Technical

How we build, for whoever is vetting us.

Each answer starts in plain language. Open "Go deeper" for engineering detail, and again for the specifics. If something you need to know isn't here, ask us.

Accuracy and testing

How do you know the AI's answers are right?

We build a test set from your own examples and measure the tool against it before anyone relies on it, and again whenever anything changes.

Go deeper
  • Ground-truth labels created with your subject-matter staff, not guessed by us.
  • Measures chosen to fit the task, whether that's catching things, extracting numbers or drafting text.
  • Quality targets set after a first look at your real data, rather than promised up front.
  • An evaluation runner you keep, so you can re-test after any prompt, model or data change.
Deeper still
  • Tuned to what matters most for the task, for example catching every important change even at the cost of some false alarms for a person to review.
  • Every change is compared with a baseline before it ships.
  • Prompts and settings are versioned and re-tested whenever they change.
  • Automated grading is always checked against human judgment.
What happens when the AI is unsure?

It says so, and a person decides.

Go deeper
  • Human review queues for anything below a confidence threshold or above a risk threshold.
  • Answers cite the source passage or page, so a reviewer can check in seconds.
  • Tools are allowed to abstain rather than guess.

Data and privacy

Where does our data go?

Only where you've agreed it can, written down before we start.

Go deeper
  • A data-flow map for each tool: what's sent, to which service, and why.
  • Commercial API access rather than consumer chat apps, with each provider's training and retention terms confirmed for your use.
  • Least-privilege access and separate credentials for every client.
  • Personal data redacted or pseudonymized before it reaches a model, where the task allows.
Deeper still
  • Tools can run in your own cloud and provider accounts, so data never passes through ours.
  • Logs kept free of personal data; retention settings reviewed per provider.
  • Anything we held is deleted at handover.

Cost

How do you keep AI costs from running away unchecked?

We estimate running costs before we deliver, design to a budget, and give you guidance on keeping those costs in check afterward.

Go deeper
  • Per-task cost measured on real samples during design, not estimated from list prices.
  • The smallest model that meets the quality bar, with cheap filtering steps before expensive ones.
  • Usage caps and alerts on every provider account.
Deeper still
  • Cost per document or decision tracked in every evaluation run, next to quality, so tradeoffs are explicit.
  • Standard efficiency techniques for repeated and non-urgent work.

Models and architecture

Which AI models do you use?

Whichever fits the job and the budget, and we keep you free to switch.

Go deeper
  • We work regularly with Anthropic, OpenAI and Google models, and with open-source models for specific use cases.
  • Provider-agnostic interfaces, so changing models is a configuration change.
  • Switching is a measured decision: the evaluation suite says whether the new model is better.
What do you actually build?

Small, well-made tools that fit how your team already works.

Go deeper
  • Document pipelines: extract, normalize, validate and store.
  • Assistants that answer from your documents, with citations.
  • Triage and monitoring workflows that route only what matters to people.
  • Agents with explicit tool permissions and human approval steps.
  • Front ends where your team already works, often a shared spreadsheet or a simple web app.
Deeper still
  • Structured model outputs validated against a schema.
  • Re-runnable steps, so a failed job resumes rather than restarts.
  • An audit trail of every model decision.

Ownership and handover

What do we have at the end?

Everything we built, and the confidence to run it.

Go deeper
  • Code in a repository you own, with integration and maintenance instructions.
  • The evaluation suite and its test data.
  • Everything set up in your own accounts.
  • Training for the people who'll run it, and a defined support window after handover.

Proof: our own product

How the Great Heron Guide was built

A pipeline that turns thousands of corporate sustainability report PDFs into a structured, queryable GHG emissions dataset, with AI making expert-level judgments about which reported numbers to use. Extraction is built on Anthropic models.

Go deeper
  • Accuracy: measured against human-verified ground truth, including an independent holdout set the pipeline was never tuned on. On validation: about 96% precision and about 92% recall, under a strict definition of a correct value. When in doubt it returns "not found" rather than a guess.
  • Chat: the assistant that answers questions about the data is tested separately, with its own set of evals, so the values and calculations in its answers are checked against the dataset. When the data isn't there, it says so rather than guessing.
  • The model reads, code decides: the model reads the reports; the arithmetic happens in code, and every value is checked against its source.
  • Cost: each report is narrowed to its emissions-relevant pages before the model reads it, which improves both accuracy and cost.
  • Scale: thousands of companies across every sector; well over a million pages of reports spanning more than two decades; tens of thousands of emissions values, each traceable to its source.
great-heron-guide.com →

A question we didn't answer?

Send it over. Technical reviewers are welcome on the first call.

Get in touch