Rexon PambujyaRexonPambujya

Jul 2025 – Jan 2026/Mumbai

Meeting-to-BRD generation platform, insurance QA fine-tuning, and chemical blend production forecasting.

• Designed and deployed an LLM-powered pipeline turning recorded meetings into client-ready BRD documents, cutting manual documentation effort by 90% across 8+ stakeholders.
• Built the transcription layer end to end — extracting audio from MP4/MP3, chunking long recordings and calling Azure OpenAI Speech in parallel threads so multi-hour meetings transcribe reliably.
• Added a review chatbot that lets stakeholders question and refine requirements, feeding the conversation back into the generator so the BRD updates iteratively.
• Shipped it as Flask REST APIs behind JWT auth, with per-session artefacts across Google Cloud Storage, Firestore and Secret Manager, and automated export to Word.
• Fine-tuned Gemma 2 2B on 10k+ SBI General Insurance records using LoRA adapters and 4-bit quantisation, building the pipeline that turned 1,000+ policy PDFs into provenance-tracked QA pairs.
• Evaluated the fine-tune on BERT-F1, ROUGE, semantic similarity and faithfulness, serving it from FastAPI and Docker endpoints with adapter hot-swap.
• Built a time-series forecasting pipeline for SABIC's ethanol blend production datasets in RapidMiner, improving forecast reliability and reducing analysis effort.

• Python
• Azure OpenAI
• LoRA
• PyTorch
• Flask
• FastAPI
• Google Cloud
• RapidMiner
• Docker
• Jenkins

Feb 2024 – Mar 2025/Mumbai

BMC Salary Program — replaced a manual, file-based salary process for Mumbai's municipal schools with an online system.

• Automated salary processing for 1400+ BMC schools, cutting manual workload and increasing efficiency by 80%.
• Integrated AI-driven microservices running a lightweight ML model that flags payroll outliers against historical data, catching anomalies before disbursement.
• Built backend services in Python and integrated REST APIs to move salary data reliably between systems.
• Developed frontend modules in React.js, Next.js, HTML and CSS, improving navigation and the day-to-day experience for school administrators.
• Set up continuous integration workflows so data deployments shipped reliably instead of by hand.

• Python
• REST APIs
• Microservices
• Anomaly Detection
• React
• Next.js
• CI/CD
• HTML
• CSS

Oct 2023 – Jan 2024/Mumbai

HSN Recommender System and EwayBill Module.

• Prepared training data for the machine learning model by annotating essential attributes across 250+ documents, then fine-tuned OpenAI's GPT-3.5 on the HSN dataset.
• Trained a model on HSN documents using the LangChain framework, improving HSN code prediction accuracy and streamlining code assignment.
• Created an EwayBill module in HTML, CSS and Angular to automate filling out E-waybill forms, reducing manual workload.

• OpenAI GPT-3.5
• LangChain
• Angular
• HTML
• CSS

Jul 2022 – Jul 2023/Mumbai

Route optimisation for logistics — clustering delivery locations to cut transport cost and speed up delivery.

• Implemented a novel clustering approach validated with the Travelling Salesman Problem, grouping geographic coordinates into single-route clusters and cutting transportation costs by 28%.
• Framed the business problem for the logistics operation, then analysed delivery data and reworked the process so routes could be measured against a clear efficiency target.
• Structured raw Indian address data and enriched it with Google Maps API geodata, adding three spatial features that made coordinate-level analysis possible.
• Identified clusters of customer locations serviceable by a single route from the cluster centroid, reducing average distance travelled per route.
• Built an interactive Folium and Streamlit dashboard with key metrics and filters, halving the time taken to draw location-based insights.

• Python
• scikit-learn
• SciPy
• NumPy
• Pandas
• Folium
• Streamlit
• Flask
• Geospatial Data