Data Platform Engineer @ BharatGen — I build the data that makes LLMs work, and the platform behind it.

I build high-throughput, cost-efficient data platforms for LLMs — from petabyte-scale ingestion to governed metadata, evaluation harnesses, and agentic delivery. Currently leading data infrastructure for India's Sovereign AI at BharatGen.

• Chitrakshara: A Large Multilingual Multimodal Dataset for Indian Languages — CVPR 2025 (VLMs4All Workshop). 193M images, 30B tokens. HuggingFace
• AyurParam: A State-of-the-Art Bilingual Language Model for Ayurveda — ArXiv 2511.02374 · HuggingFace
• BhashaBench V1: A Comprehensive Benchmark for the Quadrant of Indic Domains — ArXiv 2510.25409

• Delivered a 1T+ token pre-training corpus for BharatGen's foundation models — Common Crawl WARC, 11M books (Internet Archive), 30M+ open-access papers, 7M GitHub repos.
• Co-authored 3 papers: Chitrakshara (CVPR 2025), AyurParam, BhashaBench V1.
• ~10M Q&A synthetic-data pipeline that shipped AyurParam, the highest-SOTA Ayurveda LLM.
• BhashaBench — Indic LLM evaluation benchmark, adopted as BharatGen's standard eval framework.
• Cut crawling infrastructure 87.5% (40 → 5 machines) with zero throughput loss.
• Rewrote the tokenizer in Rust for a 3× SLURM-native training speedup.

Desktop app (Electron + React 19 + TypeScript) that turns raw talking-head footage into high-retention edutainment shorts. Uses Gemini 2.5 for multimodal transcription and GSAP animation generation, with an FFmpeg render pipeline. Solo build, end-to-end.

BharatGen's live AI product for farmers. RAG-optimized agent with tool use: location-aware search, government policy lookup, and form-filling workflows. Curated the knowledge base and shaped the agent's retrieval and tool layer.

Data engineering interview prep — PySpark notebooks, theory docs, quizzes, and company-specific patterns. Built around Zephyr Coffee Co., a fictional 200-store chain with messy data. Community-adopted.

JuiceFS + DuckDB + DataHub lineage across 1M+ datasets. Reproducible pipelines for LLM training and evaluation. Chosen over S3/Athena/Glue for cost efficiency at petabyte scale.

Applying Tiago Forte's PARA method as a gamified personal knowledge system. Side project, TypeScript.

• Pretraining Data Engineering — corpus curation at PB scale, dedup, contamination filtering, quality gating (Common Crawl WARC, Archive.org, ArXiv, GitHub).
• Synthetic Data, Evaluation & Harnesses — context-grounded Q&A generation, benchmark design, LLM-judge pipelines, and evaluation harnesses for standardized model scoring.
• Agent Building — RAG-optimized agents with tool use (location search, policy lookup, form filling); production advisory agents for real-world advisory workflows. Agent Harness for complex task.
• Model Serving & Inference — vLLM + Ray multi-node (1 → 50 nodes), ONNX optimization, Rust tokenizer.
• Data Platforms & Governance — on-prem lakehouse (JuiceFS + DuckDB + DataHub), Great Expectations, dbt tests, custom Indic-script validators.

• Languages: Python (PyTorch, PySpark, Pandas), Rust, SQL, Bash, JavaScript
• Data / Compute: Spark, Kafka, Debezium (CDC), Airflow, DataHub, DuckDB, Trino/Presto, Scrapy, JuiceFS
• AI / ML: vLLM + Ray, SGLang, Essential AI, ONNX, Surya OCR, Vertex AI, Great Expectations
• Cloud / Infra: AWS (S3, EMR, Glue, Redshift, EC2), Azure (Blob, Functions, Databricks, SQL), Docker, SLURM, GitHub Actions

• AI / Data Instructor @ upGrad — applied data engineering and LLMs.
• Head Organizer, Kafka Mumbai Community — meetups, 10+ talks, technical discussions.
• Maintainer, dataengineer_prep — 73★ open-source DE learning repo.

• Harvard CS50 — score 9.9
• 2nd Rank in Diploma — Rajokari Institute of Technology
• Extension Mania (IIT Madras) — Winner
• AWS Cloud Practitioner, Azure Data Fundamentals, SQL (Intermediate) certified

📍 Mumbai, India 🎓 IIT Madras Online — Foundation Certificate, Programming & Data Science (2022–2023) 🎓 Rajokari Institute of Technology — Diploma in Computer Science (2018–2021)

• 💼 LinkedIn