Our Mission
Bringing structure to the industrial world's dark data.
Mining operations sit on decades of unstructured intelligence — paper logs, PDFs, scattered reports. We exist to turn that dead archive into a live, queryable brain.
Two builds, four years apart.
Mining documentation is mostly tables and scans. Standard text extraction flattens a ventilation table into a run of numbers that reads like prose — not merely lossy, but confidently wrong once it has been embedded, retrieved and cited. General assistants make it worse by inventing regulatory citations that look exactly like real ones.
The first version was built for Smart India Hackathon 2023 against the Ministry of Coal problem statement, by a team, and won at the national level. CMPDI officials who judged the finals opened discussions about deploying it at scale. Those talks did not proceed.
This platform is a separate, ground-up rebuild started in June 2025 and developed solo since — hybrid retrieval over your own documents, five specialized agents, and a rule the system is held to: every claim carries the document and page it came from, and an answer that is not in the retrieved context is refused rather than guessed.
Engineering Principles
How we build the intelligence layer.
Answers You Can Check
Every claim carries the document and page it came from, and the model is instructed to refuse rather than answer past its retrieved context. A near-miss is reported as a near-miss.
Domain Precision
No model is fine-tuned on mining data. Precision comes from retrieval: hybrid vector and keyword search over your own documents, narrowed by a cross-encoder reranker before the model ever sees a passage.
Built for Awkward Documents
Mining documentation is mostly tables and scans. Extraction is layout-aware, tables survive as tables, and pages that return no text fall back to OCR.
Augment, Not Replace
We capture the institutional knowledge of retiring senior engineers and make it instantly accessible to the next generation.
Read the source
MiningNiti is open source and MIT licensed. The repository documents the architecture, the retrieval evaluation results, and a candid list of what is not finished yet.
github.com/Iammilansoni/MiningNiti