The six phases of a typical project
Discovery call
We start with a 45-minute video call (or in-person meeting if you prefer). The goal is simple: understand what problem you want to solve and whether AI is the right tool for it. We ask about your current workflow, the data you have available, the systems you already use and the outcome you want to measure.
About one in three discovery calls ends with us recommending a non-AI solution. A well-designed SQL query or a rule-based script can sometimes do the job faster and cheaper. We would rather tell you that upfront than sell you a model you do not need. If AI does make sense, we move to phase two and schedule a data review within the next five business days.
Data audit and feasibility assessment
Before we commit to a timeline or price, we need to see your data. You share a representative sample (we sign an NDA if required), and our engineers spend three to five days evaluating it. We look at volume, completeness, label quality and whether the signal we need is actually present.
At the end of this phase you receive a written feasibility report. It includes a candid assessment of data readiness, a recommended approach (which model family, which infrastructure), estimated accuracy ranges and a list of risks. If the data is not ready, we outline exactly what needs to change before a model can be trained. This report is free for projects above HK$120,000; for smaller projects we charge a flat HK$8,000 audit fee that is credited toward the project if you proceed.
Proposal and agreement
Based on the feasibility report, we write a fixed-scope proposal. It lists every milestone, its deliverables, its acceptance criteria and its price. You know the total cost before we write a single line of production code.
Milestones are typically two to three weeks long. Each one produces something tangible: a trained model, an API endpoint, a dashboard, a deployment script. Payment is due when you sign off on each milestone, not before. If we miss a deadline by more than five business days (and the delay is on our side), we apply a 5% discount to that milestone automatically.
Development and iteration
This is where the actual building happens. Our engineers prepare the data pipeline, train the model, write the integration code and set up the infrastructure. You get a weekly progress update every Friday afternoon, delivered as a short written summary with screenshots or demo links.
We run internal code reviews on every pull request and automated tests on every commit. For ML models specifically, we track metrics (precision, recall, F1, RMSE) on a validation set throughout training and share the results with you so you can see how accuracy evolves week by week. If a model is not converging as expected, we flag it early and discuss whether to adjust the approach or the scope.
Testing and validation
Before deployment, we run the system through a structured validation period. For predictive models, this means comparing predictions against real outcomes over a two-week window. For automation tools, we run them in shadow mode alongside the existing process and compare outputs.
Your team participates in user acceptance testing during this phase. We provide a checklist of scenarios to test and a simple form for reporting issues. Bugs found during validation are fixed at no extra cost. If accuracy falls below the threshold we agreed in the proposal, we retrain or adjust the model until it meets the target or we refund that milestone.
Deployment and support
Once validation passes, we deploy to your production environment. We handle server provisioning, CI/CD pipeline setup, monitoring dashboards and alerting rules. Every deployment includes a rollback plan in case something goes wrong in the first 48 hours.
Post-launch support is included for 30 days at no additional charge. During this period we fix any bugs, answer questions from your team and monitor system performance. After the 30-day window, you can either manage the system yourself (we hand over full documentation and access) or move to a retained partnership for ongoing maintenance and model retraining.
Common questions about our process
These are the things clients ask most often during or before the discovery call.
Most projects run between six and fourteen weeks from signed proposal to production deployment. The biggest variable is data readiness. If your data is clean and well-structured, we can move fast. If it needs significant cleaning or labelling, add two to four weeks. We always give you a specific timeline in the proposal, not a range.
That is normal. Almost every dataset we receive has gaps, duplicates or inconsistencies. Our data audit phase identifies these issues and we include data cleaning as an explicit milestone in the proposal. For labelling tasks (like annotating images for a computer vision model), we can either train your team to do it or handle it ourselves using our annotation partners.
You do. All source code, trained model weights, configuration files and documentation are transferred to you upon project completion. We retain no proprietary rights. The only exception is if we use an open-source library with its own license (like PyTorch or scikit-learn), in which case that library's license applies to that component.
Yes, within reason. Small adjustments (changing a field name, tweaking a threshold) are absorbed into the existing milestone. Larger changes (adding a new data source, building an extra model) require a change order with its own timeline and price. We draft change orders within two business days so you can decide quickly.
Python is our primary language for ML work (PyTorch, scikit-learn, Hugging Face Transformers). For backend services we use Python (FastAPI) or Go, depending on performance requirements. Infrastructure is typically on Google Cloud Platform or AWS, managed with Terraform. We adapt to your existing stack when possible rather than forcing a migration.
Why we structure projects this way
The six-phase process exists because AI projects fail in predictable ways. The most common failure is building a model before understanding the data. The second most common is deploying a model without a proper validation period. Our structure forces us to address both risks before they become expensive problems.
We also learned early on that weekly communication matters more than daily standups. Clients are busy. A concise Friday update with metrics, screenshots and a clear list of next steps keeps everyone aligned without eating into your calendar. If something urgent comes up mid-week, we call or message immediately rather than waiting for the scheduled update.
The milestone-based payment model protects both sides. You never pay for work that has not been delivered and accepted. We never start a phase without knowing the previous one is approved. It keeps the relationship honest and the project moving forward.
Ready to start with a discovery call?
The first conversation is free and takes about 45 minutes. We will tell you honestly whether AI fits your situation and what the next steps would look like.
Book a discovery call