Nexum

Inspiration

Construction projects in Colombia run on fragmented information. Budgets, schedules, scope documents, and supporting files often describe the same reality in different ways, which makes critical supply risk hard to detect early. We were inspired by a simple question: what if a project team could upload the documents they already use every day and immediately understand which supplies might create budget overruns or schedule delays?

That became the foundation for Nexum. We wanted to build a system that turns messy construction documentation into actionable supply intelligence before risk becomes visible too late.

What it does

Nexum is a construction supply-intelligence platform.

It ingests project inputs such as:

budget baselines
work schedules
contractual scope documents
supporting project files

Then it uses an agentic workflow to:

extract candidate supplies from those files
qualify which ones are real project-relevant supplies
preserve semantic meaning at the supply level
use tools such as web search and forecasting services to enrich supply monitoring
connect the useful subset to a monitoring layer
surface executive risk signals such as critical supply alerts, exposure, and project-level monitoring

The goal is to help project managers identify scarcity risk early enough to take preemptive action before it becomes a cost overrun or a schedule delay.

How we built it

We built Nexum as a Google Cloud-native agent system:

Gemini on Vertex AI for model reasoning
Google ADK to orchestrate the live supply-intelligence agent
MCP as the tool/runtime boundary
Arize Phoenix for tracing, inspection, and iterative improvement
Cloud Run for deployment
Cloud SQL for persistence
Next.js + TypeScript for the web application

At runtime, the architecture is: Nexum Web -> nexum-api -> Google ADK agent -> MCP-connected tools/data -> Arize Phoenix

The agent reads project inputs, extracts supply candidates, qualifies the real ones, preserves semantic supply classes, and powers the executive signals shown in the UI.

Challenges we ran into

One of the biggest challenges was that construction inputs are messy by nature. The same supply can appear differently across schedule files, budget lines, and descriptive documents, which makes deterministic mapping difficult.

Another challenge was deciding where the real product bottleneck was. At first, it was tempting to focus on forecasting and downstream insights, but Arize Phoenix helped us see that the more urgent problem was actually supply extraction itself. By tracing the workflow step by step, Phoenix showed us that if extraction and semantic classification were weak, then the rest of the pipeline could look complete while still being misleading.

That changed our priorities. Instead of only increasing coverage, we focused on making supply retrieval and classification more truthful first, then layering forecasting and monitoring on top of that stronger foundation.

Accomplishments that we're proud of

We are proud that Nexum is no longer just an idea or a prototype-level interface. Thanks to Google Cloud and Arize, we made a significant step toward turning it into a real production platform that can help the construction industry operate more efficiently.

What we learned

The biggest lesson was that agent engineering needs observability. Phoenix was not just a logging layer for us; it became the mechanism we used to understand where the agent was correct, where it was overconfident, and where it was hiding uncertainty.

What's next for Nexum

The next step is to deepen the supply-intelligence layer and move from a strong hackathon foundation toward real adoption.

We want to:

improve semantic classification coverage across more supply families
strengthen price-source linkage for the supplies we already detect
expand monitoring beyond coarse market categories
make scarcity and forecast signals even more actionable for project managers
continue using Arize Phoenix to measure and improve agent truthfulness over time

We are also proud that Nexum is already generating interest beyond the hackathon. We have started conversations with leads and potential clients, which gives us confidence that this is solving a real industry problem and not just a technical exercise.

Longer term, we see Nexum becoming a decision-support layer for construction operations: not only identifying critical supplies, but helping teams act earlier and with more confidence across budget, procurement, and execution.

Beyond construction, we also see Nexum evolving into a broader multi-industry platform by incorporating agents specialized for other sectors. The same core approach—document understanding, supply intelligence, and monitored risk signals—can be extended to industries with similarly complex operational and procurement challenges.