Open to research collaborations and AI/ML roles

I teach machines to read messy, multilingual documents.

I'm Radha Krishna, an MS by Research student and engineer at CVIT, IIIT Hyderabad. I work on Indic OCR, cross-lingual retrieval and RAG for historical archives.

See my work Download résumé Get in touch
2papers: ICVGIP '25, CVIP '26
200,000archive pages in my RAG study
3languages: Hindi, English, Punjabi
Scanned page, line 14

Making archives searchable, in the languages they were written in

I'm pursuing an MS by Research in Computer Science and Engineering at IIIT Hyderabad, focused on OCR, document layout understanding and cross-lingual information retrieval.

At CVIT I fine-tune CRNN-based OCR models, build high-throughput pipelines that process thousands of document images, and design Elasticsearch systems that retrieve from noisy OCR output.

My latest work combines layout segmentation, multi-script OCR and agentic RAG to answer questions over the 200,000-page Punjab Vidhan Sabha archive.

MS by Research, CSEIIIT Hyderabad · July 2025 – present
B.Tech, DTDP (IT specialization)Jawaharlal Nehru Architecture and Fine Arts University, Hyderabad · 2019 – 2023

Where I want to take this next

Documents are where I started, but the tools I use (vision models, representation learning, careful evaluation on noisy data) apply well beyond them. I'm looking for interdisciplinary research where machine learning meets the physical and life sciences.

Medical imaging

Robust vision models for scans where labels are scarce and noise is real, which is the same problem as my label-free OCR work.

Vision-language-action models

Systems that see, reason in language and act, with a focus on grounding and reliability.

Physics-informed machine learning

Building known physical structure into models instead of learning everything from data.

Fluid dynamics

Learned and hybrid methods for simulating and analysing flows.

Cosmology and astronomical imaging

I've already taken a first step here by detecting gravitational lensing arcs and Einstein rings in telescope images.

Computational science

Fast, scalable numerical and data pipelines, informed by my systems work in Rust and large-scale OCR.

Experience

Project Technical Personnel (L1), Junior Engineer

Centre for Visual Information Technology, IIIT Hyderabad
  • Fine-tuned CRNN-based OCR models on domain-specific data, improving accuracy on diverse and noisy inputs.
  • Built high-throughput pipelines that run OCR over thousands of document images.
  • Designed Elasticsearch search with custom analyzers and schema tuning for OCR output.
  • Cut inference latency with architecture-level optimizations and batch processing.
More responsibilities
  • Automated data preprocessing, training, evaluation and deployment workflows.
  • Integrated OCR systems and search backends into research applications used by other teams.
  • Maintained technical documentation and modular codebases for reproducibility and handoff.

Junior Engineer

Centre for Visual Information Technology, IIIT Hyderabad
  • Wrote web scraping scripts that collected and organized large datasets for model training.
  • Built and fine-tuned ML models, adapting pre-trained models to specific problems.

Subject Matter Expert (part-time, remote)

Chegg India
  • Helped students with CSE coursework, including open-source software and computer networks.

Projects

Traditional search fails on these archives because of OCR noise, language mismatch and missing context. CLiGRA combines lexicon-based search with LLM reasoning: it finds documents by keyword and semantic similarity, then answers in the language of the query, grounded in what it retrieved.

PythonFAISSVector databasesDjango RESTNLTK
Project page

A database projection operator with adaptive column reordering and four execution strategies (eager, streaming-batched, parallel-hybrid and columnar-staging). On billion-row workloads it ran about 10× faster with 256× lower memory use.

RustRayonSIMD
View on GitHub

A CNN trained on the uTHCD dataset to recognise 156 Tamil character classes. The Streamlit app lets you draw or upload a character and returns the Unicode prediction with confidence scores, plus a practice mode.

PythonOpenCVNumPyscikit-image
View on GitHub

An automated pipeline that finds Einstein rings and arc-like lensing structures in noisy, low-brightness astronomical images, using background subtraction, edge detection, the Hough Transform and contour-based geometry.

PythonOpenCVHough Transform
View on GitHub

Publications

CVIP 2026

Reading Beyond OCR Noise: Robust Cross-Lingual RAG for Legislative Archives

An end-to-end framework combining vision-based layout segmentation, multi-script OCR and agentic RAG to answer questions over degraded historical parliamentary documents. Evaluated on the 200,000-page Punjab Vidhan Sabha archive.

Read the paperProject page
ICVGIP 2025

Label-Free Adaptation of Indic Printed OCR

A self-supervised adaptation method that improves OCR for low-resource Indic scripts without annotated training data, using domain adaptation to lift recognition accuracy on multilingual printed text.

Read the paper

Skills

Machine learning and vision
PyTorchTensorFlowOpenCVCRNN / OCRFine-tuningDeep learning
Languages
PythonRustCC++JavaSQLHTML / CSS
Search and data
ElasticsearchFAISSVector databasesMongoDBMySQL
Backend and tools
Django RESTFlaskDockerAWSGit / GitHub
Core CS
Data structures and algorithmsComputer networksDBMSOperating systems
Ways of working
Project managementCommunicationLeadershipProblem solving
Pre-Regional Math Olympiad

Qualified during the second year of Intermediate studies.

Have a hard document problem?

I'm open to research collaborations, especially interdisciplinary ones in medical imaging, physics and cosmology, as well as full-time AI/ML engineering roles. Email is the fastest way to reach me.

Download résumé Write an email LinkedIn GitHub