career bridge
The gap
Put a number on the distance between your best demo and a system, and mark every unknown honestly as unknown.
You can build a working demo. You have probably built several, and one of them impressed somebody enough that a conversation started about shipping it. What you have not done is put a number on the distance between that demo and a system, and that distance is what this whole road covers.
What you do
Read the market, then audit your own work against it. Collect twelve postings for GenAI or applied AI engineers and pull out the responsibilities that are not model work, because those turn out to be most of them: latency budgets, cost per request, evaluation, versioning, being on call. Now take your best demo and interrogate it against that list in writing. What does it cost per request today, in money rather than tokens? What is the p95 latency, measured rather than remembered? What happens at fifty concurrent users? What happens when the provider retires the model, which will be sooner than you expect? What is the plan when it says something wrong in front of a customer? Most of those will have no answer at all, and the absence is the finding. Write down the answers you have and mark the rest unknown, plainly, because those unknowns are your next four stages. Then map your own last three years against the responsibility list and mark each line have it, adjacent, or do not have it. Anyone who has run a service with a latency budget has more of this than they think.
Done when
- Twelve postings collected, with the responsibilities that are not model work extracted into one list.
- Your best demo audited in writing against that list, with every unknown marked as unknown.
- Cost per request and a measured p95 latency exist as numbers, or are explicitly recorded as unmeasured.
- Your own experience mapped line by line: have it, adjacent, or do not have it.
- A one-page brief exists naming the three gaps you will close next.
What you end up with
A written audit of your best demo against a real responsibility list, every unknown marked as unknown, ending in three named gaps.
If you get stuck
The usual failure is answering the audit from memory, which produces confident numbers that are wrong. If you have never measured latency, write unmeasured rather than roughly two seconds. The absence is more useful than the guess, because it tells you what to instrument first. The second failure is deciding the demo is fine because it worked when you showed it. It worked once, for one person, on a good day, and that is the definition of a demo.