MK Solutions

The Northwind assistant greets Sarah by name and offers quick actions. It only reads what she types, never her account details.
Asked about her bill, the assistant shows a card from the billing system explaining the amount and why it went up.
The full breakdown: six months of usage, every rate and charge, the July price rise, and her balance.
Sarah types in her meter reading and billing recalculates on the spot, bringing her bill down from £169.60 to £132.01.
Still not solved, so a case opens with the whole chat attached. Sarah won't have to explain it all again.
Over on the agent desk, Sarah's case lands at the top of the queue within a second, alongside live counts and updates.
Each case shows exactly which rules set its urgency and team, point by point. Plain rules, no AI.
Customer records, the bill and meter history sit together, so agents don't have to dig through four systems.
Agents assign the case, tag what made it harder than it should be, leave a note and record how it was resolved.
Filter to chats the assistant resolved on its own. In the demo data, that's 164 conversations with no agent time needed.
Insights tracks backlog flow, expected wait time, what the backlog is made of and how long cases stay open.
Insights also flags unusual spikes in volume, tracks how often the assistant solves chats, and surfaces what customers and agents report.

Northwind is facing a growing customer support problem: more complaints are entering the system while complaints are also taking longer to resolve. This is creating a growing backlog, putting increasing pressure on support agents, and making it harder to address customer issues efficiently. Even relatively simple inquiries take weeks to resolve, further adding to the workload.

Our AI-powered customer support system helps Northwind reduce this growing backlog by resolving simple inquiries automatically and making the remaining complaints easier to manage. The chatbot can answer common questions, provide personalized bill and usage breakdowns using real-time data, and escalate more complex issues to human agents. When escalated, AI-powered triage prioritizes and routes each complaint to the appropriate system, while an agent dashboard gives staff a clear overview of their workload and the most urgent issues.

The goal is simple: resolve more complaints automatically, reduce unnecessary workload for support agents, and ensure the complaints that do require human attention are handled quickly and by the right resources.

Built With

fastapi
gemini
godaddy
langgraph
python
react
tailwind
typescript
vultr

Submitted to

Hack the Hill III
- Winner CGI Challenge — Third Place

Created by

I built the backend in Python and FastAPI, wrote the complaint-triage logic for the AI assistant in LangGraph, and handled backend deployment. I designed the agent graph so user data never entered the LLM's context window, which kept it private by design.

Oluwatoyin Faniyan
I did the data analysis for the team. Northwind's challenge rules required all diagnosis and value-case numbers to come from hand analysis, not AI, so I worked through the CSV pack in a Jupyter notebook using pandas — profiling every field, validating derived columns against their inputs, and reconciling monthly totals against the KPI file before trusting any number. That EDA is what surfaced the three findings the pitch is built on: the system-transfer penalty, the legacy-region meter problem, and the information-only complaint backlog.

Fatum Ali
Abdul Raafey
Mustafa Karimi