skip to content
[ standing by ]channel initialising
[ system online ]exer//runtime

AI, code, systems and useful experiments. I build strange things to see what actually happens after the demo.

latest transmission

exer//watch — 01

the stream fills itself

Videos are read from the YouTube Data API at build time and cached in the repository. There are none to show yet, so there are none shown — no placeholder uploads, no invented titles, no borrowed thumbnails. Subscribe and this section will be full the next time you land here.

channel universe — 08 topics

Channel topics

  1. ai

    models · evaluation · what the benchmark does not tell you

  2. architecture

    enterprise ai platforms · integration · the diagram versus the invoice

  3. code

    written, refactored, occasionally deleted

  4. agents

    orchestration · tool use · the loop that would not terminate

  5. experiments

    [ some of these will not work ]

  6. tools

    tested properly, not unboxed

  7. systems

    infrastructure · governance · the parts nobody demos

  8. ideas

    probably overengineered. building it anyway.

projects in motion

exer//build — 03 active
buildingproject 01

agent that files its own bug reports

A long-running agent with a memory of what it broke. It reads its own traces, works out which step failed, and opens the ticket. Mostly it opens the ticket about itself.

  • typescript
  • local model
  • sqlite
worked
the memory layer
failed
every retry policy, twice
next
make it stop apologising
read the write-up
diagram · 16:9 export
testingproject 02

a governance layer nobody asked for

Policy, audit and approval routing wrapped around a model gateway. Every enterprise wants it in month six. Nobody puts it in the demo.

  • python
  • opa
  • postgres
worked
the audit trail
failed
latency budget, comprehensively
next
move policy evaluation off the request path
read the write-up
diagram · 16:9 export

visual archive

rails pause on hover and on focus · click any frame to open it · reduced motion holds them still

right now

exer//runtime
building
an agent that reads its own failure traces and files the ticket itself
testing
how much of an agent stack survives contact with a real approval process
learning
where local models stop being a novelty and start being infrastructure
next
the retry-policy post-mortem, filmed at the whiteboard

field notes

exer//notes
“The architecture looked simpler on the whiteboard.”
  1. note · 001what an agent framework actually costs once you add the approvalsdraft
  2. note · 002the retrieval step everyone skips, and the week it costs laterdraft
  3. note · 003running a model on hardware that should not be able to run itdraft
Cristian Exer, photographed in an office
[ the human ]

I build AI systems that have to survive a Monday morning.

Enterprise architecture by day, experiments by night. I am interested in the gap between the demo and the deployment — governance, cost, latency, the awkward integration nobody put on the slide. The channel is where I close that gap in public.

Not a CV. If you need one, it is a link away.

[ end of transmission ]

Build it. Break it. Explain it. New transmission whenever something finishes breaking.