Yash Upadhyay

Designed LangGraph/LangChain-based agent workflows for intelligent document processing in the energy domain. Built AI pipelines with dynamic routing based on document structure, agent feedback, and tool availability. Developed OCR-centric systems to parse HT, New HT, and LT electricity bills with regex anonymization, rule-based metadata validation, and exception handling. Enabled multimodal understanding via OCR, visual analysis, and LLM reasoning. Created tool-using agents for anomaly detection, forecasting, billing error checks, and consumption insights. Integrated LLMs for semantic extraction, structured field parsing, and enrichment. Scaled multi-PDF/ZIP processing with async job handling, retry logic, and runtime safeguards. Engineered high-throughput APIs using FastAPI/Node.js with Redis caching and rate limiting. Built endpoints for energy data ingestion, solar analytics, grid diagnostics, and historical reports. Managed ETL and real-time ingestion with BigQuery and TimescaleDB. Migrated PostgreSQL to TimescaleDB for time-series IoT data. Created analytics dashboards for TOD/load/power factor trends and billing history using Matplotlib and custom charts. Built forecasting pipelines using MLflow, DagsHub, and 20+ models (ARIMA, Prophet, CatBoost, LSTM, DeepAR, TimesNet, etc.) with continual learning and outlier resilience. Developed React/Next.js frontends with TSX and modular CSS, plus mobile apps via React Native and Expo. Implemented full-stack integrations, role-based access with Firebase Auth, and in-app visualizations. Built CI pipelines using Pytest, Jest, GitHub Actions, and Docker. Enabled agentic OCR/classification/metadata workflows, LangFuse observability, and Edge-AI for low-latency grid diagnostics. Deployed on GCP (GKE, Cloud Run, GCS, Pub/Sub) via Terraform and Docker. Leveraged Vertex AI, Firestore, Cloud Scheduler, and Pub/Sub for document and forecasting workflows.

Designed systems for large-scale PDF ingestion, parallel processing, and data extraction using multi-threaded workers. Used regex parsing for field extraction, anomaly detection, and metadata tagging. Stored unstructured PDFs in GCS and structured outputs in TimescaleDB. Integrated syntactic parsers to extract tables, financial terms, and key indicators from energy documents. Built visual diff dashboards comparing OCR, regex, and LLM-enhanced extractions. Developed internal annotation tools to improve OCR accuracy and regex tuning for varied layouts. Deployed OCR models supporting PDF, PNG, TIFF, and complex formats with tables, graphs, and scans. Enabled auto classification and transformation of documents into structured datasets with minimal manual input. Integrated Firebase Auth and role-based access for mobile/web portals. Built agentic workflows for automated ingestion, transformation, and reporting based on triggers or anomalies. Developed time-series forecasting pipelines using LSTM, XGBoost, and Facebook Prophet for energy prediction. Integrated cloud-hosted APIs, models, and real-time pipelines using Google Cloud (Compute Engine, Cloud Functions, Cloud Run, Pub/Sub, Firestore). Migrated backend to GCP for IoT telemetry scalability. Built React Native + Expo mobile apps for energy monitoring, and connected React + Node.js dashboards to real-time APIs. Used Terraform for cloud provisioning and deployed with CI/CD. Ensured robust delivery through unit, integration, and end-to-end testing workflows.