HackathonAmazon

Amazon ML Challenge 2026

Amazon's national machine-learning competition for engineering students - teams get a real Amazon dataset and 72 hours to build a working ML solution, with strong performers considered for an Applied Scientist Intern PPI.

Brabble editorialUpdated 13 Oct 2026

Official page

Organiser: Amazon
Mode: Fully online (a one-day grand finale for the top 10)
Prize / outcome: ₹2,25,000 total (₹1,00,000 / ₹75,000 / ₹50,000) plus certificates and goodies; top 50 considered for PPIs
Team size: 3-4 students, one team leader
Application fee: Free
Rounds: 4

What it is

Amazon ML Challenge is a machine-learning competition run by Amazon for engineering students in India. It is separate from HackOn With Amazon: instead of building a product, teams of three or four work on a real machine-learning problem with a dataset supplied by Amazon, and are ranked on a leaderboard. According to the 2026 listing, registration ran from 7 to 20 September 2026, the ML challenge ran from 9:00 AM IST on 25 September to 9:00 PM IST on 27 September (the page calls it a 72-hour hackathon, though the listed window is about 60 hours), the results and the top 50 were announced on 2 October, and the grand finale for the strongest teams was on 7 October 2026. Teams receive the problem statement and the dataset on day one, and submit by day three, together with a one-to-two-page document explaining their approach. Reports on the 2025 edition describe industry-grade problems such as product classification, ranking and recommendation, a live leaderboard during the event, and a private leaderboard calculated on the complete test set afterwards. The scoring metric is set in each year's problem statement and is not published on the listing, so read it before you design anything. It works like an applied-science hiring assessment: the same habits that win it, such as honest validation and clear reasoning, are what Amazon scientists look for in an intern.

Rounds

01
Registration and team formation
Teams of three or four register on the challenge listing before the deadline (20 September 2026 at 11:59 PM IST in 2026). One member is the team leader, and every participant can belong to one team only. No dataset or problem is released at this stage.
02
The 72-hour ML challenge
At the start of the window (9:00 AM IST on 25 September in 2026) teams receive the problem statement and the dataset. Solutions are due by day three (the 2026 window closed at 9:00 PM IST on 27 September), as code plus a one-to-two-page document describing the approach. Reports on 2025 describe a live leaderboard during the event and a private leaderboard on the full test set afterwards, so a model that is over-tuned to the public leaderboard can drop when the final ranking is released.
03
Results and top 50
Results and the top 50 were announced on 2 October 2026. The top 50 teams are considered for pre-placement interviews for the Applied Scientist Intern role, and the top 10 teams are invited to present to Amazon scientists.
04
Grand finale
The grand finale was scheduled for 7 October 2026, from 10:00 AM to 3:00 PM IST. In 2025 the finale was a live virtual presentation to Amazon's scientists on 17 October, as reported. Finalists explain the approach, the validation and what they would do next, and answer questions.

How to prepare

Step 01

Build a validation habit before the challenge

A leaderboard rewards models that generalise, and a private leaderboard on the full test set punishes over-fitting to the public one. The strongest teams trust their own validation more than the live ranking. This advice is Brabble's synthesis, not an official rubric.

Set up cross-validation that matches how the test data will be split
Track one metric that matches the competition metric and compare every change against it
Never tune repeatedly on the public leaderboard alone
Keep a log of experiments with the score and the change for each
Fix random seeds so results can be reproduced

Step 02

Plan the 72 hours

Three days is long enough to try many ideas and short enough that a poor plan wastes them. Teams that start with a quick baseline always have something to submit.

Hour 0 to 6: read the problem, explore the data and build a simple baseline that submits
Next day: improve features and try two or three strong model families
Last day: ensemble, check for leakage, and freeze the final pipeline
Assign roles: data and features, modelling, validation and write-up
Keep the last six hours for the write-up and a clean, runnable repository

Step 03

Know the common problem types

Reports on the previous edition describe classification, ranking and recommendation on product data. Being comfortable with these shortens the time to a good baseline.

Text and tabular features: TF-IDF, embeddings, gradient boosting
Gradient boosted trees and how to tune them
Pretrained models and when fine-tuning is worth the time
Ranking metrics such as NDCG and recommendation basics
Leakage: duplicates and information that would not exist at prediction time

Step 04

Write the one-to-two-page approach document

The document is part of the evaluation and is what the scientists read first. A clear, honest page beats a long one.

State the problem, the metric and the validation scheme in the first paragraph
Describe what worked, and also what you tried that did not
Include a small table of experiment scores
Say what you would do with another week
Keep it to two pages and make the code runnable from the README

Representative practice - not leaked questions

Questions to practise

01
You are given product titles and descriptions and must predict a category from a large taxonomy. What is your baseline, how do you validate, and what would you try next?
What to show: Text classification and validation
02
Your public leaderboard score is much higher than your cross-validation score. What do you suspect, and what do you do?
What to show: Leakage and over-fitting to the public set
03
You must rank products for a search query given historical clicks. Which metric do you optimise, and how do you handle position bias?
What to show: Ranking metrics and bias
04
You have 12 hours left and three promising ideas. How do you decide what to ship?
What to show: Time management and risk
05
Explain your model to a scientist who has not seen the data in two minutes.
What to show: Communicating the approach

Who can enter

Students in full-time PhD, M.E., M.Tech., M.S., MS by Research, B.E. or B.Tech. programmes at engineering campuses in India
Belonging to the 2027 or 2028 graduating batch (the 2026 edition's window)
A team of 3 to 4 members with one team leader; students from different colleges may team up
Each participant can be part of only one team
Free to enter. CGPA requirements are not mentioned on the listing

What you actually get

The three prizes in 2026 were ₹1,00,000 for the winner, ₹75,000 for the first runner-up and ₹50,000 for the second runner-up (₹2,25,000 in total), each with a certificate and goodies, and the listing also mentions AWS credits and Amazon merchandise. The top 50 teams are eligible for pre-placement interviews for Amazon's Applied Scientist Intern role, subject to the usual eligibility requirements, and the top 10 teams, and the top 10 women-only teams, receive certificates and Amazon swag. A pre-placement interview is a chance at an interview, not an offer. Teams also get to present their work to Amazon scientists, which for many is the main reward.

Frequently asked questions

Is Amazon ML Challenge the same as HackOn With Amazon?

No. They are separate programmes. HackOn With Amazon is a team hackathon built around problem themes. Amazon ML Challenge gives teams a real dataset and a machine-learning task and ranks them on a leaderboard.

What is the status of the 2026 edition?

Every stage has passed as of 13 October 2026: registration closed on 20 September, the challenge ran 25 to 27 September, results and the top 50 were announced on 2 October, and the grand finale was scheduled for 7 October. Treat this guide as preparation for the next edition and check the listing for its dates.

Who can take part?

Students in full-time PhD, M.E., M.Tech., M.S., MS by Research, B.E. or B.Tech. programmes at engineering campuses in India who belong to the 2027 or 2028 batch (the 2026 edition's window), in teams of three or four. The listing does not mention a CGPA requirement.

Can teammates be from different colleges?

Yes, according to the listing, students from different colleges may form a team, but every participant can be in only one team.

What do I submit?

Your solution code and a one-to-two-page document explaining the approach, by day three of the challenge. The document is part of the evaluation, so it should state your validation method and why you chose the model.

Does it guarantee a job or internship?

No. The top 50 teams are considered for pre-placement interviews for Amazon's Applied Scientist Intern role, subject to eligibility. That is an interview opportunity, and the result still depends on the interview.

What is the evaluation metric and what is the dataset?

They are set in each year's problem statement and are not published on the listing. Reports on the 2025 edition describe industry-grade tasks such as product classification, ranking and recommendation. Read the problem statement on day one carefully, because the metric decides what a good model is.

Is there a fee?

No, entry is free.

Comparing options? See how Amazon ML Challenge 2026 stacks up against other AI and ML hackathons for students in India.

Dates, eligibility, team rules, prizes and the PPI and swag details were read from the Internshala listing for Amazon ML Challenge 2026 on 13 Oct 2026; the Unstop listing could not be read automatically. The dataset type, the live and private leaderboards and the 2025 finale come from third-party reports of the 2025 edition and are labelled as reported; the evaluation metric is not published on either listing. No separate Amazon-hosted page was found. Preparation advice is Brabble's synthesis, not an official rubric.

Sources and further reading

WhatsApp LinkedIn X

More guides

All guides