Continuum Assistant

Answers from your documents. Checked before you see them.

Your specifications, contracts, reports, and test records already contain the information your teams need. We make that information easier to access with an assistant that answers from your documents, cites the relevant sections, and clearly indicates when the information isn’t there.

How an answer gets from source to response

Answered from your documents

Your question is answered by reading the documents you supplied, and nothing else. The model cannot reach the internet.

Audited against the source

An LLM-as-a-judge checks the answer back against the passages it came from, claim by claim, and marks each one supported or not.

Scored for your application

Your own criteria define what a successful result looks like. Instead of relying on generic benchmarks, the assistant evaluates answers against your standards and sends anything that falls short back for correction.

Shown, or refused

A passing answer is supported by the relevant section of the source document. When a question requires an engineer’s judgement, the assistant flags it rather than making an unsupported assumption.

The NAPA airfield report assistant.

Learn about the published FAA / NAPA study on advanced technologies for airfield pavements. Ask it anything about the 51 technologies, the scoring, the gaps or the roadmap.

Reads only the report and its appendices. Nothing from the open internet.

Cites the report section under every answer.

Refuses questions outside the report and requests for personal data.

Tested before release: 29 of 29 on a question set the project team approved, graded by an independent model.

Does not yet audit each answer at answer time. That layer is in the pipeline above and is offered in pilots.

Where the assistant fits into your work

Answer the contract questions my inspectors and contractors keep asking.

Load the contract, specifications, and standard tables. The assistant answers straightforward questions with the relevant clause or table reference and a direct link to the source. Questions that require judgement are referred back to the appropriate engineer. It does not make approval decisions or approve submittals.

You give: The document set and twenty real questions
You get: A hosted assistant, tested against those questions
Timing: We scope it with you

Find the holes in my specification before a contractor's AI does.

The same pipeline run the other way. It reads your language the way a cost-minimising bidder would, lists each clause that can be read two ways, quotes it, and states the reading you did not intend. A person decides what to fix.

You give: One specification or contract package
You get: A ranked list of ambiguities with the clause quoted
Timing: We scope it with you

Tell me the state of practice on one topic, with citations.

Name the topic. We assemble the corpus from your own reports, the national research and the screened open literature, run the summary and the audit, and return a short brief with every claim marked supported or not and its scope stated.

You give: A topic and any reports of your own
You get: A cited brief with verdicts on each claim
Timing: We scope it with you

A focused pilot to test the assistant on your own work. Up to 50 documents, an agreed test set, and six weeks including fourteen days of hands-on access. The scope and fee are fixed, with known limitations disclosed upfront.

Scope a pilot

Where your documents live, and how they are handled

Nothing installs

Your team opens a link and gets started. No software installation or account management required.

Your corpus, removable

The documents live in a corpus dedicated to you. Replacing or removing one is a two-step command that takes minutes.

Anonymised logs

We keep the question, the sections used, and whether an answer was found. No names, no IP addresses, no personal data.

Your cloud, if required

Continuum hosts by default. For sensitive corpora the same assistant deploys inside your agency's own cloud tenant.

Pavement engineers who ran the AI behind the FAA airfield study.

Continuum Infrastructure Solutions is a small, minority-owned pavement engineering firm in Raleigh, North Carolina. We provide asphalt materials evaluation, pavement and overlay design, performance prediction software, data analytics, and on-site training, using mechanistic test methods and modeling to predict long-term pavement performance under real traffic and climate conditions. Founded 2024. Clients include FHWA, FAA through NAPA, and state and international agencies.

For NAPA and the FAA we screened the published literature on advanced airfield pavement technologies, evaluated 51 of them in depth, and delivered the report and roadmap that is now public. The same team built the assistant above. We know what a job mix formula says, so we know when an answer is wrong.

10,677abstracts screened

3,469full publications read

51technologies evaluated

Noor Saleh, Ph.D., and Y. Richard Kim, Ph.D., P.E. Continuum Infrastructure Solutions, Raleigh, North Carolina.

Frequently asked questions.

Do I need an IT team?

No. Your people open a link in a browser. There is nothing to install, no account for your IT department to create, and no change to your network. The one exception is if you want the assistant to run inside your own cloud account, which does need your IT and is scoped separately.

Why not just use ChatGPT or Claude?

They are very powerful but they can answer from everything they have ever read or add from the internet, and they will answer almost anything you ask. This is not a chat box over a general model. It answers from the documents you supplied and nothing else, shows the section every claim came from, and is built to say so when the documents do not support an answer.

Before an answer reaches you it is checked against the passages it was drawn from, and a second model reviews it and marks it supported, partly supported, or insufficient support. Below the threshold it goes back for correction rather than out to a person.

Our assistants still run on commercial foundation models. What we build is the workflow around one: what it is allowed to read, how it searches, and what happens to its answer before you see it. We orchestrate that workflow, and the system is prompted and tuned by engineers who work in this field, so it knows what it should answer and what it should refuse.

What do I need to give on my side?

The documents, and the revision of each one you want used
Twenty real questions with the answers you expect, including the ones it should refuse
Written permission to process the material
One engineering contact to agree the questions and review the results

Before any of that, representative samples, so we can both confirm the documents are suitable before anything is signed.

What do I get at the end?

A hosted question-and-answer interface
Answers with source references, working links into the document, and an indicator of how well the source supports each one
A validation summary covering the twenty agreed questions, including the ones it refused and the cases with insufficient evidence
A 60-minute walkthrough, a written guide to what it can and cannot do, and an export of the questions and results

What if our data cannot leave our own environment?

The same assistant can be deployed inside your own cloud account, or on your own infrastructure. That is assessed and priced separately, because it changes who runs and pays for the infrastructure.

One caveat we would rather state than have you discover: unless a self-hosted model is used, the final answer step still calls an external model provider. We can evaluate the use of open-source models or other arrangements when scoping.

What if I want our documents out of Continuum's infrastructure?

The corpus is dedicated to you, and removing it is a two-step command that takes minutes. Replacing a single document is the same operation.

What remains is the anonymised log: the question asked, which sections were used, and whether an answer was found. No names, no IP addresses, no personal data. That can be deleted as well, on request.

How long does it take?

It depends on the documents and on what you want it to do, which is why we scope it with you first. A first pilot runs from six weeks, starting once the signed scope, the documents and the reference questions are all in hand.

Most of that period is build, ingestion and testing. The last fourteen days are hands-on access, so your team reaches its own view of the assistant before final handover. Larger or harder sets, scans, complex tables and figures, take longer.

How are the answers actually checked?

Retrieval searches two ways at once. By the words in your question, and by what the question means, so the right clause is found whether or not it happens to use your vocabulary. The candidates are then re-ranked before anything is written.
The draft is checked against its own sources. References against the agreed document list, and numbers, units and specification references against the passages they came from.
A second model reviews it against the retrieved passages and marks it supported, partly supported, or insufficient support.
Anything below the threshold goes back for correction rather than out to a person.

These checks can miss errors. A number can be quoted correctly and still refer to the wrong material, and a source can itself be out of date. The indicator describes the automated assessment, not the probability that an answer is correct. Your engineers verify against the original documents before using anything in engineering work.