Professional journey

Agentic ICF Generator — Transforms Clinical Study Protocol tables into ICF Summary Tables using Skills Cards architecture with scoped MCP tool restrictions. 1–2 days manual → under 5 min.
Vision Agentic RAG — Multimodal retrieval system with DSPy serving a global medical writing team. Retrieval hit-rate@6: 60% → 85% across thousands of documents and images.
Custom Document Parser — Docling-based parser extracting tables, figures, and images from PDFs, RTF, DOCX into structured Markdown; indexed into Weaviate + GCS.
LLM-Judge Evaluation Stack — 4 specialized judges (faithfulness, completeness, patient-safety, alignment) scoring at claim level with source-snippet citations and abstention-aware metrics; MLflow-tracked runs, DVC-versioned golden datasets on Vertex AI pipelines.
Judge Meta-Evaluation — Seeded-error datasets measuring per-judge recall/precision, calibration against medical-writer annotations, cross-model verdict verification (Gemini vs GPT) — grounding raised to 1.00, completeness +7 pts.
Streaming Re-architecture — Re-engineered monolithic application with full streaming: long-running task response time cut from ~2 min to ~4 sec.
Native Word Comment Layer — Backend-driven OOXML pipeline (flat-OPC) embedding LLM reasoning and source citations as native Word comments — every AI output auditable and source-traceable inline.

Compound AI Audit Chatbot — Led team of 5, delivered POC in 2 months and production-ready chatbot in 3 months, parsing thousands of documents (PDF, PPTX, images). Answer accuracy: 70% → 94%.
Advanced Retrieval Pipeline — Azure Search reranking with DSPy-driven dynamic keyword generation and recursive retrieval. ~4-sec streaming response time.

QA ChatBot — LlamaIndex RAG + LangChain with auto-retriever composition on Vespa.ai, covering dozens of FAQ topics with sub-50ms retrieval latency.

Personalized Learning Platforms — LLM-powered (Claude on Bedrock) with hybrid search on Vespa, serving Wiley's global base of millions of learners.
Hybrid Search Engine — ANN + BM25 with spell correction, autocompletion, and multilingual search; deployed via Docker & Kubernetes with 4 APIs.
Automated Skill-Tagging — Model covering 8,000+ skills at 95% precision, with full MLOps pipelines (SageMaker, MLFlow, Lambda, Step Functions).

Trained and upskilled Engie's data science team on Amazon SageMaker and PySpark through hands-on sessions, enabling advanced data analysis and model development capabilities.

KYC Remediation — AWS Textract for ID document analysis, MRZ consistency checks via checksum validation, and CRM consistency using text mining.
Churn Prediction — Predictive risk model on customer banking activity; optimized with HyperOpt and LightGBM for improved retention strategies.
Email Sentiment & Transaction Mining — Specialized banking sentiment lexicon (adapted VADER, StanfordNLP/NLTK) and recurrent-payment/subscription detection over transactions with PySpark/Hive.

Cross-Sell Reco Engine — Recommendation system integrating business rules, seasonality, and recurrence with personalized homepage allocation; PySpark/HDFS/Hive on GCP.

Equity Risk Premia — Developed and launched ERP strategies with ML algorithms (Random Forests, SVM), index pricing models, and full integration with trading systems and FACTSET API.