An AI research organisation
DeliveredFrontier model evaluation and alignment
Evaluation of state-of-the-art LLMs on long-horizon engineering work, with grading of agentic safety, honesty and scoping.
Behaviour grading
Agentic safety
Blast radius proportionality
Honesty
Says what it knows, and what it does not
Scoping
Stays inside the task it was given
Deference
Respects the user's constraints
Simulated interface. Not client data.
What the work covered.
Step 01
Long-horizon task evaluation.
Evaluated state-of-the-art LLMs on software engineering tasks of up to 30 turns, covering system investigation, root cause analysis and refactoring.
Step 02
Code and architecture review.
Reviewed model-generated Python for API compatibility, memory management safety and regression test coverage.
Step 03
AI behaviour grading.
Graded agentic safety (blast radius proportionality), honesty, scoping and deference to user constraints.
Step 04
Expert rationales.
Wrote citation-backed rationales comparing multi-turn runs, setting senior engineer baselines for autonomous debugging.
Evaluation run · up to 30 turns
Simulated interface. Not client data.
Evaluation run · up to 30 turns
Step 01
Long-horizon task evaluation.
Evaluated state-of-the-art LLMs on software engineering tasks of up to 30 turns, covering system investigation, root cause analysis and refactoring.
Model-generated patch
Step 02
Code and architecture review.
Reviewed model-generated Python for API compatibility, memory management safety and regression test coverage.
Behaviour grading
Agentic safety
Blast radius proportionality
Honesty
Says what it knows, and what it does not
Scoping
Stays inside the task it was given
Deference
Respects the user's constraints
Step 03
AI behaviour grading.
Graded agentic safety (blast radius proportionality), honesty, scoping and deference to user constraints.
Run A
Run B
Expert rationale
Citation-backed. Senior engineer baseline.
Step 04
Expert rationales.
Wrote citation-backed rationales comparing multi-turn runs, setting senior engineer baselines for autonomous debugging.
Simulated interface. Not client data.
- What we built
- Evaluation and alignment work on frontier language models, focused on long-horizon engineering, root cause analysis and LLM behaviour alignment.
Why it matters to our clients
The same safety, honesty and scoping checks govern every agent Aelix Echo builds. An agent that acts out of proportion to its task, overstates what it knows, or ignores a user's constraint fails our review before it reaches a client.
How our agents are builtDelivered under a confidentiality agreement.
More case studies
A customs import and export team
Import and export platforms for a customs team
A medical device manufacturer
Accounts payable automation on the Power Platform
A medical device manufacturer
End-to-end document automation on n8n
A management consultancy
Recovering, migrating and extending a recruitment platform
A real estate client from Dubai
Routing new CRM leads to the right agents
Start with one process.
Tell us the process that takes up your team's time and the systems it touches. We will show you how an agent or automation would handle it.