Large language models such as GPT-4 are now used across specialised production systems, but for high-volume, domain-specific applications the operational picture is challenging: API spend on the order of $110K-550K per year for a workload of roughly ten thousand daily queries, end-to-end latencies that drift between two and ten seconds during market hours, and the obligation to send proprietary data through a third-party endpoint. Sufficiently capable open-source models can address all three constraints, provided they can be adapted to the target domain without losing the format compliance and reasoning consistency that production callers depend on. This thesis develops such an adaptation pipeline for Axyon AI's Alyx, a financial-analysis system whose language model produces stock briefs in JSON, classifies the relevance of news items, writes sector and universe commentary, and summarises longer-form reports.