In this episode of the Sunny Ray Show, host Sunny Ray talks with Lord Munjal, founder of Alphaiva AI, about the frontier of AI and where today's models still fall short. Munjal explains that his company builds infrastructure for three customer groups: frontier AI labs, vertical AI companies focused on finance, and agent building companies. Using domain experts, the company finds weaknesses in models and agents, then supplies training data and reinforcement learning environments to fix them. He discusses failures such as lost long-term context, excess tool calls, wasteful token use, and confident answers that are not correct for qualitative expert decisions. Munjal also covers how experts are paid, why benchmarks go stale within months, and how agents will need to find the best path through services like travel aggregators. He shares lessons from banking at Citi and Merrill Lynch, explains his self-funded approach as a three-time founder, and notes that word of mouth works better than cold outreach. The conversation closes with how enterprises are training their own open weight models and where listeners can learn more.
Lord Munjal explains where frontier AI models still break, and how domain experts and RL environments are being used to fix them.
What is the company you're building and what problem are you solving?
We build infrastructure for frontier AI labs, vertical AI companies in finance, and agent building companies. We find where their models and agents are weak, especially on qualitative expert judgment, then work with domain experts to provide training data and reinforcement learning environments that make those models and agents perform better.
What did banking teach you that you still use as a founder?
I spent most of my corporate career in finance and investments, including stints at Citi and Merrill Lynch. I learned how to manage investments and the risk around them, how to take calculated bets, and how to adjust when circumstances or new information change. Handling complexity was the biggest lesson.
Which failures in AI models worry you the most?
Failures vary by model and agent, but I see lost context and weak long-term memory, poor tool calling, and wasteful token use. Using 50 million tokens when 5 million would do costs the client ten times more. Failures also rise in arcane domains like astronomy and gemology, where model builders have invested less.
What have you learned about who to take money from, having funded the company yourself so far?
I'm a three-time founder with a successful exit, so I was able to self-fund. Investors should share a similar vision, and being contrarian helps because you want your thinking challenged. The horizon should be long term, since venture is not a quick buck. Investors with network and domain experience are a better fit.
How do experts in your network get paid?
There are a few models. One is paying by the hour for their time. Another is task based. For example, a PhD horticulture scientist who develops 100 questions that stump a frontier lab's AI could earn about $100 per question, so roughly $10,000 for the set.
What will you be working on in the next 90 days?
The space moves so fast that benchmarks go stale in a few months, when new models score far above what older ones did. Our experts keep finding where models still fall short, then build training material and RL environments. Those environments keep growing with more tools and states as agents take on more tasks.
What has worked best for getting in front of agent building teams, and what has flopped?
Word of mouth works best, especially recommendations from researchers building the models and agents. Industry conferences also pay off, because we can show what we do. Cold calling and mass cold email have not been productive for an infrastructure company like ours.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interviewSend a founder here and we will know it was you.
One email a week. The best ideas from the show, read through bitcoin, AI and robots, and what founders should do about them.