Appen

Last week, Appen hosted the Responsible AI Fireside Chat at our Hyderabad office 🇮🇳, which brought together leaders from Bosch, CBRE, PwC, and more. On stage, Rajesh Dhuddu (PhD), Partner & Leader, Emerging Tech at PwC, and Flt Lt Bipin Chandra Dutt Pendyala, GM India Operations at Appen, led a powerful discussion on building responsible AI. Topics included why many AI projects fail, how bias is embedded at the foundational level, and the importance of guardrails, PII protection, and human-in-the-loop. They also shared examples of cyber risks like prompt injection and highlighted frameworks shaping the industry, from India’s DPDP Act to NIST standards. Real-world applications were also explored across healthcare, finance, telecom, and compliance - from credit decisioning with built-in safeguards to multilingual model development for Indic languages. Together, these insights reinforced a clear takeaway: building trustworthy AI requires high-quality data, strong governance, and explainability, paired with practical steps that balance innovation with risk. Stay tuned for a detailed debrief.

This week in Seattle, Appen partnered with AI Circle for a fireside chat on one of the most transformative topics in AI: Retrieval-Augmented Generation (RAG). Our speakers: Vladimir Karpukhin, TPM at Google and co-author of the original RAG paper Si Chen, VP of Strategy & Marketing at Appen Together, they unpacked: - Why the original idea of jointly training retrieval and generation was a turning point and how RAG only surged in adoption once LLMs like ChatGPT opened the door for enterprise use. - The hidden challenges of real-world deployment, from scaling vector databases to adapting retrieval for domain-specific data. - The critical role of high-quality, diverse datasets in reducing hallucinations and powering reliable AI systems. - What’s next: agentic retrieval, multi-modal augmentation, and infrastructure innovations that could reshape how enterprises apply AI. Thank you to AI Circle and to everyone who joined us in Seattle. 💡 Stay connected - our upcoming recap will dive deeper into the ideas shared on stage.

Appen x AI Circle Fireside Chat - Seattle We’re excited to bring together leading voices in AI for an evening of insight, innovation, and connection. 🎤 Speakers Vladimir Karpukhin – Co-author of the seminal RAG paper; former researcher at Meta; now at Google Si Chen – VP, Strategy & Marketing, Appen What to Expect A deep dive into the origins, evolution, and future of RAG - plus candid perspectives on where AI is headed Networking with innovators: technologists, founders, and leaders shaping the AI ecosystem Happy hour atmosphere with hors d’oeuvres, beer, and wine This event is designed for AI builders, researchers, and business leaders who want to engage directly with the people shaping the next wave of innovation. Date & Time: Wednesday, September 17, 7-9 PM PT 👉 Reserve your spot now: https://luma.com/gsm4ckop

🔊 Reflections from Interspeech 2025 One of the clearest signals from last week’s discussions: the future of speech technology will depend on data that is both broader in coverage and deeper in expertise. Across conversations, a few themes stood out: ✅ Low-resource languages remain a bottleneck. As models scale and fine-tune, the need for high-quality data in underrepresented languages is more urgent than ever. True global inclusivity in speech AI requires solving this gap. ✅ Medical speech data is in high demand. From diagnostics to patient interaction, researchers are seeking domain-specific datasets - paired with annotators who bring subject-matter expertise. Accuracy in these use cases is not optional; it’s critical. ✅ Specialized domains of speech are under active research. Work around speech disorders and children’s speech shows the field is expanding to areas where precision, sensitivity, and ethical responsibility matter most. At Appen, we see these conversations as a reminder that advancing speech AI isn’t only about larger models - it’s about more representative data, guided by human expertise. We’re excited to keep working with the research community to address these challenges and ensure the next generation of speech technologies is accessible, ethical, and effective. Tammy Ann Haskins, Giorgia Marcucci, George Krasovitsky

Appen is sponsoring Interspeech 2025 in Rotterdam. 🇳🇱 The conference theme, Fair and Inclusive Speech Science and Technology, reflects the importance of ensuring speech technologies represent diverse voices, languages, and communities. 📍 Booth: 3rd Floor 📅 August 18-21, 2025 You’ll also find Appen at the Speech Science Festival on August 17, connecting with attendees and discussing advancements in speech technology. Our team - George Krasovitsky, Giorgia Marcucci, and Tammy Ann Haskins - will be available to share how Appen supports global innovators with diverse, high-quality speech and multimodal data to build safer, more reliable AI. 👉 Attending Interspeech? Visit our booth to continue the conversation.

Association for Computational Linguistics 2025 brought together the global NLP community, and Appen was there as a sponsor to connect, learn, and share ideas. This year saw record participation, with an increasingly global mix of researchers, practitioners, and industry leaders coming together to shape the future of language AI. From our booth, we connected with customers, academics, and innovators to explore cutting-edge research developments and the role of high-quality, human-aligned training data across LLMs, translation and localisation, speech and dialogue systems, and enterprise AI applications. Key themes from ACL 2025 that are shaping the NLP landscape: • Generalization of language models – Development towards models that perform reliably and fairly on unseen data, across domains, and in diverse contexts—reflecting real-world complexity and human-like intelligence • Human feedback for AI alignment – Moving beyond a single “gold standard” to incorporate diverse and sometimes conflicting human values into model alignment frameworks • Multilingual performance and cultural nuance – Addressing persistent performance and alignment gaps for low-resource languages, and building systems that work for all communities, not just the most resourced. • Responsible AI at Scale – Balancing speed, efficiency, and ethical safeguards to ensure LLMs remain trustworthy as they become more powerful. Thank you to everyone who visited our booth to exchange ideas, share challenges, and imagine what’s next for responsible AI. We leave ACL 2025 inspired and energized to continue advancing inclusive, safe, and high-performing AI systems through exceptional human-in-the-loop data. Ryan Kolln, Si Chen, Sergio Bruccoleri, Brian Jenkins, Christian Neff, George Krasovitsky

🌍 Multilingual AI must go beyond translation; it needs cultural intelligence. Language isn’t just words. It’s identity, values, traditions, and social context. And when AI systems miss that nuance, the consequences can be serious, from miscommunication to real-world harm. In our latest piece, Sergio Bruccoleri, Project Operations Director (GenAI) at Appen, explores what it takes to build culturally adaptive AI and why accurate, inclusive, and high-quality training data is essential. 🔍 Key takeaways: · Up to 47% of contextual meaning is lost in traditional machine translation · Western-centric training data risks bias and cultural homogenization · Cultural nuance is critical in fields like healthcare, public safety, and global commerce The solution? Culturally adaptive AI built with representative data and human-in-the-loop evaluation 📈 And the real-world impact? Appen helped a global tech company expand from 5 languages and 10 dialects to over 30 languages and 70+ dialects, with 250,000+ culturally ranked interactions. 🔗 Read the full post: https://lnkd.in/eNHxPgGi

🌍 Appen is heading to ACL 2025 in Vienna! We’re proud to be a Silver Sponsor at this year’s Annual Meeting of the Association for Computational Linguistics, taking place July 28-30 at the Austria Center Vienna. 📍Visit us at Booth #8 Our team will be on-site and available for meetings to discuss: 🔐 Model Safety & Evaluation – Meet with Si Chen and Christian Neff 🌐 Multilingual and Culturally-Aware LLMs – Chat with George Krasovitsky and Sergio Bruccoleri 🧠 Complex, Domain-Specific LLM Work – Dive deeper with Si Chen and Sergio Bruccoleri We’re also excited to have our CEO, Ryan Kolln, attending ACL, connecting with researchers, leaders, and innovators from around the world. Don’t forget to stop by for our Vienna-themed swag! Let’s connect at #ACL2025 - stop by the booth, say hi, or chat with Brian Jenkins about how we support real-world enterprise use cases. 🔗 https://lnkd.in/exkW_2pV

🚀 Scaling Human Feedback for Safer, Smarter LLMs   When Cohere set out to improve alignment for their enterprise-grade LLM, Command, they didn’t just want annotations - they wanted expert, real-time human feedback at scale.   That’s where Appen came in.   Through our collaboration on PANDA Plus, we helped Cohere operationalize preference-based fine-tuning with: ✅ A vetted pool of LLM-experienced annotators ✅ Real-time delivery through our AI Data Platform ✅ Custom workflows for ranking, rewriting, and feedback ✅ Support across both production and experimental model variants   In just 12 weeks, our contributors logged over 2,400 expert hours, powering fine-tuning for one of the most trusted enterprise LLMs on the market.   Read the full case study to see how we turned subjective preference into structured training data: 🔗 https://lnkd.in/eXig-JGF Questions on how this came together? Alexi Schwartzkopff from our team can share more.

📊 Leaderboard scores are helpful, but they’re not the whole story. In the latest Forbes piece, Ryan Kolln reframes the overreliance on leaderboard rankings and explains why human evaluation is essential when selecting AI models for real-world applications. Whether you're building for finance, healthcare, legal, or customer experience, standard metrics only get you halfway. Human insight helps you assess alignment with brand values, regulatory needs, and actual user expectations. Read the full article on why more companies are embedding human review into their model evaluation strategy: 🔗 https://lnkd.in/eywy5_9a