RBG.AI

Corporate at Campus! We are proud to share that our team member, Jairam Rajasekar, recently conducted a Knowledge Sharing Workshop on "AI in Speech Processing" at CSI College of Engineering, attended by over 40 students from the 2nd- and 3rd-year AI/ML and Cybersecurity programmes. The session covered the foundations and real-world applications of AI in speech, an area that sits at the intersection of machine learning, linguistics, and human-computer interaction, and one that is increasingly relevant across industries. At RBG.AI, we believe that advancing AI goes beyond building products; it also means making AI for Everyone possible by nurturing the next generation of engineers and researchers. Initiatives like this reflect our commitment to bridging the gap between academia and industry. Thank you to CSI College of Engineering for the opportunity, and to every student who attended and engaged with such curiosity and enthusiasm. #RBGai #CorporateatCampus #KnowledgeSharingWorkshop #MakingAIforEveryonePossible #CorporateAtCampus #AI4Everyone #AIEducation #SpeechProcessing #AIWorkshop #KnowledgeSharing #FutureOfAI

Exciting News from RBG.AI! We are proud to share our latest contribution from the RBG.AI Research Team: "AURA-ST: Acoustic-Unconstrained Residual Architecture for Speech Translation," published at IWSLT 2026 (23rd International Conference on Spoken Language Translation), one of the most prestigious venues in spoken language research. The Problem: Speech translation for low-resource African languages (Hausa, Igbo, Yoruba) is an extreme challenge, marked by severe data scarcity, high tonal and dialectal variance, and the absence of large pretrained acoustic models for these languages. Our Approach: • Dual-Stream Acoustic Encoder: A jointly trained semantic and paralinguistic encoder that captures both phonetic content and tonal cues critical for languages where pitch alone distinguishes meaning. • Attention-Free Prefix Architecture: Instead of resource-heavy cross-attention, AURA-ST treats projected audio as a native token prefix to a frozen LLM, enabling scalable, memory-efficient speech-to-text translation. • MLP-Targeted LoRA: We identified and resolved a structural incompatibility between standard PEFT adapters and the Gemma-4 Per-Layer Embedding architecture, enabling effective fine-tuning with less than 0.65% of the model's total parameters. Key Takeaway: AURA-ST demonstrates that modular, attention-free architectures paired with targeted parameter-efficient fine-tuning can unlock strong speech translation performance for low-resource languages without the computational overhead of traditional end-to-end systems. The PLE-LoRA incompatibility finding also has broader implications for anyone working with PEFT on Gemma-4 variants. 📖 Read the full paper here: https://lnkd.in/d98Y7q2s Congratulations, Jairam Rajasekar and Reshma Unnikrishnan. Your work is a reflection of the research culture we strive to foster at RBG.AI. We look forward to the continued impact of this contribution. We'd also like to extend our sincere gratitude to Michal Ptaszynski and Barathi Ganesh Hullathy Balakrishnan, our collaborators at the Kitami Institute of Technology, Japan, for their contributions to this research. #AIResearch #SpeechTranslation #NaturalLanguageProcessing #LowResourceNLP #IWSLT2026 #MetaLearning #LargeLanguageModels #RBGAI #AfricanLanguages #Innovation

And that's a wrap on another amazing batch of summer interns at RBG.AI! 🎉 Over the past few weeks, we have had the pleasure of welcoming 10 brilliant minds from Sri Ramakrishna Engineering College into the RBG.AI family, and what a ride it's been! 🚀 From diving headfirst into real-world challenges to shipping work that actually mattered, this batch didn't just intern; they contributed. Each one of them brought fresh perspectives, relentless curiosity, and more energy than we could keep up with. 😄 To our 10 rockstars: you came in as students, and you're leaving as builders. We hope RBG.AI gave you a glimpse of what it looks like when passion meets purpose, and we can't wait to see what you go on to create. 🌟 Wherever you go next, know that you'll always be part of the RBG.AI story. 🚀 A special thanks to our Incubators and Enablers, SREC SPARK INCUBATION FOUNDATION and Sri Ramakrishna Engineering College #Internship #RBGai #NextGenAI #FutureBuilders #AIStartup #InternLife #YoungMinds #GrowthMindset

🏆 Exciting News from RBG.AI! We are proud to announce that PAIRNotes has been selected as a Semifinalist at the Nebius AI Discovery Awards 2026, in the Medical Imaging category! Out of 646 global submissions, only 198 advanced to the semifinal. We are honoured to be among them. Radiology workflows are burdened by the time and cognitive load of producing structured, accurate reports at scale; a bottleneck that directly impacts clinical throughput and diagnostic quality. Our Solution: PAIRNotes (https://pairnotes.rbg.ai/) • AI-Driven Traiging and Report Drafting: PAIRNotes is an intelligent medical triaging and radiology report drafting engine that automates the generation of structured, accurate reports in real time. • Workflow Integration: Designed to integrate seamlessly into existing radiology pipelines, eliminating manual documentation overhead. • Scalable Precision: Built to handle high volumes without compromising report quality, empowering radiologists to focus on what matters most: patient outcomes. PAIRNotes addresses the single greatest bottleneck in modern radiology not by replacing the radiologist, but by removing the documentation burden that slows them down. This recognition validates our belief that AI-driven automation can meaningfully transform clinical workflows. See the semifinalists: https://lnkd.in/ewSkvHUe A special thanks to our Incubator SREC SPARK INCUBATION FOUNDATION. #AIResearch #MedicalImaging #Radiology #HealthcareAI #PAIRNotes #RBGAI #Nebius #AIDiscoveryAwards #Innovation #HealthTech

Proud News from RBG AI Research! 🎉 We are delighted to congratulate our team member, Jairam Rajasekar, on the publication of his research article, "SegTM-SGD: Few-Shot Emotion Classification Framework Using Meta-Learning Paradigm", published in the prestigious journal Computer Speech & Language. This work was co-authored with Dr Jyothish Lal G, Premjith B, Bharathi Raja Chakravarthi and Barathi Ganesh Hullathy Balakrishnan. Emotion recognition in low-resource settings remains a significant challenge in NLP, where labelled data is scarce and traditional supervised approaches struggle to generalise across languages and domains. Our Approach: • Meta-Learning Framework (SegTM-SGD): A few-shot learning paradigm that enables the model to generalise effectively from very limited labelled examples. • Low-Resource Emotion Recognition: Tackles the core challenge of building robust emotion classifiers without large annotated datasets. • Intelligent Language Understanding: Advances the broader goal of making NLP systems more context-aware and emotionally intelligent. Key Takeaway: This research demonstrates that meta-learning is a powerful alternative to data-hungry supervised methods, offering a practical, scalable path forward for emotion classification in real-world, resource-constrained environments. Read the full paper here: https://lnkd.in/g28EpX8G Congratulations, Jairam Rajasekar and Barathi Ganesh Hullathy Balakrishnan! 🎊 Your work is a reflection of the research culture we strive to foster at RBG.AI. We look forward to the continued impact of this contribution. 👏 #AIResearch #NaturalLanguageProcessing #MetaLearning #FewShotLearning #EmotionRecognition #MachineLearning #ResearchExcellence #RBGai

Just wrapped up three incredible days at FITT, IIT Delhi, participating in Stage 2 of the AI Grand Challenge 2025 (7-9 May 2026). Our team worked on PS-6: Language Agnostic Speaker Identification & Verification, Diarisation, and subsequent Transcription & Translation. In plain terms build an AI that can: → Identify and verify who is speaking → Separate multiple speakers in a conversation (diarisation) → Do all of this regardless of language or dialect → Then transcribe and translate the speech accurately It's a genuinely hard problem. And one that matters enormously in a country with 22 scheduled languages and hundreds of dialects, where the gap between what people speak and what AI can reliably process is still very real. Stage 2 was a full evaluation rigorous, demanding, and eye-opening. Results are awaited, but irrespective of the outcome, the experience of building this system, defending it in front of expert evaluators, and being in a room full of teams solving problems of national importance that's something we'll carry forward. Grateful to the organisers at FITT, IIT Delhi for putting together such a well-structured challenge. And to our team the effort we put in was real, and we are proud of what we built together. hashtag #FITT #SRECSPARK #RBG.AI #IITD

Just wrapped up three incredible days at FITT, IIT Delhi, participating in Stage 2 of the AI Grand Challenge 2025 (7-9 May 2026). Our team worked on PS-6: Language Agnostic Speaker Identification & Verification, Diarisation, and subsequent Transcription & Translation. In plain terms build an AI that can: → Identify and verify who is speaking → Separate multiple speakers in a conversation (diarisation) → Do all of this regardless of language or dialect → Then transcribe and translate the speech accurately It's a genuinely hard problem. And one that matters enormously in a country with 22 scheduled languages and hundreds of dialects, where the gap between what people speak and what AI can reliably process is still very real. Stage 2 was a full evaluation rigorous, demanding, and eye-opening. Results are awaited, but irrespective of the outcome, the experience of building this system, defending it in front of expert evaluators, and being in a room full of teams solving problems of national importance that's something we'll carry forward. Grateful to the organisers at FITT, IIT Delhi for putting together such a well-structured challenge. And to our team the effort we put in was real, and we are proud of what we built together. hashtag #FITT #SRECSPARK #RBG.AI #IITD