Grateful to have placed 2nd runner-up in the inaugural IBM × AWS AI League held last month, a global IBM competition with over 700 participants from across the world. Huge congratulations to Marcin Walków, you’re truly a well-deserved champion! 👏👏👏
The AI League tested advanced prompt engineering, LLM fine-tuning, strong business judgment, rigorous problem solving, and analytical skills.
We were challenged to fine-tune a compact Llama 3.2 3B Instruct model under tight time (10 hours) and compute, then produce a single slide executive recommendation across varied customer scenarios that included constraints, goals, timelines, and budgets.
The outputs from our model were determined by an AI judge scoring 50 blind questions that compared with a 90B model. I briefly topped the Leader board for a few days until the last few hours and entered the Finale in second place with a 98% win rate. That reflects how fierce the qualifier was!
In the live Finale, AI judge scoring was reduced to 40%, and human judging accounted for 60% (split between human judges at 40% and the audience, 20%)
Key takeaway – with curated data and disciplined LLM fine tuning, a small language model can outperform a much larger model on targeted tasks. It aligns with a point of view in a client Technology Radar that I co-authored earlier in June this year that I stressed on the necessity of implementing FinOps for LLM, as right sizing with fine-tuned small models can deliver comparable or better outcomes at materially lower latency and cost. It is gratifying to see the thesis validated in competition.
A small but telling lesson from the finale – a formatting hiccup (missed line breaks) made my output harder to read. Top 3 participants for the Leaderboard where AI was the judge did not win the Finale. It was a reminder to me that we build for people first. Human-centred AI, always.
Thank you to the organizers, Yee Min Chong and the AWS team for a thoughtful use case and steady support. Thank you to my leader Wan Taow for championing IBM Consulting Advantage (ICA) Assistants. I relied on ICA (with IBM’s Granite) throughout and eating our own dog food helped me iterate faster within tight timeline and achieved a good win rate.
Proud of everyone who competed. AI League has been great fun, and the highlight for me was the spirit across IBM. Even as we aimed for the Finale, colleagues stepped up to help one another during the qualifier. In my view, all the participants are winners.
On to the next IBM AI League in 2026.
Note: This post was first originally published at LinkedIn in 2025 – 2025 Inaugural IBM × AWS AI League
Date Published: 2 January 2026