GPT Integration in Doha, Qatar

Oil & Gas / Energy

A Qatar Energy joint venture (QatarEnergy + international oil company, operating 4 offshore gas platforms) needed to: make 25 years of technical documentation accessible to field engineers. The document corpus: 180,000+ documents including well completion reports, maintenance procedures, safety assessments, equipment specifications, regulatory submissions, and operational logs. The problems: (1) Findability: engineers spending 45 minutes average searching for a specific procedure or technical reference. Documents stored in: SharePoint (60%), network file shares (25%), email attachments (10%), and personal files on engineer laptops (5% — the "tribal knowledge" problem). No unified search that worked across all repositories. (2) Language: 55% of documents in English, 35% in Arabic, 10% mixed. Engineers needing information in both languages — an English-speaking field engineer needing to reference an Arabic safety assessment written by a Qatari safety officer, or an Arabic-speaking operations manager needing to understand an English equipment specification. (3) Data sensitivity: well performance data, reservoir characteristics, and production forecasts were commercially sensitive — classified as confidential under the joint venture agreement. This data could not be processed through external AI services (OpenAI, Anthropic). (4) Expertise drain: senior engineers retiring or rotating out of the joint venture — taking decades of operational knowledge with them. New engineers asking the same questions repeatedly, with nobody maintaining a knowledge base.

We built a GPT-powered knowledge management platform over 12 weeks: (1) Document ingestion: automated pipeline ingesting documents from SharePoint, network file shares, and email archives. Arabic OCR (for scanned Arabic documents — approximately 15% of the corpus). Metadata extraction: document type classification, date, author, well/platform reference, and topic categorization. (2) Bilingual RAG: vector embeddings generated using multilingual embedding models (handling both Arabic and English documents in a unified vector space). Retrieval: an engineer querying in English could retrieve relevant Arabic documents (and vice versa) — with the system providing translated summaries alongside the original document. FAISS vector store on private infrastructure (no cloud dependency for the retrieval engine). (3) GPT interface: a chat-based interface where engineers asked natural language questions — "What is the procedure for replacing the subsurface safety valve on Platform B?" or "ما هو إجراء استبدال صمام الأمان تحت السطح على المنصة ب؟" — and received GPT-generated answers grounded in the actual document corpus with source citations. The system: explicitly citing which documents it drew from for every answer — engineers could click through to verify. (4) Self-hosted LLM: for queries involving commercially sensitive data (well performance, production forecasts), the platform used a self-hosted Llama 3 70B model running on 4× NVIDIA A100 GPUs in the joint venture's Doha data center. For non-sensitive queries (safety procedures, equipment specifications from manufacturer manuals): Azure OpenAI Service on the joint venture's dedicated Azure tenant. The architecture: automatically routing queries to the appropriate model based on document sensitivity classification. (5) Knowledge capture: a "lesson learned" feature where engineers documented operational insights, troubleshooting solutions, and best practices — structured entries that fed into the knowledge base, gradually building the institutional knowledge that previously existed only in senior engineers' heads.

Search time: from 45 minutes average to: 3 minutes for a specific document or procedure. Engineers: asking the GPT interface natural language questions and receiving answers in seconds — with source document citations for verification. Usage: 120+ engineers across 4 platforms using the system within 2 months of launch. Average: 15 queries per engineer per day — indicating genuine adoption (the system actually useful, not a compliance checkbox). Bilingual effectiveness: Arabic-to-English cross-language retrieval accuracy: 84% (the system correctly identifying relevant English documents from Arabic queries). English-to-Arabic: 81%. The bilingual capability: particularly valued by mixed-nationality teams where Qatari engineers and international specialists needed to access the same knowledge base in different languages. Knowledge capture: 450+ "lessons learned" entries added in the first 6 months — operational insights that previously would have been lost when engineers rotated out. 3 entries: directly preventing equipment failures by alerting field engineers to known issues before they encountered them. Data security: commercially sensitive queries processed entirely on the joint venture's private infrastructure. Azure OpenAI Service queries: covered by enterprise data processing agreement with no data used for model training. Zero data leakage incidents. Estimated annual value: QAR 2.8M ($770,000) in engineer productivity improvement (45 minutes → 3 minutes × 120 engineers × 15 queries/day × 250 working days). Equipment failure prevention: 3 incidents avoided × estimated QAR 500,000-2M per incident = QAR 1.5-6M in avoided unplanned downtime. Development cost: QAR 650,000 ($178,000) over 12 weeks. ROI: development cost recovered within first 2 months from productivity improvement alone.