As giants like Visa, Mastercard, and Google launched their own agentic payment protocols, we focused on what startups actually need: fast, compliant, and trustable APIs that let AI agents transact safely.

This talk is a story of how we moved from building complex financial guardrails to enabling real payments-and what we learned about trust, autonomy, and the next financial interface. | | Prathamesh Devadiga | Teaching an AI a Language It’s Never Heard | Low-resource languages such as Tulu are virtually absent from large language model training data, creating a vacuum towards underrepresented languages and its potential impact. This talk explores we’ll explore how an AI can learn such a language from minimal data, without traditional fine-tuning, through self-learning, self-judging, and synthetic data bootstrapping. We discuss novel approaches such as LLM-as-a-judge evaluation, self-play dialogue generation, cross-lingual transfer from related Dravidian languages, and prompt-based linguistic emergence. | | Tushar Goyal | The LLM is thinking. Wait, how is it thinking? | We’re exploring how AIs actually think. Not just bigger models, but new methods that let them reason step by step like humans. We’ll talk about things like continuous reasoning, diffusion-style LLMs, and why reinforcement learning suddenly made AI good at math and code. I’ll keep the explanations simple, show why these techniques work, and share a few fun stories of the people who made the breakthroughs. | | Atharva Arya | AI for academic simulations | We are building a system that will build scientifically accurate graphs/ interactive simulations for academia to ease learning of complicated concepts for high school/university students- or for educators to create interactive simulations or content on topics they are an expert in.

The talk will include failures/successes on building the agent, the architecture, what worked and what didn't. Takeaways will include building better tools for LLMs, RAG tips, validation systems etc and will touch on ways to minimize cost while getting outputs as good as the better models.

Along the journey to build the agent, I stumbled on various ideas and principles on building agents and will be highlighting the most significant ones. | | Krishnan Ramalingam | Animease : How I'm Making Anime Creation Effortless | Animating a story means managing over 70 files for a single minute of video. On top of it, users are often confused on which voice model to use, which image model to use, how to prompt, how to get the lipsync right, and how to keep track of shots images and their assets. Animease is an all-in-one studio that transforms a writer's story into a finished vertical anime micro-drama, automatically handling everything from character generation to final video. With a simple editor for customisation, creators can iterate on their vision and produce a finished video without touching any other tool. | | Sindhoora Raja | AI-Powered Scent and Breath Detection | In this talk, I’ll explain how I’m using AI and chemistry to analyze the chemicals we exhale or emit, similar to how trained dogs detect diseases or scents. I’m building a GC-MS and machine learning pipeline to study these volatile compounds from sources like surgical masks and rank them by how detectable they are. I’ll walk through how the system identifies useful chemical signatures and then designs candidate mixtures for validation, with the goal of enabling simple, non-invasive health and scent detection systems. | | Rahul Bansal | Going from 0 to 3K revenue during loss funk | Sharing my experiments that I did to take Dictation Daddy from 0 to 3K revenue.  | | Krishna Dvaipayan | Personalised Learning Engines that Adapt Like a Human Tutor | We are building Ved AI, a personalised learning engine that models each student’s learning DNA to generate adaptive learning paths and content. Our work focuses on automating pedagogical insight: analysing learning behaviour, constructing structured learner datasets, and dynamically training models that adapt to individual cognition in real time. The talk will dive into how personalised learning systems can move beyond static videos and fixed curricula, touching on automated content understanding, reinforcement-driven pedagogy alignment, and the next frontier - AI that generates adaptive educational videos tuned to student attention and retention. | | Biraj Patel | The Itch to build an LLM Server, a Vector DB, and a Chrome Extension | I came to Lossfunk curious about how LLM servers actually work under the hood. Built one from scratch with static batching and hand-coded auto-regressive loop. Then while reading self-attention blogs, I got annoyed copying snippets to Claude, so I built my own AI sidebar that now has 100+ users. Then I got curious about vector databases and built one in the browser using IndexedDB. | | Rajeshree | Finding optimal drone flight paths using RL | In this talk, I will break down how RL can replace traditional rule-based navigation in autonomous drones. I will show my attempts about how a drone can learn to navigate forests and unstructured environments without being explicitly told every possible obstacle. I will walk through the challenges of real-world deployment, from perception and planning to handling uncertainty, and share what actually works when moving RL from simulation to live autonomy. |