tiered assistants
production assistants run on three model tiers. cheap models answer instantly, the expensive one wakes up only where money is at stake.
~5% of calls reach the premium tier
* field notes from production
the layer between
frontier models
and production.
i build agent systems that survive production and inference that never gets a gpu bill. what actually works ends up in the two channels below.
01 · the main channel
borkiss notes
nerd notes on the ai scene. what works, what's hype, what it costs in tokens.
02 · the community
modelhost
the local inference community. run models on your own hardware. setups, benchmarks, no gpu invoices.
everything below is proof. the channels are the point ↓
01 · what's running
consilium
sixteen agents argue in isolation, a judge assembles the final answer. the argument itself does the work.
78% of the gain comes from the loop, not the models
inference in the browser
llms and video detection run entirely on the user's device. no server, no invoice. 3 live demos ↗
6.5× faster · 724 MB instead of 2.4 GB
50B+tokens a quarter, moving through my agent pipelines
8300+commits a year. my agents write the software
142★bs-p, an avx-512 market-making kernel
02 · the path so far
-
head of ai, an agency (nda)
ai strategy, agent pipelines, safety classifiers and the tooling underneath. 50b+ tokens a quarter say it works.
-
founder & head of engineering, fdaa.dev
we embed ai engineering into clients' operations. the system stays with them, code and access included. open source on the side.
-
quantitative trader
market-making kernels, signals, data pipelines. the adversarial habits from that era are the exact failure modes the agent scene trips over now.
-
co-founder, holypoly.ca
cto & cao hat. prediction markets, community, zero boring meetings.
03 · elsewhere
04 · write to me
the fastest way in is telegram.
write here. the rest is noise.