Science does not work directly from induction, it requires a theory in addition to observations. That provides a mechanism that forces statistics to apply. As you repeat an experiment, the odds it will ultimately fail to repeat in the future do go down. They never get to zero. You could still be wrong. But that becomes less and less likely as you successfully reuse the theory over time.
Even if you don't do the p-value computations, you are safely covered by the Rule of Succession, as long as you have a falsifiable theory, and you challenge it. You have some odds p < 1 of being wrong, and if you successfully repeatedly use your result n times the odds it will eventually fail are <= p^n, which converges to zero when n is infinite. Since 'being falsifiable' implies you would recognize when you failed, you are legitimately sampling a distribution, and this is not induction, it is math.
Each time you 'fail to falsify' you 'reject the null hypothesis' and in doing so you are building a statistical basis for believing your model and its underlying assumptions are less and less likely to be wrong to a significant degree. If you do keep actually compute the p-values, as people in very slippery sciences like psychology and sociology force themselves to do, you can go back and figure out just how unlikely ultimate failure is becoming due to seeing the outcome in various instantiations.
Statistical convergence is not really induction in the pure philosophical sense, it is deduction with probabilistic truth values, that never really reaches 100% reliability.
So he is 'right' that science is never logically watertight. And we already knew that, given that it constantly changes and evolves. But he is wrong to say the procedure is not deductive.