To the Admissions Committee: I taught the applicant Machine Learning and supervised their final-year NLP project. I recommend them for your MS in Artificial Intelligence.
Their code-mixing classifier scored 89% on a held-out set and then failed the moment a real non-Twitter speaker used it — and instead of hiding this, they made it the centerpiece of their report, dissecting how their scraped training data had encoded assumptions about who its users would be. A student who studies their model's real-world failure more carefully than its benchmark success is exactly who belongs in a serious AI program.
They reimplemented backpropagation from scratch before permitting themselves a framework, and derived the attention mechanism by hand until they trusted it. That refusal to treat neural networks as spells is rare in a fresh graduate. I recommend them without reservation.
To the Admissions Committee: I lead engineering at a computer-vision startup and managed the applicant, an ML engineer, for two years. I support their MS in Artificial Intelligence application.
The applicant took a defect-detection model from a paper to a factory floor, where it held until dusk lighting tripled our false-negative rate. They patched it with augmentation heuristics that worked — and then told me plainly they didn't fully understand why, which is why they want graduate training in domain shift rather than more blog posts. That honesty is why I'm recommending them out rather than trying to keep them.
They cut our inference latency 60% and built our active-learning loop, so competence isn't in question. What they lack is the theory their job never taught, and they know precisely which theory. I recommend them strongly.
To the Admissions Committee: I supervised the applicant's research assistantship on low-resource language modeling. I recommend them for your MS in Artificial Intelligence.
Assigned to improve a Telugu NER model with barely any labeled data, the applicant resisted the urge to throw a large model at it and instead worked the linguistics — leveraging morphological structure and careful augmentation to beat our baseline with a fraction of the parameters. They understood that in low-resource settings, cleverness about the data beats scale, and they proved it empirically.
They document experiments so others can reproduce them and treat negative results as informative. They are ready for AI research and I endorse them with enthusiasm.