Nothing earth shattering in Engineering at Affinity Answers in 2025

After not doing anything earth shattering in 2024, our streak continued into 2025. This year had two tracks — the heritage track and the contemporary track.

The heritage track

Great civilizations have lost the learning by not recording what they do and did; we do not want to pass into the history like that. So, this year we took time to record them. Disney (not the brand) talks about how the art and science of whipping up Audience Recipes here at Affinity Answers. And then like the 1000 pillar temples, we have the 100 thousand entity catalog — carefully built — entity by entity for over 15 years, we had to tell that story — Soundar and Disney tell that here. Some architectures stand the test of time and audience tech stack had stood the test of time — Santhosh and Ashish record the nuances of the tech stack that processes millions of records daily here.

The contemporary track

The popular columnist Morgan Housel writes about Magazine Architects — It’s a derisive term architects use for their colleagues who design buildings that look beautiful, grace magazine covers, and win awards, but lack functionality for the tenants. Swapna’s confessions of a sub-optimal design is a landmark in balancing between elegance and maintainability while refusing to be a “magazine architect”. M Kavitha follows the same approach in automating ETL pipelines with human-in-loop for efficiency — emphasizing there is no utopian complete end-to-end automation nor a dystopian zero automation — everything has a middle ground.

The long shots

At Affinity Answers, we always have a track that works on what could be next. Needless to say it is the powering the products with LLMs. The duo of Lakshay and Ashish have given birth to several internal tools using LLMs and one promising application — the AdChoreo. Serious AI applications that users use to make bottomline changing decisions better be right — they found things that are normally talked about in detail — the context and horror of math in LLMs — read them if you want to build serious LLM based applications.

Sometimes byproducts of long shots are in the short-term more useful than the long-shots themselves (like Post-it). Data partnership manager S Kavitha used the know-how gained above by using AI to build a brand catalog in record time; we cannot really write about it here as it has some secret sauce that is too early to make it to public domain.

Meanwhile, Agentic Advertising was born in mid 2025 with the advent of Ad Context Protocol. We built a signal discovery agent scratching the surface of the AdCP protocol. Read Santhosh’s writeup on how we built it.

We started the year with AdTech:Delhi (we are likely to in 2026 as well). Here is one on that.

Finally, as you can see, nothing earth shattering this year as well; bring it on 2026!

PS: No Gen AI tool has been used to create this post; the (real) photo is from an exotic location we had been to for our off-site!