Harsh Soni

My account-opening automation for a regulated bank runs in production, and a GenAI triage layer for credit-card service requests is in UAT. I also work on Kyndryl’s internal multi-agent platform: orchestration, task planning and pluggable tools.

• Paper forms → clean records for a regulated bank. An end-to-end OCR pipeline on AWS Textract that tells form types apart, extracts the right fields for each, validates and encrypts sensitive data, then pushes clean records into downstream systems. Taken from a working PoC to the production process the bank now runs on.
• GenAI triage for credit-card service requests. A triage layer on AWS Bedrock that classifies each request against the bank's taxonomy, splits tickets that bundle several issues, and extracts structured details. Runs as an unattended daily job.
• Internal multi-agent platform. A shared platform that standardises how AI agents get built and deployed across client engagements; I contribute to it.
• Regulated-environment platforms & accelerators. Multilingual document pipelines (OCR, translation, structured extraction) on on-prem and cloud platforms for regulated environments, GPU inference with vLLM and quantization, packaged as reusable frameworks and accelerators. Technical delivery lead and mentor on the team.

I built demand forecasting for 23,000+ SKUs in 9 global regions; it raised forecast accuracy 30–35% and added USD 3M in sales. Then GenAI tools for regulatory dossiers and text-to-SQL. As technical lead, I mentored three data scientists and interns.

• Global SKU-level demand forecasting. Forecasting for 23,000+ SKUs across 9 global regions with 30+ models (AutoARIMA, SARIMAX, XGBoost, Prophet). Pipelines from SAP BW through Azure Data Factory and Delta Lake into Databricks, with MLflow tracking, CI/CD and automated retraining, plus an LLM layer that explains forecasts in plain language. Result: USD 3M incremental sales · USD 1.3M saved in inventory costs · +30–35% forecast accuracy.
• GenAI regulatory dossier automation. A Generative AI platform that automates regulatory dossier creation in crop protection. OCR, Azure OpenAI, LangChain, HuggingFace and FastAPI behind Azure SSO. Result: USD 23K saved in manual processing.
• LLM Text-to-SQL. Business users ask enterprise data questions in plain English. Semantic search finds the right tables, the LLM writes the SQL, and the answer comes back summarised.
• Customer churn model. A Random Forest model on SAP sales data that flags customer segments at risk of churning, and sized a potential 8–12% revenue uplift for retention.

I built an NLP model with semantic search in R, 88% accurate, and a Shiny dashboard HR managers used to explore survey themes. I later extended it into my TISS dissertation.

I analysed COVID-19’s social, economic and psychological impact on beneficiaries of the Uday programme, and built the dashboard and recommendations from it.

Dissertation: an NLP system for HR survey analytics, extended from my HPCL internship.