PipelineDealUnderwritingMarketAI analysisIC memo
CREUW AI
A commercial real estate platform that follows one deal from first look to investment committee.
AI & Machine Learning: Intelligence that does real work.
We put AI where it earns its place, inside workflows, decisions and product features, and engineer it to be measurable, affordable to run and safe to depend on.
Every useful AI system is one sentence long.
Data becomes understanding, understanding becomes a decision, a decision becomes an action. Pick a use case and read it across.
Data
A customer email
Hi, our order arrived damaged. Can we get a replacement shipped this week?
Understanding
- Intent
- replacement
- Issue
- damaged item
- Urgency
- this week
- Order
- found in history
Decision
Route to Returns
Confidence 0.94
Above 0.85 it acts. Below, a person decides.
Action
A ticket is opened and a replacement drafted. An agent approves it with one click.
Data
A scanned invoice
INVOICE 2291 — 14 line items — Net 30 — Total 18,420.00
Understanding
- Supplier
- matched to record
- Total
- 18,420.00
- Terms
- 30 days
- Lines
- 14, all read
Decision
Matches the purchase order
Confidence 0.97
Any mismatch goes to accounts, with the difference highlighted.
Action
Entered in the ledger and queued for payment. Nobody typed anything.
Data
A shopper's history
Viewed trail shoes and a rain shell. Bought running socks. It is October.
Understanding
- Interest
- trail running
- Owns
- socks, not shoes
- Likely next
- head torch, gloves
- Price band
- mid
Decision
Show the head torch first
Relevance 0.81
Never suggest what they own. Never guess below 0.6.
Action
Three suggestions on the product page, each with the reason it is there.
Data
Three years of orders
Weekly sales by region, with holidays and promotions marked.
Understanding
- Trend
- rising steadily
- Season
- peaks late November
- Promotions
- lift, then a dip
- Noise
- two outlier weeks
Decision
Order 1,240 units
Likely range 1,090 – 1,410
Always shown with its range, so a planner can overrule it.
Action
A purchase suggestion reaches the planner on Monday morning, ready to adjust.
Data
A question, typed plainly
what's our policy on returning opened items
Understanding
- Meaning
- returns, opened goods
- Sources
- 3 policy documents
- Access
- checked for this user
- Best match
- Returns policy, §4
Decision
Answer from §4
Grounded in a source
No source, no answer. It says when it does not know.
Action
A direct answer, with the paragraph it came from one click away.
Data
A request, in a sentence
Move Thursday's site visits to Friday and let the clients know.
Understanding
- Task
- reschedule
- Affected
- 4 visits, 4 clients
- Conflict
- 1 on Friday
- Needs
- calendar, messaging
Decision
Propose a plan
Waiting for approval
It may draft anything. It sends nothing without a yes.
Action
The calendar is updated and four messages go out, after one confirmation.
Should this be AI at all?
It usually earns its place when
- There is too much of something for people to read, sort or check
- The same judgement is made hundreds of times a day
- A good-enough answer now beats a perfect one next week
- You have examples of what right looks like
It usually does not when
- A rule, a query or a better form would do the job
- A single mistake is unacceptable and nobody can review
- There is no data yet, and no way to collect it
- The goal is to have AI, not to change an outcome
We ask this first, on every project. Sometimes the honest answer is no, and we will say so.
Models are sometimes wrong. Design for it.
The useful question is never whether a model makes mistakes. It is what happens to them. Move the bar and watch.
- handled automatically
- 9
- sent to a person
- 7
- mistake slips through
- 1
Lower the bar and more is automated, including more of the mistakes. Raise it and people review more. Where it sits depends on what a mistake costs you, so we build it adjustable and measured.
What ai & machine learning means here
AI and machine learning development at SOLOGEN means building systems that read, classify, predict, recommend and act inside a product: language-model integrations and copilots, document intelligence, intelligent search, automation, and models trained on your own data.
Who usually needs it
- A team drowning in documents, messages or tickets
- A product whose users now expect it to be smart
- An operation with years of data nobody has put to work
- A leader who suspects AI could help, and wants an honest answer
What it turns into.
The same sentence, in seven shapes. Most products need one or two of them.
- AI copilots
- Assistants inside your product that answer from your own data and can take actions, with limits you set.
- Intelligent workflows
- Multi-step processes where software reads, decides and routes work, with people in the loop where it matters.
- Document intelligence
- Invoices, contracts, forms and reports read automatically and turned into structured records.
- Recommendation systems
- Suggestions that reflect what each person is actually looking for.
- Classification and prediction
- Models trained on your data to sort, score, forecast and flag.
- Intelligent search
- Search that understands meaning, not just matching words, across your content and records.
- AI integrations
- Language and vision models connected to the product and systems you already have.
Engineered like software, because it is.
Evaluated.
Each feature has a test set built from real examples and a target agreed in advance. Quality is a number you can watch, before and after release.
Observable.
What the system was asked, what it answered and what it cost are logged, so a strange result can be traced to its cause.
Affordable.
Running cost is estimated in planning and designed down: smaller models where they are enough, caching, batching.
Replaceable.
Models improve and prices move. A clean boundary around the model means it can be swapped without rebuilding the product.
Models, data and what serves them.
Backend
The logic, APIs and integrations behind the interface.
The backend carries your business rules. A well-structured one makes new features cheap and keeps the product stable when usage climbs.
- Node.js
- Fast, event-driven services and APIs that share a language with the frontend.
- NestJS
- Structured, testable Node.js architecture for larger systems and teams.
- .NET
- Enterprise-grade services, especially where they meet existing Microsoft systems.
AI / ML
The parts of a product that read, see, predict and decide.
AI choices affect accuracy, running cost and how much you depend on a single provider. We choose for the task and keep the option to change.
- Python
- The working language of machine learning, data pipelines and model serving.
- AI APIs
- Language and vision models integrated into products and workflows.
- Machine learning models
- Models trained on your own data for prediction, ranking and classification.
- Computer vision
- Detection, inspection and document understanding from images and video.
Databases
Where your product's information lives and stays correct.
Data outlives code. The right model keeps information consistent, queries fast and future reporting possible.
- PostgreSQL
- A dependable relational default for structured data and complex queries.
- MySQL
- Proven relational storage, often the right fit alongside existing systems.
- MongoDB
- Flexible document storage for data whose shape changes often.
Cloud & Infrastructure
Where the product runs, scales and recovers.
Infrastructure decides reliability and running cost. We fit the platform to your scale, your team and any systems you already use.
- AWS
- Broad, mature infrastructure for products that need room to scale.
- Azure
- A natural fit for organisations already built around Microsoft.
- GCP
- Strong data and machine-learning services alongside general infrastructure.
- Firebase
- Managed backend services that get mobile and real-time products live quickly.
- Supabase
- A managed PostgreSQL backend with authentication and storage built in.
Before you ask.
We are not sure AI is right for our problem. Can you help us decide?
Yes, and sometimes the answer is no. We look at the task, the data you have and what a mistake would cost. If a rule, a query or a better form would do the job, we will say so.
Can you add AI to a product we already have?
Yes. Integrating AI into an existing product is a core part of this service: connecting models to its data and its permissions, with our web and mobile engineers working alongside the AI team.
How do you know an AI feature is good enough to release?
Every feature gets an evaluation set built from real examples and a target agreed with the people who will rely on it. Quality is measured before release and tracked afterwards, so it is something you can see.
What does it cost to run?
It varies with volume and the models used, so we estimate running cost during planning and design around it: smaller models where they are enough, caching, and batching. You see the figure before you scale.
What happens to our data?
That is decided with you at the start: what leaves your systems, which providers are acceptable, what is retained and what is not. The architecture follows those constraints.
How do we get started?
Describe the workflow or feature you have in mind in the project brief, or ask to talk to an AI specialist if you would rather test the idea first.
