Cut
Cost/latency report
Baseline, levers, quality on the frozen set, £ per task.
AI Performance Optimisation
TTFT, TPOT, cache, batch, route — cut cost and latency with quality inside noise, and prove it on a frozen set.
Created by Baljeet Dogra
Course Objectives
- Instrument a request until you know where the milliseconds go.
- Cut cost about in half on a real workload without measurable quality loss.
- Choose cache, batch or route with numbers, not slogans.
- Load-test the path you will actually ship.
What you will produce
Test
Load profile
p50/p95, saturation, the knob you will turn first in an incident.
4 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 Measure Week 1
TTFT, TPOT, queue, retrieval, network. A table before a lever.
- Breakdown
- Tracing
- Tokens
- Queues
02 Levers Weeks 2–3
Prompt cache, semantic cache, batch, trim, cascade, spec decode if it earns it.
- Cache
- Batch
- Route
- Trim
03 Prove it Week 4
Frozen eval, load test, capstone report.
- Quality
- Load
- SLOs
- Capstone
Who this is for
Engineers with a live LLM feature
The bill and the p95 are both too high. You need levers.
Performance engineers
You know web perf. Tokens are a different budget.
Prerequisites
- You have an LLM-backed path to measure — or we provide a lab system
- Comfortable with logs and dashboards
Not a fit if
- Premature optimisation on a demo
- Kernel-level CUDA work — out of scope
Related: Applied GenAI Engineering · LLM API Integration
Questions
GPUs required?
Only if you self-host. Managed APIs are enough for caching, routing and trimming labs.
Is 50% a promise?
It is the lab target on the provided workload. Your production number is evidence, not a slogan.
Ready to start?
Four weeks to a measured cut in cost and latency. Create an account to enrol.
Enrol now