Hello, welcome to my personal page.

I am Ravi Theja, Applied AI and OSS Research engineer.

Ravi Theja D

Currently at MistralAI, previously LlamaIndex. I build LLM systems for enterprise document and knowledge workflows, and ship open models on the side. My recent work includes Navarasa (showcased at the Google I/O 2024 keynote), BRAG, and a hackathon-winning RL agent for enterprise HR.

Experience

MistralAI — Applied AI

Palo Alto, CA · Feb 2025 – Present

  • Built enterprise AI systems using Mistral LLMs, OCR, and embedding models to automate document and knowledge workflows.
  • Designed evaluation datasets from real customer workflows; surfaced OCR failure cases and worked with modeling teams to harden the models.
  • Cut invoice reconciliation time by 99% (2 hours → 1 minute) by architecting an LLM-driven document understanding pipeline with structured extraction and code-mapping.
  • Built multi-agent systems for automated PRD generation from meeting transcripts, earnings call analysis, financial reporting, and industrial knowledge retrieval. Blog · Cookbooks

LlamaIndex (Founding Engineer) — AI Engineer & Developer Advocate

Remote · Oct 2023 – Jan 2025

  • Built evaluation modules for RAG systems and introduced LLM-as-judge for retrieval quality.
  • Implemented GraphRAG, CorrectiveRAG, AdaptiveRAG, and Mixture of Agents as LlamaPacks.
  • Authored the O’Reilly course Building RAG Applications with LlamaIndex.
  • Improved LlamaCloud retrieval on complex queries via sub-query planning and metadata filtering.
  • Advised RAG solutions for ByteDance, EY, NetApp India, Albus, Atomic Works, and Videoverse.

Independent AI Research — OSS work

Bangalore / San Francisco · Jan 2024 – Present

  • Hackathon wins, open models, and benchmark contributions — see Projects.

Glance – InMobi — Senior Machine Learning Engineer

Bangalore, India · Mar 2021 – Oct 2023

  • Shipped the Glance TV Screen Saver: automated wallpaper creation with CLIP embeddings, sentence transformers for image search, and generated headlines.
  • 85% reduction in content creation time (20 min → 3 min) via an AI-assisted pipeline for Glance TV.
  • Deployed GPT-3 based paraphrasing-comment and auto-poll systems — +32% watch time, −30% editorial workload.
  • Built DropoutNet and autoencoder + ALS recommenders that lifted engagement on cold and sparse user segments.

Earlier experience

India · 2013 – 2021

  • TCS Innovation Labs — Research Engineer (2019–2021). Attention-based BiLSTM with knowledge-graph features for humor detection in edited news headlines; published at COLING-2020 (SemEval workshop).
  • Quadratic Insights — Data Scientist (2016–2017). Hierarchical text-mining + Naive Bayes to auto-route a bank’s customer-complaint emails across three department levels.
  • Hindustan Petroleum — Operations Officer (2013–2014). Validated ML sales-forecasting models and built features for oil-product distribution planning.

Education

M.S. Computer Science, IIIT Bangalore — GPA 3.77 / 4.0 · 2017–2019

B.Tech. Electrical & Electronics, NIT Warangal — GPA 8.14 / 10.0 · 2009–2013

Projects & OSS

Meta OpenEnv Hackathon SF, Winner

RL · GRPO · Llama 3.2-1B

An OpenEnv-compatible RL environment simulating enterprise HR onboarding across 6 apps, 25 tools, and 77 tasks. GRPO training on Llama 3.2-1B with a rubric reward delivered +67% mean task score and +162% on complex multi-step tasks.

Code

Navarasa 2.0

Gemma-7B / 2B · featured at Google I/O 2024

Gemma finetuned for 15 Indian languages. Featured in Google’s Gemmaverse, showcased at the Google I/O 2024 keynote, and ranked top-6 for Indian languages by Microsoft Research.

Gemmaverse · Blog · Models · Code

BRAG, Small Language Models for RAG

LoRA / QLoRA · <$25 per model

A series of small models tuned for retrieval-augmented generation. Outperforms Cohere Command R+, Qwen2, and Llama 3.1, and closely matches GPT-4-Turbo on ChatRAG-Bench. Each model trained for under $25.

Blog · Models

MMMU benchmark in SGLang

VLM evaluation

Integrated the MMMU benchmark into SGLang so vision-language models can be evaluated with a standardized harness.

PR

Automatic Knowledge Transfer, Hackathon Winner

LlamaIndex · D-ID

Summarizes codebases and generates explainer videos to cut onboarding effort for large engineering teams. Also presented at PyCon India 2023.

Blog · Code · PyCon talk

Publications

Summarizing Short Medical Conversations

ACL 2023 · MEDIQA-Chat workshop

Research paper on generating concise summaries of short doctor–patient conversations, contributing to the MEDIQA-Chat shared task on clinical dialogue summarization.

Paper

Humor Recognition in Edited News Headlines

COLING 2020 · SemEval-2020 workshop

Attention-based BiLSTM architecture combining knowledge-graph and lexical features to detect humor in edited news headlines.

Paper

Achievements & Talks

Skills

Python · LLMs · GRPO / RL fine-tuning · RAG · LlamaIndex · Multi-agent systems · LLM evaluation · NLP · SQL · Deep learning