Built by one engineer,
across six research phases
DiagAI is a solo research project by Omkar Thorve — built to answer one question: can a combination of LSTM anomaly detection, multi-agent LLM reasoning, and semantic knowledge graphs produce better root cause analysis than a human expert working alone? After two years and six research phases, the answer is yes.
The researcher
Omkar Thorve
Researcher · Engineer · Builder
Designed and built the entire DiagAI stack end-to-end — from data preprocessing and LSTM autoencoder training through knowledge graph ontology design, LangGraph agentic pipelines, REST API, and the production dashboard. Every phase was researched, implemented, and validated independently.
The mission
Industrial equipment should never fail without warning. DiagAI is the AI layer that makes predictive maintenance accessible to every manufacturing plant, oil platform, and utility grid — not just those with dedicated data science teams.
Technology
Anomaly Detection
LSTM Autoencoder + Random Forest ensemble (F1 0.947)
Agentic Reasoning
LangGraph 4-agent pipeline · Groq Llama 3.3 70B
Knowledge Graph
OWL/SWRL ontology · Neo4j · semantic embeddings
Backend
FastAPI · MongoDB Atlas · Python 3.11
Frontend
Next.js 14 · TypeScript · Tailwind CSS
Infrastructure
Render · Docker · REST API
Principles
Explainability over black boxes
Every anomaly score, RCA report, and maintenance recommendation includes a traceable reasoning chain. You always know why DiagAI flagged something.
Real data, real validation
Models are trained and benchmarked on established industrial datasets — AI4I 2020 and MetroPT — not synthetic toy examples.
Engineers first
DiagAI is designed for the people who actually fix machines, not for dashboards nobody reads. Every feature starts with a maintenance engineer use case.
Continuous improvement
The Learning Agent is a core system component, not a marketing claim. Feedback from every confirmed diagnosis improves future accuracy for your fleet.
Research journey
Research begins. First LSTM autoencoder trained on the AI4I 2020 dataset to detect manufacturing anomalies.
Phases 1–3 complete: feature engineering pipeline, anomaly detection system, and OWL/SWRL knowledge graph ontology.
Phase 4: Knowledge graph embeddings and cross-domain failure pattern transfer learning across industrial datasets.
Phase 5: LangGraph 4-agent RCA pipeline (Diagnostic → Reasoning → Planning → Learning) ships. 84.6% success rate on validation.
Phase 6: Full evaluation suite — ablation studies, cross-domain benchmarks, F1 lifted from 0.542 to 0.947 with ensemble scoring.
Production deployment on Render. REST API, MongoDB Atlas backend, and Next.js dashboard go live.
DiagAI platform launched publicly with full product documentation and API access.