2025 Trends, Applied AI Challenges, and What to Look Forward to in 2026
Your Presenters
Aishwarya Naresh Reganti
Founder & CEO, LevelUp Labs
Early AI researcher at Alexa and Microsoft
35+ published research papers
Led 30+ AI implementations for AWS clients across legal, tech, banking, and medical
AI consulting clients include Deloitte, Microsoft, and Hitachi
Kiriti Badam
Building Codex at OpenAI
Building Codex, a software engineering agent
Previously built AI/ML + infrastructure at Google for ads-scale systems
Founding engineer at Kumo.ai (Forbes AI 50 startup)
We're also educators.
We create free and paid resources to help practitioners level up their AI skills.
Free AI Courses
Free Live Sessions
Paid Cohorts
What to Expect from This Session
You'll hear a lot of terms today. It's okay to feel overwhelmed—that's why we're recording this so you can revisit it later.
It's okay to not understand every word. We're keeping it as simple as possible so you can build a high-level story.
This is not an exec slide deck. No McKinsey-style reports, no name-dropping, no pitching numbers and figures.
This is practitioner-focused. Written by people actually working in this space, with realistic expectations—not a sales pitch.
What was the real breakthrough of 2025?
It wasn't just new model releases.
Models got better, but that wasn't what moved the needle for teams actually shipping AI.
It was plumbing.
Standards emerged. Integration got easier. The boring work of making agents actually work finally started paying off. The unglamorous infrastructure work became the competitive advantage.
The teams that shipped weren't the ones with the best models.
They weren't stuck contemplating which model to use. They knew how to connect everything together:and that's what mattered.
A Few Honest Lessons from 2025
Most of your time goes to integration. Not prompts, not model selection:connecting systems and handling edge cases.
Reliability beats capability. A predictable system is far better than something accurate but chaotic.
The model is the easy part. The hard part is everything around it:context, tools, evaluation, deployment.
Start narrower than you think. Build up to complex agents: making 10-step agents on day one only makes debugging harder.
What Most Teams Build
Impressive demos
Works in notebooks, fails in production. 95% never ship.
→
What Actually Ships
Boring reliability
Predictable, observable, recoverable. Does less, works always.
The Applied AI Stack
Four layers of the stack, plus the challenges that cut across all of them
What We'll Cover
Input Layer : From prompts to context engineering, meta-prompting, and multimodal
Model Layer : Foundation models, long context, RLVR, fine-tuning, and hybrid reasoning
Application Layer : Agents that actually ship, tool calling, and patterns that work
Output Layer : Trust as engineering, reliability math, and security frameworks
What's Still Broken : Hallucinations, RAG stagnation, and the production gap
Road Ahead : What 2026 looks like and how to prepare
