AR Panel Talk Project Deep-Dive

Applied Research opener. Present a project or paper to a panel of three senior researchers. The depth probe is the entire round — high-level slide-walks fail; the round expects derivations, ablations, and failure modes. Requirements 5-10 minutes presenting one project or paper of the candidate's choosing. 40-50 minutes of Q&A. The panel cycles questions; each researcher follows up on a different axis. Coverage expectations: Problem framing — why does this problem matter? What's the gap in prior work? Method — derive the key equations or training objective on the spot; don't gesture at them. Experiments — what were the ablations? What didn't work? What is the minimal set of experiments that supports the claim? Limitations — where does the method break? What's the failure mode you observed? Extensions — what would you do next if you had three more months? Notes The most common failure mode is presenting at too high a level. Saying "we used contrastive learning" without being able to write the loss on the whiteboard is a fail signal. Researchers want to see that the candidate genuinely owns the work, not summarises it. Pick a project where you can defend every architectural decision, including the ones that were the obvious default. "Why Adam not SGD?" should not stop you. Have an explicit failure-mode story. Researchers grade highly on candidates who can name what didn't work and explain why. A presentation with no failure analysis reads as either dishonest or shallow. Bring derivations: at minimum, be able to write the loss function and the gradient. For RL projects, the advantage estimator. For supervised learning, the link between the loss and the calibration / decision rule. For paper presentations, the panel may have read the paper; this raises the bar. Stick to a project where you have hands-on experience, not just one you find interesting. Preparation Rehearse the 10-minute pitch three times with a peer; force them to ask the hardest possible question after each section. Build a one-slide cheat-sheet of the key equations and have it open in case you blank. Prepare a structured failure-mode story (one paragraph): what failed, what you observed, what you hypothesised, what you tried, what worked and what didn't. Pre-rehearse the extension story; the round usually closes with this and it sets the lasting impression.