3, and PySpark jobs on Amazon EMR engineered features such as usage, tenure, plan changes and upgrade eligibility;
• Created the central customer analytics dataset in Amazon Redshift, combining behaviour, billing, CRM and campaign data into one source for training and scoring;
• Added social-network features with Spark GraphFrames (PageRank, connection count, communities, neighbouring churn) based on call and messaging relationships;
• Evaluated the model on ROC-AUC, precision, recall and lift, focusing on finding high-risk customers for campaigns;
• Worked with T-Mobile’s marketing, customer intelligence and data engineering teams to review campaign results and retention trends, feeding them back into features and model retraining.