Direct LLM Integration: Replaced Genie’s constrained setup with direct GPT and Gemini connections through Vertex, enabling control over model selection, prompting, and retrieval-augmented generation (RAG).

SQL Generation & RAG Refinement: Completed and optimized the Vanna.ai integration, improving conversational SQL accuracy and adding RAG to preserve context across multi-turn chats.

Data Security & Multi-Tenancy: Implemented governance to ensure SQL queries respected client boundaries and protected patient-sensitive data, supporting multi-tenant use cases.

Conversation Management: Built CRUD controls for prior conversations, allowing analysts to refine, re-run, and manage historical queries.​

Platform Integration: Delivered API endpoints, deployed MCP clients, and connected directly to Databricks data storage for secure query execution.

Schema-Driven Extraction: Designed custom JSON schemas to capture key entities such as commodity, deposit type, grade/tonnage, location references, and exploration stage. This ensured consistency across varied document sources.

NLP + LLM Pipeline: Applied large language models to parse press releases and assessment reports, map textual mentions into schema fields, and resolve ambiguities in deposit references.

Geolocation Integration: Developed logic to map extracted coordinates or descriptive location references into GIS-ready lat/long polygons, ensuring compatibility with Mapbox and PostGIS systems.

Quality Control: Incorporated validation layers to flag uncertain georeferences and normalize attribute values (units, deposit classifications) to industry standards.

Workflow Alignment: Delivered schema outputs that could plug directly into the client’s exploration data platform, providing searchable and mappable insights.

Gene Mention Disambiguation: Applied LLMs to distinguish true gene mentions from proteins, enzymes, and diseases; iteratively refined prompts to maximize precision without losing recall.

Hybrid Modeling for Scale: Trained a local LLaMA model on GPT outputs, enabling low-cost inference for most cases while routing ambiguous mentions to GPT for high-confidence adjudication.

CNV Detection & Normalization: Extended the pipeline to identify CNVs in literature and mapped them into one of two accepted notations, ensuring consistent representation across the CKB.

Workflow Integration: Delivered modular services integrated directly into client’s curator pipelines, supporting both Mastermind search and CKB curation.