Bhavtosh Rath's Blog

I recently landed a new data science/ML role, and wrote down the parts of the process that were actually in my control — apply-early habits, recruiter screens, behavioral prep, and why my first few interviews were rough on purpose.

A recent Netflix paper tries to separate taste from recommendation effect, then actually checks its own answer against a real experiment instead of just trusting the math.

Semantic-ID generative retrieval and HSTU-style sequential generation solve the same problem in opposite ways. Both are already in production at scale, and both share a bottleneck neither one has solved.

Generative recommenders look like they could finally bring real personalization to small retailers who cannot afford an ML team. An entrepreneur's honest stress-test of whether that's a business or just a nice property of the model.

RecCompl, Kunlun, TransX, and MTGenRec are Meta, LinkedIn, and Meituan quietly admitting that serving generative recommendation at real QPS means rebuilding the LLM serving stack from scratch, for embeddings instead of tokens.

Recommendation is converging on the same generative, transformer-based modeling that reshaped NLP — semantic IDs and sequential transformers are replacing two-tower retrieval at Netflix, Kuaishou, Meta, and beyond.

A look across Meta, Google Research, and Netflix Research shows the same shift happening in parallel: recommendation systems are starting to represent user preferences as language again, and explainability is coming along with it.

Part 3 of the value-router series: a Black-Friday-style demand simulation shows that a router calibrated for "normal" traffic can be told the calendar and still only improve a little — while a slow-path budget that simply reacts to the day's own estimates, with no idea a peak is happening, holds up almost perfectly.

Part 2 of the value-router series: why a routing decision is only half the story, and what a lightweight decision log and calibration check reveal about who actually gets the expensive path.

A routing pipeline that decides fast-path vs. slow-path treatment for retail items using both estimated difficulty and estimated value — and an eval that shows why difficulty alone is the wrong signal to route on.

I built a researcher-strategist-executor agent pipeline that personalizes item serving from simulated clicks — with every decision logged and explainable. Here is how it works, and what the eval numbers actually said.

Summarizing how Netflix leverages in-house LLMs to deliver highly personalized recommendations.