enterprise ai, and what the good models let me ship

this is the largest company i've worked at since the advent of 'ai being good at writing code' (Opus 4.7 was really the tipping point as everyone else eventually caught up to this standard), thus I was able to ship a shitload of projects that were immediately useful. even one of these would've only been 80% complete (and that would've been amazing!) within the duration of my tiny tenure if I did any of this before 2023, let alone before opus-level-models became the norm for programming.

so here's all of it, loosely in the order that it mattered.

x360 assistant

Answers business questions on existing enterprise applications. Product and infrastructure supporting 5-6+ different applications through a single npm package installation.

The single npm package install is tuff as shit. Five or six (and more) different applications, five or six different teams, and nobody gotta rearchitect shit to get the assistant in (well maybe if they want url chat sharing ;o) ).

knowledge corpus + hydration

For the x360 Assistant, we essentially are feeding large markdown files into frontier models to generate a response.

Yes, fine tuning local on prem models IS ABSOLUTELY the right solution here, but in an enterprise environment that is hell to move quickly on and the azure/aws 'fine tuning' solutions hardly give more attractive pricing for the amount of developer hours to make it work.

This consisted of me exploring and reading literature on how to best figure out how to generate 'the core knowledge base' as well as allow non-developers to make contributions to it.

One of the biggest challenges was keeping the knowledge base (per each enterprise application) updated while the other developers update the enterprise applications that the knowledge bases are generated from. This is absolutely hell without just fine tuning weights as I am just trying to get the same results in a super roundabout way with many constraints, however it works!

context window memory compression

Largely an implementation of the approach from here: https://arxiv.org/abs/2605.17304 (the paper frames it as typed, source-grounded semantic atoms with their own little compression language).

Similar to json, but without reliance on spaces, tabs, nesting. long.and.explicit.keys which counterintuitively actually perform wonderfully for preserving a conversations context window.

ai health telemetry

Metrics to help capture current and future changes to our enterprise AI infrastructure. Much of this was 'what do we log, where do we log it, how do we communicate it?'

I try to base everything either in: another (bigger) company's blogpost, or a current-ish research paper.

Considering that my x360 Assistant is likely the largest usecase of AI use for our internal tooling, it will help capture impact of changes to it over time.

agentic engineering debugger

Essentially a suite of javascript and cursor hooks that use cursor's api to surface prompts vs token usage.

Came about as I was really one of the only 'orchestrate' users (having picked it up from the opusplan model) which uses a stronger model to plan and weaker ones to implement (in the interest of cost efficiency).

This sees use to reduce overhead on certain project's super expensive cursor prompts and calls, surprisingly helpful and wonderful for the governance grab game (many just do policy without shipping shit so shipping shit that helps policy feels good).

uat tear-up-tear-down

Devops tools, even at Azure's attractive pricing, are expensive to leave up. Needed a way to hand a business user a simple URL that runs code to rapidly get and iterate on all of their feedback, yeah?

Essentially most of the work was already well established through process, this last piece was just bridging the gap to make it more turnkey (e.g asking cursor to interact with various scripts in Azure CLI to hopefully reduce the amount of clicks within the devops portal)

the people, actually

I had the most technical manager (director) ever in my half a decade of work experience, along with a very methodical senior developer both acting as mentors and peers while implementing all of these projects.

I have never met a bunch of friendlier, happier, easy to work with people that are genuinely excited to show up to work each day to ship legitimate impact to real world industry (BEYOND just pushing data over the wire).