🛢️ DrillSense: AI-Powered Operational Intelligence from Daily Drilling Reports

DrillSense is a domain-specific LLM + RAG system engineered to deliver AI-based operational insights and alerts from Daily Drilling Reports (DDRs) and associated petroleum engineering documentation.

It leverages a dual approach combining Supervised Fine-Tuning (SFT) for domain knowledge injection and a Retrieval-Augmented Generation (RAG) pipeline to ensure outputs are accurate, explainable, and contextually grounded.

---

🚀 Key Capabilities

DrillSense is built on a Large Language Model (LLM) fine-tuned on proprietary drilling data and enhanced by retrieval-based context injection. It provides the following core functionalities:

• Automated Summarization: Generates concise summaries of daily drilling operations and activities.
• Domain-Specific Q&A: Answers complex questions regarding PVT (Pressure-Volume-Temperature) data and well procedures.
• Context-Aware Alerting: Provides LLM-based risk or anomaly alerts based on contextual information from the reports.
• Semantic Retrieval: Enables highly accurate, domain-specific search and retrieval of past operations and documents.
• Full-Stack Deployment: Integrated user experience delivered through a React + FastAPI deployment stack.

---

⚙️ Technology Stack

Layer	Technology	Purpose / Details
Frontend	React.js, Tailwind CSS, Axios, React Query	Responsive UI for query submission and results visualization.
Backend	FastAPI, Uvicorn, Python 3.9+	High-performance, asynchronous REST API layer for serving requests.
Vector Database	FAISS (HNSW)	High-speed similarity search and vector storage for RAG.
AI / ML	Vertex AI (SFT), SentenceTransformers	Model fine-tuning, managed deployment, and text embedding generation.
RAG Pipeline	Naive RAG (Retrieval $\rightarrow$ Prompt Composition $\rightarrow$ LLM)	Context injection mechanism to ground the LLM's responses.
Deployment	GCP VM, Nginx, Vercel	Backend hosting/proxying (GCP), Frontend hosting (Vercel).
DevOps	Docker, GitHub Actions	Containerization and CI/CD for streamlined deployments.
Model Hosting	Vertex AI Model Registry/Endpoints	Centralized, scalable hosting of the fine-tuned LLM.

---

🧠 AI & Data Workflow

1. Supervised Fine-Tuning (SFT)

The foundational LLM is fine-tuned to master the specific language and concepts of petroleum engineering.

• Training Data Sources:
  • volve_daily_drilling_report
  • Whitson PVT Manual
  • volve_alpaca (Used for instruction-tuning augmentation)
• Data Preprocessing:
  • Cleaned placeholder values (e.g., -999.99 converted to null or excluded).
  • De-identified sensitive or proprietary data.
  • Normalized field names and domain-specific tokens.
• Training & Evaluation:
  • Split: $80%$ Train, $10%$ Validation, $10%$ Test.
  • Training conducted on Vertex AI.
  • Achieved an initial $\approx 68%$ task accuracy on the held-out validation set.

2. Retrieval-Augmented Generation (RAG) Pipeline

The RAG pipeline provides up-to-date, precise context for the fine-tuned LLM.

• Document Processing:
  • Chunking Strategy: 300–600 tokens with 80–100 token overlap.
  • Embedding Model: all-MiniLM-L6-v2 (high-speed, effective sentence embedding).
• Retrieval:
  • Storage: FAISS index utilizing the Hierarchical Navigable Small World (HNSW) algorithm.
  • Search: Top-K semantic similarity search ($\text{K}=5$) to retrieve the most relevant chunks.
• Prompt Composition (Fixed): $$ $$$$\text{"Use ONLY the provided context to answer the question. Context: } {\text{retrieved_chunks}} \text{ Question: } {\text{user_query}}" $$ $$$$ $$
• Output: The LLM's final response is generated and displayed with source citations for verifiability.

---

🧮 Key Performance Metrics

Metric	Achieved Value	Description
Task Accuracy (SFT)	$\approx 68%$	Accuracy on the held-out validation set for the core tasks.
System Error Rate	$< 5%$	Proportion of requests resulting in a server or infrastructure error.
$\text{P90}$ Latency (Target)	$< 3.0 \text{ seconds}$	Target time for $90%$ of queries to be completed.

---

🛠️ Infrastructure & Deployment

• Frontend Deployment:
  • Deployed and managed via Vercel.
  • Utilizes React/Tailwind and communicates with the backend via REST.
• Backend Deployment:
  • Hosted on a GCP VM Instance.
  • Reverse-proxied through Nginx for security and load balancing.
  • FastAPI application served by Uvicorn and managed by Supervisor.
• Domain Configuration (Hostinger):
  • drill-sense.com $\rightarrow$ Frontend (Vercel)
  • api.drill-sense.com $\rightarrow$ Backend (GCP VM)

---

📊 Monitoring & Maintenance

• Continuous Monitoring: Metric logging handled by Prometheus for collection and Grafana for visualization.
• Automated Retraining:
  • Weekly evaluation process.
  • Triggers a full model retraining if accuracy drops $>5%$ or the hallucination rate exceeds $>5%$.
• Deployment Safety: Canary deployment ($5%$ traffic test) is used before a full-scale model rollout.
• Security: Endpoints are secured with API keys and enforced rate limits. Logs are sanitized (no PII/sensitive data) before storage.

---

🧱 Project Deliverables

• ✅ Cleaned & validated JSONL datasets for SFT.
• ✅ SFT training and validation scripts.
• ✅ FAISS embedding and index builder utility.
• ✅ Modularized FastAPI backend with all defined endpoints.
• ✅ React + Tailwind responsive frontend interface.
• ✅ Dockerized backend deployment environment.
• ✅ Integrated monitoring and auto-retrain pipeline.
• ✅ Functional prototype hosted on Hugging Face Spaces.

---

🧰 Example Usage

Backend Query Example

# Run backend locally
uvicorn app.main:app --reload

# Query the /generate endpoint
curl -X POST "http://localhost:8000/generate" \
    -H "Content-Type: application/json" \
    -d '{"query": "Summarize the daily drilling activities for 2023-07-12"}'

The frontend automatically connects to the same /api routes via the api.drill-sense.com subdomain. Users can also interact directly by pasting their drilling reports on the web interface.

---

📈 Future Work & Roadmap

• Advanced RAG: Upgrade from naive RAG to a hybrid RAG architecture incorporating techniques like vector search and cross-encoder re-ranking.
• Proactive Alerting: Integrate a classical ML model, such as LightGBM, for advanced time-series anomaly detection to augment LLM-based operational alerts.
• Feedback Loops: Implement a robust mechanism for domain feedback and human-in-the-loop validation to continuously improve the SFT model.
• Scalable Vector Store: Migrate the index deployment from local FAISS to a managed, scalable service like Milvus or Pinecone for enterprise readiness.

---

🧑‍💻 Author

Abdullah Imran AI Engineer | Data Scientist

Email: [abdullahimranarshad@gmail.com] LinkedIn: https://www.linkedin.com/in/abdullah--imran/

GitHub: https://github.com/poetabdullah

Project developed for Avions AI

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.