Senior Applied ML Engineer in at Ellkay LLC

• Develop schema and document discovery, clinical entity extraction, semantic classification, terminology recommendations, and data-quality anomaly detection.
• Define clinical relevance labels and benchmarks for source-to-ELLKAY's canonical data model and FHIR mappings, archive retrieval, and interoperability recommendations.
• Combine domain rules, embeddings, supervised models, and LLMs. Preserve patient context, units, dates, negation, and source references.
• Prototype archive search and cited answers. Specify abstention and SME review when evidence is incomplete or clinical meaning is ambiguous.
• Design pilot experiments and analyze reviewer acceptance, error severity, task time, and performance across source-system families.
• Train and adapt models where labeled data and evaluation justify it. Compare against simple baselines, and retain source evidence for extracted facts and recommendations.

• 5+ years in applied ML, NLP, or equivalent production engineering experience, including model development and evaluation.
• Strong Python and SQL, applied statistics, experimental design, supervised learning, information retrieval, and NLP using modern ML frameworks.
• Hands-on experience with healthcare datasets, EHR/lab workflows, HL7 v2 or FHIR, and clinical terminology.
• Experience curating labels, building representative evaluation datasets, analyzing error severity, and preventing leakage across customers or source families.
• Ability to combine clinical rules, embeddings, trained models, and LLMs, and measure confidence and abstention.
• Ability to work directly with clinical SMEs and communicate model behavior to product and engineering teams. A relevant degree or equivalent practical experience is acceptable.

• First 90 days: define the selected healthcare use case, produce an SME-reviewed benchmark with error severity labels, and establish manual/rules/model baselines.
• By six months: deliver a discovery, extraction, or retrieval pilot that meets agreed clinical quality thresholds and reports precision/recall, evidence relevance, coverage, and abstention.
• By 12 months: target 30% less analyst discovery/review time on comparable pilot tasks, with no increase in critical errors. Demonstrate performance across two source families and preserve provenance for all exposed clinical facts.
• Quality, performance, and business-impact targets will be calibrated to the selected use cases and their initial baselines. Delivery depends on the corresponding platform capabilities and approved customer pilot scope.

Awards: