What’s included
What’s included
Quality baseline
Defensible metrics on accuracy, groundedness, safety and latency you can report to customers.
Built with
- Amazon Bedrock Evaluations
- Amazon Bedrock
- Amazon CloudWatch
Why it matters
Why it matters.
Without evaluation you cannot tell an improvement from a regression, so every change becomes a debate about vibes.
Context engineering (what you send, in what order, to which model) is usually the single biggest lever on cost per request.
Teams routinely send every request to their most expensive model. Routing by task shape typically cuts spend sharply without hurting quality.
What’s included · how it runs
Baseline. Harness. Optimise. Handover.
Step 01
Baseline
Sample real traffic, agree metrics and measure where you stand today.
What’s included
Quality baseline
Defensible metrics on accuracy, groundedness, safety and latency you can report to customers.
Step 02
Harness
Build golden datasets and automated evaluation into your pipeline.
What’s included
Evaluation harness
Golden datasets from real traffic plus automated scoring on Bedrock Evaluations, wired into CI.
Step 03
Optimise
Redesign context and routing, proving each change against the harness.
What’s included
Context engineering
Prompt structure, retrieval, history, memory and token budget redesigned against the harness.
Model benchmarking and routing
Compare models on your workload and route by task so cheap models take the easy traffic.
Step 04
Handover
Leave the harness, dashboards and the tuning playbook with your team.
What you walk away with
- Automated evaluation harness in CI
- Quality baseline and ongoing scorecard
- Context and routing redesign
- Measured cost per request reduction
Case studies
Customers we’ve delivered this for.
Draft · pending HES approval
Case studyHuman eSources
Make the hiring call with AI.
How Human eSources turned nearly three decades of talent science into AI that explains every job match, live on Amazon Bedrock behind TruTalent, its new recruiting platform.
- Industry
- HR tech, talent assessment
- We built
- Talent Scout AI
- Runs on
- Amazon Bedrock
Keith Gaudreau
VP Product Strategy
Case studyMould Detect AI
Computer vision for mould detection.
[How the client got the outcome: what Assembly built and where it runs. One sentence.]
Talk to an engineer[Client logo][First Last][Role]
Case study[Client]
[The promise] [in 2 lines].
[How the client got the outcome: what Assembly built and where it runs. One sentence.]
Talk to an engineer[Client logo][First Last][Role]