Google’s Gemini has matured considerably since its rocky launch. After spending several months using it alongside Claude and GPT-4 for data engineering tasks, I have a clearer picture of where it earns its place — and where it falls short for the kind of regulated, high-stakes work I do at CIHI and for financial services clients.
What Makes Gemini Different
Gemini’s primary differentiator for data work is its native integration with the Google ecosystem and its genuinely strong multi-modal capabilities. If you’re working with BigQuery, Looker, or Google Cloud in general, Gemini’s context-aware assistance inside those tools is more seamless than anything Claude or GPT currently offer in that specific stack.
For Azure-heavy shops like ours, that advantage is less relevant. But the underlying model capabilities are still worth evaluating on their own terms.
Schema Inference and Data Profiling
One area where Gemini surprised me positively: schema inference from messy real-world data. I fed it a sample of a complex HL7 FHIR health data export — the kind of deeply nested JSON that makes most tools struggle — and asked it to infer a flattened schema suitable for a PySpark DataFrame.
The result was 80% correct on the first pass, which is impressive given the complexity of the format. It correctly identified the nested resource types, inferred data types conservatively (opting for string where it was uncertain), and even flagged a few fields that appeared inconsistent across records.
Compare that to manually writing the schema: for a 40-field nested structure, that’s easily an hour of careful work reduced to 10 minutes of review.
SQL Generation for Analytics
Gemini’s SQL generation is strong — arguably the best I’ve tested for complex analytical queries. When I described a reporting requirement in plain English (“show me the average claim processing time by province, broken down by diagnosis category, for the last 3 fiscal years, excluding records flagged as test data”), it produced a correct multi-CTE query on the first attempt.
The key qualifier: it needs a good schema description to work from. Provide it with clear table structures and it performs well. Without that context, it makes reasonable but incorrect assumptions.
Where It Falls Short for Enterprise Use
Context window use: Gemini Advanced has a large context window, but I found it occasionally lost track of constraints stated early in a long prompt — particularly compliance constraints like “never include raw patient IDs in the output.” This is not acceptable for PII-sensitive work.
Consistency: Results varied more than I’d like between runs. For documentation tasks where reproducibility matters, this is a friction point.
Regulatory awareness: When I asked about PIPEDA-specific data handling requirements, the responses were generic and occasionally incorrect. Claude handles Canadian regulatory context better.
My Recommendation
For teams on Google Cloud working primarily with BigQuery and Looker, Gemini is a natural fit and worth prioritising. The native integrations reduce friction significantly.
For teams on Azure/AWS working with sensitive health or financial data in Canada, Gemini is a useful secondary tool — particularly for schema work and SQL generation — but I wouldn’t make it the primary AI assistant. The compliance context handling and consistency aren’t quite there yet.
The LLM landscape is evolving fast. I’d re-evaluate Gemini every six months — Google has the resources to close these gaps quickly.