An overview of the work that we are doing in the software sustainability space.#software #sustainability #greensoftware #green. Our group: @SERC_IIITH @iiit_hyderabad #Grateful.

Huge congrats and best wishes to all graduating students of the 24th Convocation of @iiit_hyderabad Had a great time working with many of them!!. These are moments beyond words. First set of honors students graduated from #SA4S!!! Many congratulations to Arya Marda, Meghana

💥 Some news that’s truly special to share – our paper 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝐄𝐧𝐞𝐫𝐠𝐲-𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐭 𝐂𝐨𝐝𝐞 𝐯𝐢𝐚 𝐋𝐚𝐫𝐠𝐞-𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 – 𝐖𝐡𝐞𝐫𝐞 𝐚𝐫𝐞 𝐰𝐞 𝐧𝐨𝐰? has been accepted at @ICSEconf 2026 (Core A* - Main track) 🎉 .Authors: Radu Apsan

🎓 Happy to share that my first PhD student, Rudra Dhar (co-supervised with Prof. Vasudeva Varma) has successfully completed his comprehensive viva 🎉. His research focuses on an exciting and timely topic: 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐟𝐨𝐫 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞

🎯 The takeaway:.SLMs can serve as efficient, domain-specialized agents — and are a step toward greener AI. But more research and practice is needed. Thanks to @qualcomm_in EdgeAI lab at #iiith for supporting this work. And shoutout to @vinzstoico for presenting at #EASE2025.

📊 Key findings:.❌ Most SLMs struggled in zero-/few-shot (uncontrolled outputs).✅ Finetuned Deepseek-1.3B & Phi-3-mini hit >85% accuracy.🛡️ Prompt injection broke few-shot, but fine-tuned models stayed resilient.📱 Edge deployment:.→ ~5× less memory.→ ~10× more latency (6/n)

🚀 What we did:. 1. Benchmarked 5 SLMs on a 60k function-calling dataset. 2. Proposed 5 novel metrics for syntax + semantics. 3. Compared zero-shot, few-shot, and finetuned settings. 4. Evaluated prompt injection robustness. 5. Deployed on Qualcomm QIDK with 4-bit GGUF (5/n)

🔹 LLMs are powerful — but too generic, memory-heavy, and not ideal for on-device use. 🔹 SLMs offer domain adaptability, edge deployment, and smaller footprint — if they perform well. 💡 We asked:.Can SLMs really match up?.And are they robust enough to handle real-world issues.

In the age of agentic AI, the ability to generate function calls like:. getWeather(location=‘Paris’, date=‘tomorrow’).…is foundational. But here’s the catch 👇 (3/n).

This work was a wonderful collaboration, exploring the role of SLMs in function calling and their robustness in function calling, a critical capability for today’s agentic AI systems. (2/n).

🎉🎉🎉 Delighted and proud to share that our paper. Small Models, Big Tasks: An Exploratory Empirical Study on Small Language Models for Function Calling.was accepted and presented at @ConfEASE (Core A) in the AI Models/Data Track!.📄 Preprint: 🌐

Finally!! After so many heartbreaks. finally … what a Win for SA!! Proteas have cracked it this time… Gem of an innings by Markram. not to forget the fight from Bavuma. South Africa are the World Test Champions 🏆 Very well deserved!! Super happy.

🚀 𝐖𝐞’𝐫𝐞 𝐇𝐢𝐫𝐢𝐧𝐠: 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫 (𝐀𝐈 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠).Are you someone who’s equally excited about research and building real-world AI systems?. We’re looking for a 𝐑𝐞𝐬𝐚𝐫𝐜𝐡 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫 to join our team!. 🔍 What we’re looking for:

📢 Announcing our upcoming tutorial on 𝑺𝒐𝒇𝒕𝒘𝒂𝒓𝒆 𝑨𝒓𝒄𝒉𝒊𝒕𝒆𝒄𝒕𝒖𝒓𝒆 𝒊𝒏 𝒕𝒉𝒆 𝑨𝒈𝒆 𝒐𝒇 𝑨𝒈𝒆𝒏𝒕𝒊𝒄 𝑨𝑰 accepted for 19th European Conference on Software Architecture (ECSA) @ECSACONF, Limassol, Cyprus. Excited and looking forward to running this tutorial

RT @ANRFIndia: @ANRFIndia announces call for Proposals under Advanced Research Grant (ARG) and scheduled to open from June 12, 2025. The ca….

Spoke on Engineering Agentic AI Systems: From Lab to Land at the 11th TechForward Seminar by.@iiit_hyderabad,@HYSEA1991 & @ServiceNow.125+ attendees. Grateful for the opportunity & amazing discussions! .Slides 👉 @SERC_IIITH.#AgenticAI #MOYA #LLMs #sa4s

🎤 Excited to share! I’ll be speaking at an upcoming event on a topic we’ve been deeply exploring — Engineering Agentic AI Systems: From Lab to Land. Looking forward to diving into how we’re turning ideas into impact especially our work with @MontyCloudInc  .#sa4s #AIAgent #SaaS.