Case / 01
CorpGPT: enterprise search
Employees spent 15–30 minutes searching across fragmented internal sources. In the CorpGPT pilot, this fell to 5–10 minutes.
Project catalog
Published: 09Projects and engineering decisions
Problems, architectural decisions, and constraints in one catalog designed to grow with new case studies.
Case / 02
SQL Change Agent: pre-release review
Manual review of SQL data-mart changes took about 90 minutes. The agent reduced analysis to 35 minutes and helped cut post-release incidents by 35%.
Case / 03
RCA Agent: data incident diagnosis
Diagnosing data degradation required manual checks across five sources and took 4.5 hours on average. A stateful agent reduced MTTR to 2.1 hours.
Case / 04
Document Intelligence: construction records
A monthly flow of 20,000–30,000 construction documents required manual transcription. A hybrid pipeline processed about 70% without human intervention.
Case / 05
DOM.RF Document RAG: current regulations
Finding the correct version of a regulation took 10–15 minutes. Version-aware RAG reduced this to 1–2 minutes without mixing revisions or table conditions.
Case / 06
PIK Service Desk: routing with fallback
Short construction-site tickets could contain several problems at once. A multi-label pipeline cut routing errors by 35–40% while retaining manual fallback.
Case / 07
FESCO: supply-chain reconstruction
Up to 40% of the event stream was noisy or contradictory. A graph-constrained Transformer reduced the Manual Correction Rate from approximately 70% to 15%.
Case / 08
1C-Rarus: CPU ticket routing
A transformer added only 1–2% F1 at the cost of a GPU and higher latency. A hybrid LightGBM pipeline automatically routed 70–75% of tickets.
Case / 09
1C-Rarus: B2B lead scoring
The model prioritised B2B leads and accounted for drift in the incoming flow. In an A/B test, conversion rose by 10.5% and first-contact time fell by 25%.