Tushir Sahu
@TushirSahu
MEXT Scholar · MS CS @ Osaka University · AI Researcher
Building intelligent systems — from computer vision and LLMs to production ML pipelines. Turning research into products.
Building AI That Delivers
Bridging the gap between research and production-ready AI systems.
// bio
I'm currently an MS (CS) student at Osaka University, supported by the prestigious MEXT Scholarship — one of Japan's most competitive government fellowships. My research spans Medical Imaging, LLMs, MLOps, and Computer Vision, with a focus on self-supervised and unsupervised learning in real-world settings.
Most recently an R&D Computer Vision Intern at Omron (Kyoto), with prior work at GE Vernova and VRAIN Solution. B.Tech (ECE, CGPA 8.4) from IIIT Jabalpur. My paper DOSE was accepted at WACV'27, adding to publications at CNIOT'24 (Tokyo), ISES'24, and APSCON'24. Actively looking to contribute to teams pushing AI out of the lab and into production.
University of Osaka
MS — Computer Science · Apr 2026 – Mar 2028
IIIT Jabalpur
B.Tech — ECE · Jul 2021 – Jul 2025
// tech stack
// publications
Publications
Hover to fan out ↑
// location
University of Osaka
Projects Built
Months Experience
AI Models Trained
CGPA (B.Tech)
Where I've Worked
Research, internships, and production AI — from academia to industry.
R&D Computer Vision Intern
- Reduced AXI over-detection by 11.6% by integrating a lightweight bottleneck architecture into the baseline model.
- Evaluated 8–9 adapter-layer variants through ablation studies to balance feature preservation and computational efficiency.
- Adapted model domain distributions to significantly improve anomaly-detection accuracy on NG solder-ball inspection.
Digital Technology Intern — MLOps
- Contributed to the Model Verification Manager (MVM) platform, raising unit test coverage to 97% across multiple modules.
- Designed a pipeline-level ExitHandler to ensure consistent failure detection of pods across the MLOps workflow, capturing and logging critical errors to reduce debugging time.
- Improved error traceability and observability through structured and semantic logging practices.
AI Advanced Researcher
- Created a fine-tuned prompt encoder for MobileSAMv2 to improve segmentation performance of VRAIN's product.
- Developed a targeted training script incorporating the fine-tuned prompt encoder into MobileSAMv2.
- Replaced YOLOv8-seg with a Multi-Head Attention mechanism and FCN for dynamic box generation and refinement.
Computer Vision Intern
- Led development of a multi-threaded video processing pipeline for real-time traffic video analysis.
- Designed a tailored OCR model optimized for recognizing text on low-resolution 2-wheeler number plate streams.
- Fine-tuned models to operate efficiently on continuous data streams, minimizing bandwidth and latency.
Research Intern — Computer Vision
- Proposed SPEED-Net, a lightweight modified UNet with residual pyramidal feature extraction for colonoscopy images.
- Implemented Dilated Involutions to reduce boundary loss while keeping the parameter count in check.
- Achieved DSC of 0.953 and IoU of 0.872 — published at ISES'24.
Deep Learning Research Intern
- Developed HDA-Net, a hierarchical attention model for diabetic retinopathy grading on 85K fundus images — 82.45% accuracy.
- Proposed SHAB (Spatial Hierarchical Attention) and CHAB (Channel Hierarchical Attention) blocks.
- Employed inter-hierarchical feedback connections between SHAB and CHAB to improve discriminative power.
What I've Built
AI-powered products, research tools, and full-stack systems — from lab to production.
Automated LLMOps pipeline that detects semantic drift in technical documentation and auto-reingests stale content — keeping RAG knowledge bases fresh without manual intervention.
- Agentic tool-calling loop over hybrid retrieval, hardened with guardrails for prompt-injection, system-prompt leakage, and citation verification
- CI quality gate scores auto-generated eval sets with Ragas against a committed baseline — failing the build on measurable regression
- Qdrant vector store with Ollama-served local LLMs; drift triggers automated re-ingestion and re-evaluation
Corrective RAG state-machine built with LangGraph: retrieval is graded, corrected, and falls back to autonomous web search when the knowledge base can't answer — deployed as a containerized LLMOps service.
- Custom hybrid search with Reciprocal Rank Fusion and self-grading retrieval evaluator
- Fine-tuned HuggingFace embeddings via contrastive learning on synthetic data
- Containerized LLMOps: FastAPI backend + Streamlit UI with LangSmith tracing and CI/CD
Production-grade MLOps system for manufacturing: edge inference on factory floors with continuous monitoring, drift-triggered retraining, and automated deployment — the full ML engineering lifecycle, not a notebook.
- Kafka edge streaming → ZenML pipelines → MLflow registry with strict evaluation gates and automatic rollback
- Active-learning loop: low-confidence predictions routed to human review, then back into retraining
- NVIDIA Triton edge serving, FastAPI inference API, Terraform/Kubernetes IaC, GitHub Actions CI/CD
Salient Pyramidal Enhancement Encoder-Decoder Network for colonoscopy image segmentation — a modified UNet with residual pyramidal feature extraction and Dilated Involutions. Published at IEEE ISES'24.
- DSC 0.953 and IoU 0.872 on colonoscopy segmentation benchmarks
- Dilated Involution blocks capture long-range context at lower parameter cost than convolution
Hierarchical Dual-Attention Network for diabetic retinopathy detection and severity grading. Channel and spatial hierarchical attention modules highlight clinically relevant retinal regions, adapting hierarchical ideas from NLP to medical imaging.
- Dual channel + spatial attention improves lesion localization over baseline CNNs
- Grades DR severity into clinical stages to support treatment decisions
Real-time pedestrian safety system on live Shibuya crossing footage: detects people outside crosswalk polygons and raises alerts.
- SAHI sliced inference refines small-object bounding boxes in dense crowd scenes
- Custom polygon regions for precise out-of-crosswalk detection on live video
AI-based applicant filtering system that analyzes resumes and ranks candidates against job descriptions using NLP and classical ML.
- CountVectorizer + cosine similarity ranking served by a Flask ML API
- Full-stack delivery: Node.js/Express backend with a Vite + React frontend
Personalized movie and series recommendation platform with a content-based ML engine served as a microservice.
- Flask microservice serves the recommendation model behind an Express API
- React/Vite frontend with content-based filtering via cosine similarity
Tech Stack
The domains I research and the tools I build with.
Domains of Interest
Languages
8Frameworks & Libraries
5Peer-Reviewed Research
From top-tier venues like WACV to IEEE, ACM, and international workshops — spanning computer vision, medical imaging, and IoT.
Papers
Venues
A* (WACV'27)
DOSE: Dominance-Guided Ordinal Severity Estimation from Sparse Relative Annotations
Location based portable 5G enabled Internet of Things IoT EDGE device to measure, monitor and manage environment sensor data
SPEEDNet: Salient Pyramidal Enhancement Encoder-Decoder Network for Colonoscopy Images
ConstructNet: A Deep Learning Object Detector for Construction Site Surveillance
Let's Build Something
Open to research collaborations, internships, and full-time roles in AI / Computer Vision.