G7 GovAI Grand Challenge

Presenters Jameson Jones-Doyle, Jonathan Nuss, and Tyler Kolody

0:01

[Jameson] My name is Jameson Jones-Doyle. I'm the CEO of Axon Technologies and I've been an entrepreneur for over a decade and specialize in an advanced technologies that help people and the communities that are they are a part of to achieve greater results.

0:39

[Jonathan] Hi, my name is Jonathan Nuss. I'm a cofounder of Axon Technologies, the company behind Innervation, and I've been involved in the early days working with Tyler and with Jameson on the legal and corporate strategy behind this venture. What excited me the most from the start was Innervation’s focus on multi agent systems and the opportunity to apply that technology in ways that make systems more transparent, more fair and ultimately better for our society.

1:07

[Tyler] Hi, my name is Tyler Kolody. I'm a founder and the CTO of Axon Technologies, the creators of the Innervation Multiagent framework that we're looking at right here. For the GovAI Grand Challenge, we focus on problem 4, which was clients have a wide variety of communication needs and interactions with the government can exacerbate existing barriers. Design a solution to communicate with people in a way that is accessible to a wide range of languages, cultures and abilities.

1:41

As a predominantly disabled-owned organization, this was something that we were very excited to tackle. Focusing on ableist language and debiasing of government material. We hope to make it easier and more effective for those in government to create material that is inclusive and have it be in such a way that it can be validated and is transparent so it can be adjusted, corrected and improved over time without overly burdening those in government in the process.

2:14

So, to get started, we're just going to do a little test run here. Right now, there are three different configurations, standards that it is measuring against, but these can be swapped out to be different languages, different standards, even focusing on racism versus ableism or any other sort of bias that you wish.

2:38

This gets spread into four different agents that we can see here, each one responsible for one of the standards or the framing agent, which does not have a specific standard, but it's looking at the high-level framing and how the assumptions the text is making at a high level that might be ablest. These are all then aggregated within this agent, which also has programmatic validation steps that while they can't detect ableism, if only,

3:11

they can validate that the output is of the right structure and has content that matches. So, this ensures that the earlier input from these upper agents is included or sort of attached as part of the output, as well as that the output is of a particular format and structure that makes it easy to parse and to understand.

3:36

So, we can see that this is running right now. Currently it's saving the attachments of the previous outputs to the final output before we'll give its final answer in a minute or two.

3:49

But every step of the process is entirely auditable, entirely transparent. There's no abstraction or obfuscation, and that sort of is at the core of what makes Innervation unique. We believe very strongly the AI shouldn't go everywhere. It should go in the places that it's supposed to, that it makes sense to you, and that it should be for the empowerment of the user. Even if it's doing a lot of the work, even if it's very automated, the user and those that the user is making decisions about have to be first and foremost in the design.

4:24

So, we see that it completed here and it actually failed the validation step on the first try. We can see that it got a validation failure and then it corrected itself and got a success on the next one, which is exactly what we want to see.

4:41

So from a technical perspective, we're using Django, Docker, and Kubernetes to build our build our back end and React is our front end, but users don't have to, or clients don't have to use our front end. We have a comprehensive REST API and documentation that is used in our own front end but is exposed to clients as they want. So, while we do have things like large text and colorblind modes that are very nice and very useful, if the situation calls for it to be integrated into an existing

5:21

interface, that is entirely doable. We meet the client and we meet the user where they are. We don't ask them to come to us.

5:30

So, it's entirely interoperable. You can also, of course choose any models that you wish, any suppliers that you want to use or use open models allowing for sort of like full customization, which is particularly important for privacy centric uses, which this isn't necessarily, but the government often needs to and must protect user data and we can sort of ascend to that.

5:57

So, we hope that this will make government work more efficient in that it, like there's a lot of dimensions of effectiveness or efficiency as a democratic institution, the Canadian government, it's critical that they are able to sort of hear from all constituents. And there's no segment of the population that it is made it harder to speak to and we hope that this will make it easier to sort of just by default be using inclusive language, be working in inclusive ways. And as for next steps, while it is a bit pie in the sky, so was this three years ago and here we are.

6:41

It would be a dream to have this actually used at the level of MPs and their constituents to make it easier for MPs to both understand what their communities at large are concerned with what problems and issues they have and want to see addressed. As well as being able to then pull out the specifics that individuals have given of their individual problems. Obviously anonymized, but to basically give you the best of the high level

7:11

and the individual scales of a democratic institution. So, to just sort of bring this home, we can see here that the output is exactly as we want. We have the original text. We have a detailed analysis and risk assessment of each of the changes that were made. And then finally, we have the actual de-biased content, which then can be served and used after it's been checked and sort of given the thumbs up by a human.

7:44

So, thank you so much for listening and I hope you're as excited as we are to continue this partnership.

Presenter Wolfram Bopp

0:01

[Wolfram] Hi everyone, my name is Wolfram Bopp. By day I work at BMW as an AI expert. As a hobby

0:09

and to learn, I take part in AI hackathons and charitable projects. I really enjoy diving into novel complex

0:17

problems and it's really nice to build solution that deliver real measurable value for the general public.

0:25

We choose to tackle the challenge,

0:28

design a solution that anticipates future needs and resource allocation to reduce waste and improve service by

0:37

ensuring resources are available where and when they are needed with infrastructure data.

0:44

So let's have a look.

0:47

[Unnamed narrator] Every year, governments spend billions managing roads and bridges. And every year, billions are wasted.

0:57

We repair infrastructure too early,

1:00

wasting money, or too late, forcing emergency repairs that cost 10 times more.

1:08

Most governments manage billion-dollar portfolios using static age-based depreciation

1:15

curves that simply guess when a bridge might fail. What if we could predict structural failure before the cracks

1:24

appear? What if we could optimize every single tax dollar? Introducing ResiliAlloc,

1:32

the G7's first sovereign infrastructure intelligence platform.

1:39

ResiliAlloc isn’t just a dashboard

1:43

It's a dynamic physics engine. We rejected the industry standard of using static spreadsheets.

1:52

Instead, we built a modern Lakehouse architecture. Our swarm engine ingests real time signals from seven global APIs.

2:05

We combine the static asset backbone from Odin with dynamic stresses. We pull

2:12

historical freeze thaw cycles from Open Meteo to model thermal shock. We pull

2:20

live traffic patterns from TomTom to measure structural fatigue and crucially,

2:28

we calculate medical critically identifying bridges that, if closed, would

2:34

block ambulances from reaching a hospital. But here is where ResiliAlloc wins.

2:42

competitors will tell you, "We can't predict risk because the open data doesn't have condition ratings."

2:51

We proved them wrong. We used a federated ground truth strategy. We

2:57

trained our AI on over 6,000 validated engineering records from the USA, the UK, and Quebec.

3:07

Using a Kaggle winning stacking ensemble,

3:10

we transferred this intelligence to local assets.

3:15

We implemented pseudo-labeling, allowing the model to teach itself the nuances of the local environment.

3:24

The result, we beat the government standard model by 26%.

3:31

We moved from guessing to precision engineering.

3:37

Let's look at the mission control dashboard. The system analyzes nearly 5,000 provincial assets. On screen, you

3:47

see a sea of red. Our AI has identified 140 critical assets that are facing

3:56

imminent failure due to high traffic load and thermal shock.

4:02

Watch what happens when we apply a 25 million dollar emergency budget. The AI doesn't

4:10

just pick random bridges. It runs a NEPSE optimization algorithm to select

4:16

the projects with the highest return of investment for public safety within budget. Take asset 4594.

4:26

The AI flagged this not just because it's old, but because the combination of

4:33

heavy trucks and severe freeze cycles makes it a ticking time bomb.

4:40

ResiliAlloc ensures we fix it today. Precision policy making.

4:46

This is the power of ResiliAlloc. With a single click, the optimization engine

4:52

analyzed nearly 5,000 assets and rejected 99% of them to find the most

5:00

important projects within an example emergency budget of 25 million.

5:07

These projects aren't just random choices. They represent the mathematical optimum, the highest possible reduction

5:16

in public safety risk for every dollar spent.

5:22

We utilize 96% of the available budget with surgical precision. And unlike

5:30

other platforms, ResiliAlloc isn’t a black box because we use rigorous machine learning driven by explicit

5:38

feature engineering. We can trace every single prediction exactly back to its root cause. That means total

5:47

transparency for stakeholders and it ensures that every dollar spent is 100% explainable.

5:58

But the dashboard goes further. It puts the power of what if analysis directly into the hands of policy makers.

6:08

You can adjust risk tolerance thresholds in real time, instantly visualizing how

6:15

changing safety standards impacts the budget.

6:21

ResiliAlloc transforms complex engineering physics into clear actionable policy. It ensures that

6:30

whether you have 25 million or 1 billion, every cent is deployed where

6:36

it matters most. This is how we build a resilient future.

6:44

[Wolfram] To address your other questions. We appreciate the constructive feedback and have addressed the areas of improvement

6:51

and next steps regarding explainability while mapping out clear next steps for data scalability and user adoption.

7:02

Regarding explainability, we updated our code and demonstration videos so user can now explicitly see the root cause

7:10

driving every risk assessment. Such as the impact of freeze-thaw is 61% and traffic loads is 23%.

7:19

Availability - while immediate expansion was constrained by the short-term public availability of broader G7 data sets,

7:29

our platform architecture is fully prepared to ingest and validate new country specific data as soon as it becomes available to us.

7:41

Regarding user adoption, meaningful integration into real world government workflows requires a deep understanding

7:48

of stakeholder needs. While the tight time frame did not allow for this feedback cycles yet, engaging directly

7:57

with end users to gather this critical input is the next phase of development.

8:05

Ultimately, our vision is to evolve this AI engine to seamlessly integrate into G7 systems, empowering governments to

8:14

build safer, more resilient, and equitable funded infrastructure.

Presenter Simone Vagnoni

0:01
Morning to everyone. I’m Simone Vagnoni, a PhD Researcher at CIRSFID-ALMA-AI, University of Bologna and the Ontology Engineering Group, Universidad Politécnica de Madrid. And I will thank my supervisors Monica Palmirani and Victor Rodriguez Doncel

0:16
and the G7 GovAI Grand Challenge in particular to give me the opportunity to present my work: a deliberation knowledge graph bridging institutional and civic democratic discourse through fora in time and space, using different AI tools.

0:34
We wanted to tackle the first problem statement – managing high volumes of information, starting from the democratic data divide.

0:46
The problem is that when citizens debate climate policy on social media, while parliamentarians discuss the same topic in formal session, these conversations remain isolated islands of democracy.

1:01
So, there are a lot of data, of deliberation data, from the official portals to forum online or participatory platforms like Decidim, Polis, and so on. But there is no way to analyze them across platforms.

1:21
This is a missing missed opportunity for democratic insights. Right now, citizens and institutions operate in silos.

1:31
So, we did the deliberation knowledge graph — a unified framework that connects arguments across platforms, participants across contexts, topics across institutions, evidence across sources—

1:45
and the interface to show the potential of this.

1:53
We based the knowledge graph on deliberation ontology divided in three models. The process model: how deliberation unfolds, stages, timelines, organizations, from agenda setting to decision-making. Participant model: who participates, citizens, experts, parliamentarians, organizations, roles and relationship across context. Then there is the argument model: what gets said? Arguments, positions, evidences,

2:24
support / attack relationships that help us to go further in the level of analysis using natural language processing a posteriori.

2:35
This can help to do argument mapping, policy detection to measure the quality of the discourse and the relationship between the different nodes of the network and the topic modelling, and so on.

2:54
So, we, from different data sources, extract the raw data, harvesting from web scraping to API connectors, connect them to the ontology and upon them, services like the one that I will show you right now.

3:13
You can see the citizen interface, very minimal, which there is a semantic engine where you can search for whatever topic you want to compare to the different

3:25
platforms and fora that you have here. Like Consultation Have Your Say from Europe, European Parliament itself, Polis platform, Decidim, and so on. So, if you search for example for voting,

3:41
then due to our semantic search, because we embed all the participatory platforms and processes in an embedding space using sentence-BERT, we can search for every word, not keywords searching but for semantic. And also sort in any case, in any way.

4:18
Here we have the semantic similarity score, the number of contributions, etc. Here for example are Canadian Electoral Reform.

4:29
Every node here is a contribution that is connected to every other node and has its metadata. Text, the date,

4:43
author that in this case it's pseudonymized. The agree and disagree count that it's particular for this platform. Here we can see a visual of the knowledge graph and here we have also topic modelling.

5:03
Since we mapped the semantic of the discussions, we can map across platforms, what are the more contributed and talked about topics. You can see here are some examples.

5:23
In the next version we are doing argument mapping to enhance the research of the quality of the discourse. Returning to the presentation, there are a lot, very lot, of different platforms and we want to connect them in a linked democracy fashion,

5:48
across them and through the institutions to map what citizens want to what will effectively be done.

6:01
The real-world applications are cross-platform argument tracking, democratic health monitoring, argument quality assessment

6:12
and bridging institutional civic divide. That is the main problem in my opinion of the nowadays democracy.

6:24
But we acknowledge some challenges. Of course, the arguments exist in context and not just the structure. So, extracting them out of context could lead to misinterpretation. There are problems of privacy and contextual integrity,

6:44
but because also the data is, we harvested data legally and legibly, but the

6:57
cross-platform linking raises new concerns. There is a possibility, for example, to hypersuasion. More data enables more sophisticated manipulation. It's a problem that nowadays we face, in particular with private platforms that do it for personal and private interest, but

7:20
also would be if we agglomerate data. Our approach is transparent limitations, ethical design principles.

7:29
So, the bigger picture is an epistemic democracy where citizens see how their voices reach institutions, researchers track democratic evolution, journalists follow argument chains, and parliamentarians understand public sentiment.

7:44
So, to end, I have a call to action. For researchers, to join us in expanding the ontology and data sources. For practitioners, implement deliberation standards in your platforms. And for citizens, demand transparent, connected democratic processes. And of course, participate, have your say.

8:05
Thank you for your attention and stay connected. And here are the QR codes of the platform and my profile on LinkedIn. Thank you.

Presenter Satoko Watanabe

0:00

Hello, we are team 1902. Today we'd like to present our proposal, GovGov - government data utilization for combating misinformation.

0:11

Here is our team. We are students from Waseda University, Japan. We all belong to the debate club at Waseda and as a new challenge we decided to try participating in a hackathon.

0:24

The challenge we aim to address is the issue number four. Clients have a wide variety of communication needs and interactions with the government can be challenging and frustrating and exacerbate existing barriers.

0:39

In today's digital landscape, misinformation is ubiquitous. We identified a critical problem. The proliferation of uncorrected false information which impedes discourse rooted in fact. To safeguard a healthy democracy, we must establish two

0:56

pillars. A mechanism for real-time fact checking on social media and a streamlined path to authoritative data.

1:05

In the era of advanced AI and pervasive connectivity, social media and generative AI have become primary information sources. However, these platforms present significant structural challenges.

1:20

For instance, rumors such as, "What are we going to do if the illegal immigration problem increases because of JICA?" “Apparently, it's easier for foreigners to receive welfare benefits.” frequently circulate on X. Users often attempt to verify these claims by summoning AI tools.

1:41

Easy at Grok fact check. Yet, these systems often yield inaccurate results. This leads to polarized, unproductive debates detached from reality.

1:54

Furthermore, an increasing number of users rely on generative AI as a primary search engine. However, these models often ingest unreliable sources such as personal blogs or AI-generated slop compromising the integrity of public discourse.

2:14

So how do we restore constructive dialogue? The answer lies in the provision of high integrity information.

2:22

The government must take a proactive role in comm in communicating directly with the public. A true government to consumer G2C communication model. To be effective this information must be both comprehensive and precise.

2:40

Japan's Administrative Program Review sheets represent the only official cross ministerial database covering all budgeted projects.

2:50

Given their exhaustive and verified nature, they hold immense potential as a platform to counteract misinformation.

2:59

To leverage this resource, we developed two products integrated with the review sheets. GovGov, a next integrated bot, and GovGovWeb, a web-based AI chat interface.

3:14

First, we introduce GovGov, the official government fact checking bot. When a user encounters suspicious information on social media, they simply mention at GovGov Japan to request a

3:26

verification. The Bot identifies the most relevant administrative projects from the review sheets and feeds their specific data into an LLM. This RAG system ensures the output is grounded strictly in official facts.

3:44

Next is GovGovWeb. Currently, citizens face a dilemma. Manual web searches are cumbersome, while generative AI responses are untrustworthy.

3:55

GovGovWeb addresses this by offering a suggestion feature for vague inquiries and restricting its knowledge base exclusively to government documents ensuring maximum reliability.

4:08

Please watch this one-minute demonstration. Here, the user asks about the Ministry

of Economy and Trade Budgets and AI responds to the question.

4:30

This is the GovGovWeb and user asks about the budget of the Digital Agency in Japan.

4:40

While we have launched GovGov and GovGovWeb as initial proofs of concept, we believe two further steps are imperative to modernize government communication.

4:51

First of all, AI SEO optimization and MCP server implementation, making government data machine readable.

4:59

Second of all, centralization of public records which integrating laws and council minutes into a single source of truth for chatbot responses. Our ultimate vision is a society where

5:13

citizens can collectively endorse effective policies and constructively critique those requiring improvement.

5:21

By utilizing review sheets as a foundation for evidence-based discussion, we aim to foster consensus building that transcends social division.

5:30

Thank you so much for listening to our presentation.

Presenter Paul Brown

0:00
We are from Permion presenting DSME. Arun Majumdar is our CEO and expert in AI. Paulo Moura authors Logtalk, which makes DSME run. Barry Evans makes the neuro symbolic AI work and I'm Paul Brown, the Chief Architect.

0:14
Public service must manage high volumes of information, often manually. Tools are needed to aid with ingesting, exploring and understanding unstructured documents. Permion's Digital Subject Matter Expert, or DSME, is a multilingual, agentic, trustworthy AI application for rapid assessment of corpora. As a Canadian policy analyst, I can use DSME to contribute to a world class regulatory sandbox for strengthening sovereign intent, economic prudence and legislative integrity.

0:44
Today, our corpus contains policy documents on two themes, Energy and Artificial Intelligence. Toward a national AI energy test harness and minimum energy performance standards for AI for fulfilling net zero emissions mandates.

0:59
With our corpus loaded, DSME provides two views that are useful for finding documents that might be of interest. So, if, for example, I was interested in national policy versus global policy, well, I can find the documents which discuss those things more frequently than the others.

1:18
We also have a similar view, but for entities.

1:21
So here we have found all the things that are discussed frequently within the documents and these are your candidates for the stakeholders.

1:29
We can also add in all of the ones that are related to a single source document for an exploratory view of the potential stakeholders across a whole corpus of policies.

1:41
To help an analyst understand policies across many documents, DSME will extract all of the deontic imperative sentences and then reduce them to an actor like “models” and “power”, an action like “require” and “process”, and an object like “laptops” and “inputs”.

2:02
To see where the information came from, if we hover over an action, you'll be given the paper and the sentence.

2:09
The actions are colour coded, so we have red for obligations, we have orange for recommendations and black for prohibitions.

2:18
Furthermore, actors can be added and removed at will to allow for exploration.

2:29
Policies can be quite long to read, therefore DSME provides summative information including key words, key terms, which are multiple words, long key entities, and there your candidates for stakeholders, as well as summaries. So this first summary here is derived from a structural analysis of the text and the second summary is derived from semantic similarity to the key words. If we provide our own key words like “AI”, “energy”, and “Canada”, then we can create a summary around that topic.

3:02
And that's created using our LGM, which is our method of extracting the information at its most relaxed setting of three. We can pass that to an LLM, so the LLM can rewrite it to make it a little bit easier to read, but that can introduce hallucination. That's why we provide this Trace scores view, which shows where the text overlaps from the LLM's answer on the right with the original text on the left. And you can see the semantic similarity between the sentences as well. This sentence here doesn't seem to have any overlap and so that could be a candidate for a hallucination.

3:39
Similar to doing a summary, you can also ask questions directly of the document. So we're going to ask if AI can help reduce energy consumption and we'll get an answer out. Now this is using our LGM to extract the text straight from the document and you can trace these sentences back to the document to see them in context by scrolling through the highlighted sentences.

4:01
And these are all directly taken from the document in context. Now if we introduce the LLM, there are different levels for different levels of relaxation and that will give you those answers. And again, you can check them for hallucinations as you can with the summaries.

4:21
The key thing a policy analyst would like to know from any particular document is what are the policies? Therefore, DSME extracts the deontic imperative clauses and finds an actor, an action and an object. It will give you the modality as well as the text of the sentence. It can tell if it's negated and if there's a condition, and you can trace that text back to find it within context of the document.

4:48
We also provide these filters so we could find all the negated ones to find the prohibitions, or we could find perhaps all of the obligations that have a condition.

5:03
This is to help an analyst quickly explore the policies within a document.

5:10
In order to help an analyst understand the domain of discourse of a particular document, DSME provides a couple of tools that provide structured information. So this particular one shows how strongly related particular key terms and key phrases are to one another and what they are related to. So AI is related to high impact, technology, energy, use, eligibility, and Canada.

5:36
This shows how strongly things are related, but not how they're related. How they're related comes from the proto-ontology. So we can see the Canadian government contributions have a part that's provincial and territorial, and we can trace this back of course to the source text to see where this information comes from. This will help an analyst explore the domain of discourse of a particular policy.

6:02
DSME provides a HTTP API that can be used to automate tasks for system integration, as well as a Python library for interacting with that API. In this example, we're going to use that Python library to do some automated triage, and so we create a temporary project and upload an AI and energy corpus that contains 90 sources from around the world. We're going to filter those sources based on interesting places and interesting topics using the keywords that are the synsemantic set to reduce that down to 35 sources. We're then going to find the location of each of those sources,

6:32
delete the temporary project and create a new one called North America AI and Energy, uploading all of the sources that we found to the new project. Now, you could continue with HTTP API or the Python library, but should an analyst wish to continue in the GUI, they'll find the project they're waiting for them.

6:51
Thank you for watching. We look forward to further collaborations with prime contractors, institutions, and G7 governments. The DSME is a multilingual, agentic, trustworthy AI application poised to strengthen our sovereign intent, economic prudence, and legislative integrity. For further details, do get in touch at permion.ai.

Presenters Ana Ohde, Nigel Cullumbine, Tor Gisvold, Zhiqian Huang

0:00

[Ana Ohde] I'm Ana Ohde, health economist. I make sure our solution delivers measurable economic

0:07

value and sustainable impact at population level.

0:13

[Nigel Cullumbine] I'm Nigel Cullumbine, Chief Operating Officer. With over 20 years experience in health and care leadership, including CIO and COO roles, I specialize in

0:22

translating operational challenges into practical digital solutions.

0:27

[Tor Gisvold] Hello, I'm Tor Gisvold, CTO. I lead our technology strategy and my task is to

0:34

build secure, scalable AI first and cloud first platforms.

0:40

[Zhiqian Huang] I'm Zhiqian Huang, CEO and co-founder. My background is in NHS operational crisis management and I pioneered the use of

0:48

process mining to uncover patient flow insights across health systems.

[Text on screen that is not spoken]

Together, we combine operational, technical and economic expertise to build an operational intelligence layer that helps governments allocate resources more effectively.

1:01

[Unnamed narrator] The G7 challenge recognizes the difficulty governments face in predicting needs and allocating resources effectively within complex

1:09

systems. In emergency healthcare, this challenge is visible every day.

1:18

Hospitals must deploy limited staff space and equipment in real time. Yet decisions remain largely reactive, often

1:27

relying on individual experience rather than on demand operational intelligence linked to measurable outcomes.

1:34

Our solution aims to address this problem.

[Text on screen that is not spoken]

Process Mining /noun phrase/ A data-driven, analytical methodology used to discover, monitor, and improve business processes by extracting and analysing knowledge from event logs within IT systems.

1:43

[Unnamed narrator] With process mining, we can analyze every patient flow individually within seconds. Something traditional analytics

1:51

struggles to do. Our proprietary approach to categorize these flows into structured routes is the foundation for intelligent resource planning. As an

2:01

example, we helped an A&E realize that the delays to blood processing was actually due to portering availability,

2:08

not laboratory capacity.

[Text on screen that is not spoken]

How can all hospitals access this insight?

2:18

[Unnamed narrator] Loganberry, our AI companion, allows operational teams to interrogate these insights in natural language. Loganberry

2:27

is tailored to each hospital and their users. It is user friendly regardless of their digital maturity and analytics

2:33

capabilities. It triangulates structured flow data, hospital standard operating procedures, national guidance, and the

2:42

local contextual information without users having to interpret dashboards.

2:48

We have also incorporated feedback from the judging panel by strengthening explainability and responsible AI guide rails. We have also removed language barriers for adoption worldwide.

3:05

The AI responses are now context aware and it's able to show its reasoning clearly. Our AI is configured to respond

3:12

within scope when asked a question outside scope and knowledge base. It does not replace human judgment. It supports it. Process mining data is

3:20

supplied with patient identities obfuscated by the hospital prior to processing. And our intelligence module

3:28

does not collect or process identifiable patient data.

3:33

We have partnered with an enterprise technology provider that gives us full control over our software stack and the data residency enabling alignment with

3:41

national legislation including GDPR and the EU AI act. Its architecture is secure and cloud native currently operating at scale in a regulated

3:50

industry. This gives us confidence that we can scale from a single hospital to national deployment while maintaining governance, security and compliance.

[Text on screen that is not spoken]

How can hospital use insights to improve resource allocation?

4:08

[Unnamed narrator] We have developed a real time operational intelligence prototype since our original submission.

4:19

This module enables staff to log operational issues as they happen with AI assistance to capture context consistently

4:27

rather than relying on fragmented messages or informal handovers. Our solution links each issue to its potential impact on patient flow. This

4:36

will be the first time that tested operational knowledge of how hospitals are actually managed day-to-day is captured in a structured form for

4:45

leadership teams. Our solution provides suggestions on what they need to consider before mobilizing additional resources. This creates visibility and

4:54

reduces the cognitive load of managing complex real time operational pressures.

4:59

AI proposes a status for the hospitals which can be challenged keeping executives firmly in control.

5:16

This is the prototype architecture for production deployment, enterprise controls and an intelligence layer are added including authentication,

5:24

governance, process mining models,

5:26

knowledge integration, and explainable AI. This operational intelligence is designed to feed into the Loganberry intelligence layer, strengthening

5:35

insight with real world context. Because it operates on timestamped operational data and configurable data residency

5:42

models, it can be deployed securely across jurisdictions.

5:47

It will now reflect how each hospital truly functions locally.

5:59

[Text on screen that is not spoken]

Captures Operational Context

Links Issues to Flow Impact

Tracks Perceived vs Actual Impact

Reduce Board-Level Cognitive Load

Codify Tacit Operational Knowledge

Thank you, Canada!

Merci, Canada!

Presenters Steven Tang, Akemi Izuko, and Oliver Chen

0:01

[Steven] Hi everyone. We are Steven, Akemi and Oliver, a team of students from the University of Alberta in Edmonton, Canada.

0:08

I'm Steven and on this project I worked on enhancing the user experience and agentic engineering.

0:16

[Akemi] I'm Akemi and I focused on design and project planning while ensuring our project complies with the highest level of privacy standards.

0:23

[Steven] What really brought the three of us together for this project is a shared frustration. As government services move online, citizens are required to navigate a complex digital landscape to access them.

0:31

While most welcome the shift because it streamlines the process, allowing forms to be filled out at home, it presents a significant barrier for those with accessibility needs. The digital world is inherently

0:47

visual, relying heavily on screens and text. This is a major issue. A recent study by Statistics Canada found that 75% of Canadians over the age of 55 have poor vision without correction.

1:01

[Akemi] Exactly. We're looking to address the accessibility issues people face when interacting with their government. We've all experienced how irritating it is to fill out government PDF forms.

1:09

The tiny boxes and small fonts create massive friction. This isn't necessarily a fault of the forum designers. They just have to pack a massive amount of information into a limited space, but it impedes access to public services for many citizens. This directly aligns with the

1:25

G7 Gov AI Challenge. The Challenge states clients have a wide variety of communication needs and interactions with the government can be challenging and frustrating and exacerbating existing barriers. While we were asked to design a solution to communicate with people in a way that is accessible to a wide range of languages, cultures, and/or abilities.

1:41

To answer this, we chose to target a widespread issue of visual impairments.

1:49

This is a critical barrier because filling out a PDF is often mandatory as a prerequisite for accessing government services. With ReadyFormAI, the visual interface effectively disappears and the form itself becomes a conversation. You simply speak, the system listens, understands, and responds with absolute clarity.

2:05

This is a seamless integration of voice and data creating an experience that is not just accessible but truly effortless. Let's see it in action.

[ReadyformAI] Hello. Welcome to ReadyFormAI.

(In French)

Hi. Welcome to ReadyFormAI - French

Bonjour. Bienvenue sur ReadyFormAI - français

2:19

Hi. Welcome to ReadyFormAI.

I'm here to help you fill out your grain receipt. I can help you in French and in more than 40 other languages. To start, I need some information about your grain transaction. Could you give me your name and your producer address?

[Oliver] Yes, my name is Oliver Chen.

[ReadyformAI] Perfect Oliver, I've noted your name. What is your full address, and what types of grain are your delivering today?

Bonjour. Bienvenue sur ReadyFormAI.

Je suis là pour vous aider à remplir votre reçu de grains. Je peux vous assistez en français et dans plus de 40 autres langues. Pour commencer, j’aurais besoin de quelques informations sur votre transaction de grains. Pouvez-vous me donner votre nom et adresse de producteur.

[Oliver] Oui je m’appelle Oliver Chen.

[ReadyformAI] Parfait Oliver, j’ai noté votre nom, quelle est votre adresse complète et quels types de grains livrez-vous aujourd’hui.

(In English)

2:42

[Oliver] Uh, let's switch back to English.

2:54

[ReadyformAI] Got it. Switching back to English. I have your name as Oliver Chen. What's your complete address and what type of grain are you delivering today?

3:02

[Oliver] Uh, 123 Farm Lane and I'm delivering wheat.

3:09

[ReadyformAI] Perfect. I've added your address and set the grain type to wheat. Uh, what's the delivery date and location? And do you have a contract reference number?

[Oliver] Uh, let's go to the weight section.

3:22

[ReadyformAI] Here we are in the scale records section. I'll need the weight measurements for your wheat delivery.

3:26

What's your gross weight, vehicle weight, and unload weight? All in tons, please.

[Oliver] That should be 5, 10, and 20.

3:36

[ReadyformAI] I've entered those weights, but I want to double check. Typically, the gross weight should be the largest number, total weight of vehicle plus grain, and vehicle weight should be the smallest, empty vehicle weight. You mentioned 5,

3:48

10, and 20 tons. Could you clarify which is which? Is it gross weight 20 tons, vehicle weight 5 tons, unload weight 10 tons?

[Oliver] Oh, yeah. Right.

[ReadyformAI] Perfect. I've corrected the weights.

4:01

Gross weight 20 tons, vehicle weight 5 tons, and unload weight 10 tons. The system has calculated your net weight automatically. What's the dockage

4:09

amount? And do you have a scale ticket number?

[Oliver] As the user speaks, the AI navigates complex tables, autoscrolls to the exact field they're filling out, and instantly updates the text. Users can easily verify the inputs because the buttons are big, and the AI will read back any field if asked.

4:24

Once done, the user clicks the button and the filled out form is exported digitally as a standard PDF ready for the government.

4:34

Back to you, Akemi and Steven.

4:37

[Akemi] As you saw in the demo, what makes ReadyFormAI exceptional is how we handle complex PDFs and multilingual voice navigation while keeping things fully private.

4:45

Since government forms are packed with PII, personally identifiable information, in our initial submission,

4:52

we relied on standard open weight models that weren't fully isolated. Today, our MVP runs entirely on SOCK 2 compliant compute. We use modal as a proof of concept, demonstrating that this entire pipeline can be easily and securely deployed directly onto government servers.

5:09

On the technical side, making this fast, private, and capable was a fun challenge. We built our self-hosted pipeline using Pipecat, a provider agnostic framework for multi-stage LM pipelines.

5:18

For speech to text, we chose NVIDIA's parakeet mode. It's incredibly light on compute, has the lowest board error rate among open weight models, and understands 25 languages. For text to speech, we used Quen 3TS, which supports 10 language outputs, including all official G7 languages.

5:37

[Steven] Imagine an onboarding experience where language barriers simply fade away. You challenged us to build a tool for everyone to reach a truly global demographic, and we took that to heart. Today, ReadyFormAI brings a seamless, deeply human experience to your users.

5:55

Now speaking all the G7 official languages and many EU languages, the experience feels like magic.

6:03

Our voice agent warmly greets the user, cycling through the G7 languages. The moment they reply, it instantly recognizes their native tongue and creates a natural, frictionless conversation.

6:15

It even adapts on the fly if they switch languages mid thought. The AI effortlessly follows along, staying perfectly in sync. While the user enjoys a flawless, localized experience, our backend quietly does the heavy lifting.

6:30

Instantly translating and formatting their spoken words into standardized PDF fields. It's truly a global solution and as simple as having a conversation.

6:42

To wrap it up, we want to address two key questions. Akemi, how does our solution make the government work more efficiently?

6:50

[Akemi] Ultimately, governments want equitable accesses to their services. Without ReadyFormAI, visually impaired citizens face an unfair choice.

6:58

Either sacrifice their independence by asking a public service agent to read and fill out a form for them or just give up and not access the service at all.

7:06

Because PDF forms are the gateway to almost every public service, our solution significantly improves access, empowers citizens, and frees up government support workers. This leads to the second question. Steven, what's the ideal state of our final solution?

7:20

[Steven] Yeah. So, while our web extension removes a lot of friction already, the user still has to intentionally seek it out.

7:27

Our ideal next step is for governments to adopt ReadyFormAI directly into their web infrastructure.

7:34

So, when a citizen clicks on a PDF form on a government website, they should have the option to directly open it in ReadyFormAI.

7:44

Also, while we use the secure cloud for our MVP to ensure absolute privacy, our code is completely open and optimized to run on small, inexpensive GPU servers hosted on sovereign government infrastructure.

8:00

[Akemi] We believe AI shouldn't just be a chat box on a homepage. It should look be a tool that actively supports human endeavors. Thank you for your time and we look forward to seeing the G7 adopt solutions like ReadyFormAI.