वास्तवै चल्ने उत्पादनस्तरको एमएल प्रणाली बनाउँदै।

Four layers before the LLM: the gatekeeper pattern (in English)

How a stack of cheap classifiers cuts a document-IDP bill by an order of magnitude without ever waking the model.

The cheapest token is the one you don't send (in English)

Routing, retrieval, and a small set of decisions about which questions a model should never see.

Production ML is mostly not ML (in English)

A field report from a three-agent IDP pipeline: what breaks first, what's worth automating, and where humans still earn their seat.

A three-agent IDP pipeline on Amazon Bedrock with a four-layer gatekeeper in front. Filters 92% of input before any LLM token is spent, routes the rest by intent, and closes the loop on human corrections.

• · Python
• · AWS Bedrock
• · EfficientNet-B0
• · OpenCV
• · MCP
• · Neo4j
• · EventBridge
• · Amazon A2I

A vision pipeline for live surface-defect detection on machined parts, published at NAMRC/MSEC 2025. Combined classical image features with a small CNN to keep inference under 30 ms on workshop hardware.

• · Python
• · PyTorch
• · OpenCV
• · ONNX Runtime
• · Jetson Nano