TL;DR: I entered HackerRank Orchestrate wanting to win, then faced an AI judge asking how Praxi Clew actually worked, and I sought help mid-interview too. I finished with 66.1 out of 100 and a much more concrete understanding of what I need to learn. Next time, explaining the implementation needs to be part of building it.
I built a financial decision tool with AI assistance, submitted it to an AI hackathon, and then sat down for an interview with an AI judge.
At that point, I had spent hours asking AI whether the system worked. Now AI was asking me how the build worked.
What could possibly go wrong?
This time, understanding had an appointment.
HackerRank organized the September 2026 edition of Orchestrate, a 24 hour hackathon. Its Buy or Wait? challenge asked whether someone could afford an expense while meeting future commitments. HackerRank supplied the assignment and data; I named my implementation Praxi Clew. The official repository and problem statement are available if you want to explore it.
The build system used AI to extract facts from messages and images, then Python to forecast balances over 90 days and evaluate payment options. Astra handled orchestration, coordinating the AI extraction step with the Python forecasting. Antigravity, Google's agentic coding tool, implemented the code, running Gemini 3.8 as its model in the IDE. ChatGPT Astra reviewed the artifacts. I directed the work and increasingly insisted on evidence, limited scope, and an end to the correction cycle.
The submission required a code archive, recommendations for 250 requests, and a development transcript. Then came the 30 minute AI Judge interview, with the camera on and my submission available to the judge.
I knew there was an interview. Knowing something is on the schedule and being ready for it are apparently separate activities.
It took me a while to register how impressed I was by the interviewer itself. It started with introductory questions, followed up on what I said, pushed back on answers, and used those answers to shape the next questions. I thought it was very intelligent. The conversation kept asking me to go further into what I had offered.
HackerRank's September event announcement identifies the AI judge as Chakra. Its public explanation of AI interviewing describes adaptive follow up questions generated from candidate responses, with scoring against a defined rubric. That description fits what I experienced. I cannot identify the exact model or voice stack used in my session from the sources I reviewed.
The September event page describes the interview as a 30-minute voice interview where you "walk the AI judge through your architecture, your decisions, and your tradeoffs." HackerRank's account of the first edition says the interview probes whether a participant can explain code-level implementation details and demonstrate technical ownership, and discusses honesty about AI assistance and uncertainty. A rubric from an earlier Orchestrate event names depth of understanding, trade-off awareness, and failure-mode reasoning as dimensions, alongside honesty about AI assistance: whether a participant can clearly distinguish what they designed from what an AI tool generated for them. These sources are useful context for the kinds of abilities Orchestrate interviews have evaluated, but they should not be treated as the official rubric for the September edition.
During the interview, I repeatedly and clearly told the AI Judge that I had used AI throughout the project.
I wanted to win, always do! Even near submission, I was reading about previous Orchestrate results and asking whether they offered an edge. Learning in public is lovely and I so enjoy it. Also, I love me a leaderboard.
Before the interview, I did not think I would do well. I was an AI assisted builder about to be graded by an AI judge. I said as much to ChatGPT, then asked it to help me prep for it. Concerned and nervous, but still participating. That was where I was.
I still bombed several interview answers, because the real questions were more specific than anything I had prepped for. I brought questions into the chat and asked AI for help answering the AI judge. The system built with AI needed AI to explain itself, mid-interview, in real time. That assistance belongs in the account. I do not have the full spoken transcript, so the questions described here come from what I recorded in the chat. The explanations we developed are not a record of answers I delivered unaided.
One question concerned variable expenses. What value had we used to estimate them?
I understood why a financial forecast needed to account for groceries and transport. But which value did our implementation use? An average? The most recent transaction? Something more conservative?
The submitted function used the median of up to five recent amounts in the recurring series. A reasonable choice to reduce the influence of an unusually large or small purchase. Also a choice with limits: typical spending is not a guaranteed upper bound.
That is a short explanation once you have checked it. During the interview, I was asking for that check.
Another question concerned the order of events. If salary and bills landed on the same day, what happened first?
The code credited confirmed income, deducted existing expenses, then deducted the proposed purchase payment. That assumes income is available before the outgoing payments that day. It does not model the exact times a bank settles transactions.
This answer exposed a problem with the explanations I had been relying on. An earlier report said debits came first. Inspecting the submitted function showed credits came first. Both descriptions sounded plausible. Only one described my submission.
The questions also reached the search for spending changes. Python tried combinations of up to three permitted changes, such as stopping an eligible expense or reducing it to its minimum allowed amount. It checked payment plans against the forecast. Protected categories were excluded.
Then came verification. How did we know the outputs were correct across all 250 requests?
That question needed a careful answer. The public dataset gave us 25 examples with expected answers. The other 250 were the submission requests. We could check their format, permissions, schedules, and balances against our forecast. We could replay exported evidence and reproduce the identical output file. We could not claim agreement with expected answers we had never seen.
The build had already supplied a useful example of why that distinction mattered. In request 185, a 12% rent increase became rent of €12. The validator passed because it used the same incorrect rent as the forecast. Comparing the extracted fact with the original message exposed the error. Correcting rent from €451 to €505.12 moved the recommended payment from December to January.
This was also why I could not answer every question with "we validated it." Validated what, against which inputs, and with what limitations? Those details belonged in the answer.
The question about what I designed myself called for a different kind of answer.
I design my projects, vet the implementation, and troubleshoot bugs with the agents. HackerRank supplied this assignment, and AI proposed and wrote substantial parts of the technical solution. I was still responsible for the project as a whole. That includes whether I can explain the code, not just direct it or evaluate what came back.
This example is from mid-build, well before the interview. I stopped and thought, wait, what are we doing here? We were going in circles: another confident report, another correction, another review. I asked for a reset and a complete assessment of the files and requirements. I also required guardrails so a targeted fix would not become permission to change unrelated behavior.
That intervention changed the review process. We assembled the actual code, output, requirements, and extracted evidence so the next assessment could examine them together. I was troubleshooting how we were evaluating the system as well as directing corrections to the system itself.
Needing help with the estimation and validation details is different from being absent from the design and review work. I want this account to preserve both: the decisions I owned and the implementation knowledge I still needed to deepen.
The interview ended okay but I was definitely not confident. Just relieved I completed it. Then when Orchestrate ended, I got my score.
My reported placement was 139 out of 3,062 participants, approximately the top 4.5%.
I wanted to win but didn't. I am pleased with that placement anyway which surprised me for sure!
The transcript received full marks. I cannot tell which individual decisions earned those marks, but it was my strongest evaluated component. The interview, code, and output scores were relatively close. The results do not support blaming everything on interview nerves.
The public sample results had already shown limitations: 80% agreement on payment method, but only 12% on the exact safe payment amount. The forecast needed work. So did my ability to explain it, especially this! The score made both more concrete.
For my next build, I want explanation to become a checkpoint while the implementation is still taking shape. When an agent hands me an answer, I want to be able to describe back what a component does, which inputs it uses, and where it can fail. I do not yet know exactly how I will check that against the code every time. That is what I need to figure out next. Which will take time. It will probably interrupt the satisfying pace of the agents firing on all cylinders. But, I think that is time I need to spend.
For the next Orchestrate or hackathon or challenge, I will read and review the interview rules as carefully as the build requirements, including what assistance is permitted during any live assessment. Knowing my own implementation well enough to survive questions more specific than anything I prepped for is the harder problem, and the one I actually need to solve. Next time, I am saving the receipts: the chat logs, the test outputs, every place where I need to know the reasoning, not just the result.
HackerRank's transcript logging instructions are worth reading before starting, too. The development record was a scored part of this event. These are the September edition's details; check the organizer's current instructions for future contests.
I still build with AI. I still want to compete. Next time someone asks which value we used for variable expenses, I would like to have already explained that choice to myself.
Have you had to explain a system you built with substantial AI assistance? Which question sent you back to the code?
This is my personal participant account. HackerRank organized Orchestrate and supplied the Buy or Wait? challenge and data. Scores and placement come from my result display. Antigravity assisted with implementation, and ChatGPT assisted with review, interview questions, and this article.
AI Assisted. Human Approved. Powered by NLP.