BuildingAIsystemsthatshiptoproduction — notdemos.

As the sole AI Engineer on this project, I own end-to-end R&D for a production medical coding pipeline targeting a market projected to reach $8.4B by 2033. I avoided a flawed vanilla-LLM baseline — sub-70% coding accuracy — and instead designed a two-stage retrieve-then-rank architecture grounded in three authoritative CMS sources, with Medicaid and Medicare compliance rules baked in. The result is a 7.66% error rate against a 10–20% manual-coding industry baseline, in a sector where coding errors cost the healthcare industry over $1B annually.

Shipped a multi-agent LLM pipeline on Azure Databricks for natural-language-to-SQL over 20K+ product records, containerized with Docker for reproducible runs across dev and staging. Built a RAG assistant over 20+ documents using transformer embeddings, FAISS vector indexing, and GPT-4 intent routing — 17% better answer accuracy than an embedding-only baseline. On the analytics side, A/B tests and causal regression quantified incremental campaign lift for a top-3 pharmaceutical client, enabling spend optimization that lifted ROI by 4.2%.

Pulled 10M+ transaction records from Snowflake to engineer a hybrid recommendation system pairing Market Basket Analysis with Word2Vec embeddings, served via vector similarity search — drove a 6.7% lift in Average Order Size. Built an entity-resolution layer (NER plus embedding-driven cosine similarity) to reconcile partial company names from Azure SQL into canonical entities, and trained in-warehouse additive and multiplicative time-series forecasts via Snowpark to 12% MAPE for inventory and production planning.

Built churn-prediction models on user behavioral data at a Series-A Ed-Tech with 10M+ app downloads, generating risk scores for retention initiatives via pandas, scikit-learn, and AWS SageMaker. A parallel telemetry analysis on AWS S3 surfaced latency bottlenecks in the funnel, driving targeted system optimizations that cut app load times 10% and payment-page churn 12%.