• Conducted domain research on Indian GST rules and Tally workflows to define data requirements before building a multi-stage document processing pipeline — orchestrating a modular LLM workflow (FastAPI, Django) with pdfplumber to extract fields from unstructured PDF invoices and variable-schema Excel files, normalising layouts into a canonical JSON schema.

• Deployed and prompt-engineered a self-hosted Gemma 3 (27B) LLM on an on-prem NVIDIA DGX Spark to extract and validate invoice metadata (GSTIN, HSN codes, tax heads) against GST rules, cutting per-invoice latency from 9s to 2s.

• Engineered an entity-matching stage resolving extracted parties, tax heads, and stock items against a synced master-data catalogue — applying text normalisation and similarity matching to auto-tag records, detect duplicates, and flag unmatched entries for review.

• Built a modular, responsive React and TypeScript component library adopted across 10+ production screens.

• Implemented a high-performance Image Zoom feature using Tailwind CSS, directly reducing checkout drop-off rates.

• Hardened state management with proper loading, error, and empty-state handling, maintaining a zero-defect PR record via Git.

• Built a Python data pipeline (YouTube Data API v3, BeautifulSoup, pandas) to extract and analyse weekly competitor channel metrics, driving a ranking improvement from 5th to 2nd position over 5 weeks.

• Fed competitor data into Gemini API prompts for AI-driven analysis with structured output, enabling data-driven content decisions at scale.

• Built an automation script to pull Google Form response data, auto-fill and validate exam results on the portal, cutting per-student processing time from ~30–35s to ~8–10s.