CSE Roadmap

Core tier · about 4 hours · 10 parts

Part 4 of 10 · 2 min read

What the estimates measured

Capacity and cost for two workloads, how they scale, and how a first draft scores against the full rubric.

How to read the first table: both columns use the same formulas on different assumed inputs; compare the bottom row across columns to see which workload is more token-hungry per run.

Estimate (assumed inputs) Support platform Research agent
Runs per day 100,000 5,000
Peak runs per second 5.8 0.35
Concurrent runs at peak (Little's law) 231 312
Workers needed 12 32
Tokens per run (in / out) 28,350 / 1,500 2,301,000 / 30,000
Cost per run, no cache / cached $0.11 / $0.05 $7.35 / $1.35
Monthly model cost, no cache / cached $322,650 / $138,375 $1,102,950 / $202,635
Tokens per minute at peak vs 2M limit 10.4M (0.19x headroom) 48.6M (0.04x)

What to notice: the research agent needs fewer concurrent runs than the support platform's token demand might suggest. But each of its runs is nearly 100 times more token-hungry. That is why its headroom (0.04x) is worse even at much lower traffic.

How to read the scaling table: each row is the same support platform at a multiple of its tenant count; workers and tokens both grow, but not at the same rate as tenants.

Support platform at 1x / 3x / 10x tenants Workers Tokens per minute
1x 12 10.4M
3x 35 31.1M
10x 116 103.7M

What to notice: both columns scale a little faster than tenant count (3x tenants gives roughly 3x workers, not less). So there is no hidden economy of scale here. Growth has to be met with caching and routing, not assumed away.

How to read the rubric table: it is one snapshot (a typical five-minute first draft), not a comparison between rows.

Design review rubric (14 items) Covered by a typical 5-minute first draft
Requirements, queue and workers, streaming, cost controls, deploy 5 items
Missing async job API, durable execution, reliability, tenant isolation, versioning, observability and evals, safety and approvals, data and retention, incident response

What to notice: every one of the 9 missing items maps to a module already covered earlier in this level. The gap is not unknown material. It is material that a quick sketch skips past under time pressure.

{{svg:architecture|The support platform with the module that covers each box.}} {{svg:results|Capacity and cost estimates, scaling, and the review rubric.}}