From Wildcard to World Stage: My Journey to 1st Runner-Up in the AWS AI League Championship
An inside look at the new AWS AI League through the lens of a competitor moving from DeepRacer to model fine-tuning and agentic AI challenges. This post captures the experience, technical takeaways, and future possibilities for builders exploring the next era of AWS competitions.
This year’s re:Invent was nothing like my previous ones. What started as a last-chance wildcard attempt turned into one of the most intense and meaningful competition runs I’ve ever had. Most of you know me from the DeepRacer League, but with its sunset and the rise of the new AWS AI League, everything changed for me this year…
It’s been a wild ride this past week at #AWS #reInvent2025. With DeepRacer ending and a brand new league taking shape, I felt compelled to jump in. Since I couldn’t attend any Summits, the wildcard at re:Invent was my only path - and huge thanks to my company Aqfer and Daniel Jaye for rolling the dice and sending me to represent.
I still knew very little about the new AI League format except that engineers would be fine tuning smaller models to compete against larger frontier models. After talking with fellow DeepRacer winners who had seen it earlier in the year ( Daryl Jezierski , Mark Ross , and a very generous Medium article from Sai Ram Naragoni ), I finally had a better idea of what to expect. In the days leading up to re:Invent, I began preparing by thinking through what synthetic data generation strategies I might use depending on the challenge.
The Wildcard Qualifier
Monday morning at the MGM. 100 participants. 5 spots available. The challenge was to fine tune a 3B model to outperform a 90B model on a re:Invent builder guide. We generated training sets, adjusted hyperparameters, ran fine tuning jobs, and submitted SageMaker endpoints to the leaderboard. It was also great seeing other former DeepRacer competitors there! Robin Castro Tomasz Ptak YUTAKA KASHIWABARA
The competition was fierce. At the end of the 4-hour window, I ended up in 1st place and earned a spot in the semifinals - along with my good friend Mark Ross , who took 5th!
And yes, I finally got to meet Jeff Barr, who was kind enough to take a picture with me at the top of the leaderboard.
The Semifinals
With that, my goal for re:Invent was complete. I made it to the semifinals, joining 29 others who had earned their spots throughout the year: 5 from the wildcard, 5 from AWS Summits, and 20 champions from internal corporate competitions around the world. We received information that evening about the next challenge, something to do with agents rather than LLMs alone. The top 3 out of 30 participants would move on to the finals.
After late-night travel and anticipation, the adrenaline of the morning competition, two more workshops during the day, and a long walk to the Encore, I was honestly exhausted and was happy to be a low-key participant in this new challenge… until they introduced it:
The goal: build an agentic system to conquer the Dungeons of Agentic Interactions by completing challenges using defined tools, defeating the boss, not dying, and finding the treasure chest - earning as many points as possible within a set time limit.
Swami’s keynote this week had a theme close to my heart:
“Remember the first program you ever wrote? Remember that sense of accomplishment, freedom, and the endless possibilities?”
For him, that program was a scientific calculator.
For me, it was a pathfinding algorithm written in Pascal… to solve a maze.
After the rules were read, I was no longer exhausted, and competitive focus settled in its place. Synthetic data generation would be no help here - just pure agentic application building. Using Bedrock AgentCore and a VERY nice UX provided by the AI League team, 30 of us built out our agentic workflows using an orchestrator agent, subagents, memory, tools, and guardrails to see whose agent could complete the dungeon crawl the fastest and with the most points.
I was near the top of the leaderboard multiple times, but with the 3-hour window wearing down, my agent was close but not close enough to clinch a top spot. Until a few last-minute modifications and a hail-mary submission with only five minutes left, my username, rosscomp1, showed up as #2 on the leaderboard - guaranteeing a podium position and a spot in the finals.
I had just made the finals. My re:Invent competition spree was not over after all.
The Finals: Live on Stage
The next evening was the finals, live at the Expo on a stage in Theater 3. Our established agents would face multiple mazes, all with different time limits and new challenges. Three rounds. Forty-five seconds to adjust your runtime prompt, then 4–5 minutes to watch your agent navigate the dungeon. Round 1 tested our foundations. Round 2 raised the stakes with multiple challenges between start and treasure. Round 3 introduced a boss battle right before the chest, requiring the coordination of multiple subagents and tools.
It was a fantastic event to take part in, being MCed by none other than Nicole and Blaine, who also were announcers of the DeepRacer League! My agent did as well as I could have expected and landed me securely in the 2nd spot on the podium, a place where I am very happy and humbled to be.
🥇 Hemanth Vediyera (Cisco Systems): 25,709 points
🥈 Ross Williams (Aqfer): 19,933 points
🥉 Deepesh Khanna (CapitalOne): 12,810 points
Congratulations to Hemanth and Deepesh. It’s been an honor sharing the stage with you! We were also shown in Swami’s keynote - a dream come true!
The prizes were not trophies this time, but championship belts, which I proudly wore throughout the rest of the conference on my shoulder at the demand of the DeepRacer developer community… but honestly, I probably would have done it anyway!
One More Podium
The next day, I also took part in another workshop around fine tuning a Nova model using SageMaker and deploying on Bedrock - a more polished, efficient training and deployment process similar to Monday’s wildcard, but this time on the topic of building a Vegas event guide. I ended up placing 2nd in that competition as well!
You can also find more info about the events on Tomaz's blog posts here:
https://builder.aws.com/content/36GfcVoL7ogmLLrG0EmJv9zZVeX/aws-ai-league-last-chance-qualifiers-semifinals-remaining-sessions-and-year-2026
https://builder.aws.com/content/36KAiofgazTuOH4StPlyIXpczJV/aws-ai-league-2025-finals
This was my 6th re:Invent, and it was surreal for me. Myself and others were sad to hear of the sunset of the DeepRacer League, but what’s rising in its place has real potential, and many of us are very excited about the possibilities. Fine tuning LLMs to be specific experts, or creating agents to complete an endless variation of possible challenges in the form of a dungeon crawler game - we are so very excited about the future of the AWS AI League.
A huge thank you to the AWS team, who are always there to guide us, help us, and advise us. We know how much work went into making this experience as impactful and engaging as possible, all while making sure we were having fun doing it. In Werner’s final keynote (the end of an era!), he said:
“Your best work is hidden.”
Just know — we see it, and we appreciate everything you’ve built for us this season. We know what it takes, and we are so thankful for the opportunity you give us, and for the chance to know you along the way.
Keep building these amazing games… we’ll show up to play them.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article