Mithun Shivakoti

Gilbert, United States

• ▸Deployed a DistilBERT-based NLP model for voice AI classification achieving 98% accuracy with less than 100 ms inference latency in production.
• ▸Integrated Human-in-the-Loop validation pipelines using Argilla and Label Studio, reducing model iteration cycles and improving annotation throughput.
• ▸Redesigned backend architecture by extracting user operations into a dedicated microservice with secure TCP communication and database isolation, enabling independent scaling and improved maintainability of the core ML platform.

Hyderabad, India

• ▸Built an OCR model to automate captcha solving, scaling throughput from 70,000 to 700,000 daily.
• ▸Engineered an adverse media classification module (96% accuracy), improving corporate risk analysis speed.
• ▸Automated summarization of 30+ page reports into 1-page briefs, cutting analysis time by 90%.
• ▸Built and deployed a RAG system integrating Gemini 1.5 (GCP) and GPT-4 with vector-based document retrieval, delivering regulatory compliance assistance to 50,000+ businesses.

Bengaluru, India

• ▸Developed regression-based optimization pipeline (GPR + Ridge + ElasticNet) reducing signal loss to below 1%.
• ▸Designed modular experimentation pipelines improving reproducibility and accelerating model validation.

Colombo, Sri Lanka

• ▸Developed and evaluated supervised classification models for early autism risk detection using structured clinical data, achieving over 96% validation accuracy through feature engineering and cross-validation.