ML Engineer at ZIEERS Systems PVT LTD (2024-08 – Present)
- Wrote a data pipeline on AWS S3 to pull raw judgment PDFs, clean and chunk them using pdfplumber, and produce instruction–output pairs in JSONL format for supervised fine-tuning.
- Set up and ran the full training loop on Kaggle using the Trainer API with cosine LR scheduling and gradient accumulation; tracked model quality using ROUGE-1/2/L, legal keyword hit rate, and citation detection rate.
- Pushed the fine-tuned model to HuggingFace Hub and wired up a Gradio demo on HuggingFace Spaces so the model could be tested live via a public URL.
- Noticed the deployed model was hallucinating on retrieval heavy queries and scoped a RAG-based fix using ChromaDB and BAAI/bge-large-en-v1.5 embeddings to ground responses in actual judgment chunks.
- Worked closely with the team across the full project lifecycle – from defining what the system should do, to scoping data needs, to shipping the final model.
Intern at ZIEERS Systems PVT LTD (2024-01 – 2024-07)
- Sourced and assembled a legal corpus from 6 sources – Supreme Court & High Court judgments from AWS S3, Constitution from cdnbbsr.s3waas.gov.in, Acts from indiacode.nic.in, Law College Texts from archive.org, and Tribunal orders from indiankanoon.org.
- Wrote scripts to convert PDFs and scanned legal documents into JSONl using pdfplumber, accounting for inconsistent layouts across different legal sources.
- Parsed the JSONl output into structured metadata, segments, and chunks to make the corpus usable for LLM training and retrieval pipelines.
