Challenge

Add an AI Agent that keeps output accurate, consistent, following strict editorial standards (neutral, precise, no exaggeration).

Discovery

We started by learning:

Onlyfacts domain and the climate areas it covers
How users currently consume data (charts, tables, articles)
How the data is stored and how the website serves it today

The pilot (POC)
To help Onlyfacts quickly assess initiative, we built a fast proof of concept: an 'Everyday' Onlyfacts agent focused on queries in the Emissions and Transport sectors. The pilot integrated directly with their existing database and returned answers that included visualisations similar to what users already saw on the website.

How we helped them

Built two AI agents, one customer-facing answering direct queries. Another, deep research agent to design complex reports.

Built in a few weeks with tight iteration loops, the pilot validated the direction to proceed.

Discovery

We started by learning:

Onlyfacts domain and the climate areas it covers
How users currently consume data (charts, tables, articles)
How the data is stored and how the website serves it today

The pilot (POC)
To help Onlyfacts quickly assess initiative, we built a fast proof of concept: an 'Everyday' Onlyfacts agent focused on queries in the Emissions and Transport sectors. The pilot integrated directly with their existing database and returned answers that included visualisations similar to what users already saw on the website.

Built in a few weeks with tight iteration loops, the pilot validated the direction to proceed.

OUTCOMES

Scalable infrastructure, realtime analytics, and continuous content delivery.

Everyday Onlyfacts agent
Answers day-to-day, single-area queries (e.g., 'Petrol vehicle registrations in Victoria') using direct database access and the right visualisations.

Deep Research agent
Handles multi-step questions that require a sequence of tasks over sizeable data, followed by a crisp synthesis.

Example:
'Provide a comprehensive analysis of how Australia is progressing toward net zero emissions - compare national trends with state performance, identify the biggest contributing sectors, and assess whether current trajectories will meet 2030 targets.'

Everyday Onlyfacts agent
Answers day-to-day, single-area queries (e.g., 'Petrol vehicle registrations in Victoria') using direct database access and the right visualisations.

Deep Research agent
Handles multi-step questions that require a sequence of tasks over sizeable data, followed by a crisp synthesis.

Example:
'Provide a comprehensive analysis of how Australia is progressing toward net zero emissions - compare national trends with state performance, identify the biggest contributing sectors, and assess whether current trajectories will meet 2030 targets.'

Thoughts from the client

"The Avesta team doesn’t just build software — they build trust. The speed and quality of our relaunch was beyond expectations."

— <First name (Initial), Designation>

"The Avesta team doesn’t just build software — they build trust. The speed and quality of our relaunch was beyond expectations."

— <First name (Initial), Designation>

Snippet from a Generated Report

How we engineer for the long-term

Since minor regressions can break trust (wrong numbers, wrong chart type, missing citations, or off-brand wording), we built repeatable evals and regression tests.

These checks covered factual accuracy, brand voice constraints, numerical consistency, citation presence, and chart-rule compliance before changes shipped. This also helped us evaluate and use newer LLMs as they were released, passing on the cost advantage for the same quality to the Onlyfacts team.

Services

Safe AgentOps with guardrails and Evals

Brand-safe responses: outputs tuned to Onlyfacts’ editorial constraints, with a focus on precision and neutrality.
Grounded data access: answers produced from direct database queries, backed by a robust integration layer.
Rule-driven visualisations: chart rendering that follows clear data-viz principles and avoids inconsistent chart-type choices.

Context engineering for the deep agent design

Curated data access for efficient context: direct database calls were replaced with an MCP server that strips unnecessary fields and returns only what the agent needs
Planning-linked context summarisation: prior context summary at planning checkpoints, reducing buildup while preserving decision-relevant info
Controlled context pruning during execution: older context was removed after defined steps to prevent context rot without impacting output quality