42 / Foundation Models & Science
Foundation Models for Enterprise SQL & Time-Series Forecasting
We trained 2 of our own foundation models and put them on public leaderboards: one that writes enterprise SQL and one that forecasts business time-series, proving an Indian lab can compete at the frontier.
- Built by
- 360 Labs Research
- Stack & approach
- ResearchFoundation ModelsText-to-SQLTime SeriesLeaderboards
The problem
Most models break on enterprise data: huge real databases with 1,000+ columns, multi-step queries over 100+ lines of SQL across different dialects, and time-series forecasting across unlike domains, and reported scores are often inflated by benchmark leakage.
Solution & architecture
2 360 Labs foundation models on public leaderboards. Omni 1.0 (enterprise text-to-SQL) ranked #5 on Spider 2.0-Lite at 66.36, the only Indian entry, beating every Claude/GPT/Gemini agent submission. OmniScient (time-series forecasting) was evaluated across all 98 TIME benchmark tasks, with fine-tuning corpora fully disjoint from benchmarks and their upstream sources: leakage discipline by design.
Research highlights
Leakage discipline
Fine-tuning corpora disjoint from benchmarks
Enterprise-shaped
1,000+ column schemas, 100+ line SQL
Indian frontier model
Only Indian entry on Spider 2.0-Lite
