Updated
DealWizard.ai: a GenAI product from concept to production
Summary
A GenAI platform that turns a property listing into a structured investment analysis: layout checks, area comparables, a refurbishment plan, risks and returns for the chosen strategy.
Key figures
- Estimated 6 to 8 hours to about 3 minutesfor the analysis
- 99.97%production uptime
- 3strategies (Buy-to-Let, Flip, HMO)
Context
Property investors research every deal by hand: listing details, local comparables, room sizes and layout, refurbishment costs and financial projections. It is slow and inconsistent. The idea was to bring that whole process into one guided flow, with a report an investor can check.
Challenge
Build an AI system whose analysis people can verify. That meant three things at once: outputs grounded in real data instead of model memory, a cost per analysis that works for a small product, and reliability good enough to use every day.
Starting point
A manual analysis that took an experienced person an estimated six to eight hours per deal.
My role
I co-founded the venture with Maciej Brzek and shared product leadership with him: the idea, the roadmap and the priorities were ours. As system architect and AI technology lead I designed the AI architecture, built the application and its automations hands-on in low-code tools, and owned delivery, quality and operating cost.
What I did
Start from the process and the outcome, not the model.
I defined the workflow first (listing, strategy, structured report) and the outcome an investor needs, then chose where AI helps and where deterministic logic is better.
Ground the model in real data, and buy what is not worth building.
Model output is grounded in multiple third-party data sources through external APIs, specifically to reduce hallucination risk. For data collection we used a scraping provider rather than building our own: quick and cheap.
Design for consistency and cost.
RAG and CAG architecture, model selection and routing (simple tasks go to cheaper, faster models), caching of stable content, and control of prompt and context length.
Build the safety net.
Evaluation and validation of answers, production observability, feedback loops and an uptime target. The platform ran at 99.97% uptime in production.
Use low-code where speed matters.
The first version was built in Bubble.io, with n8n for workflows and Firebase for data. It let a small team ship and iterate quickly.
What exists today
- A live public product at dealwizard.ai.
- A guided flow from listing to report for Buy-to-Let, Flip and HMO strategies.
- Layout and room checks, area comparables, a refurbishment plan, risk notes and return metrics.
- A published example analysis.
- Production monitoring, evaluation and feedback loops.
Outcome and evidence
| Measure | Value | Basis | Scope note |
|---|---|---|---|
| Analysis time | Estimated 6 to 8 hours to about 3 minutes | My estimate of a typical manual analysis compared with a typical automated report | Estimate, not a measurement |
| Uptime | 99.97% | Production platform | Over its live period |
| Strategies covered | Buy-to-Let, Flip, HMO | Product scope | Public on dealwizard.ai |
What the figures do not show
Both times are estimates, not measurements: six to eight hours is my estimate of a typical manual analysis, and about three minutes is a typical automated run. It is not a controlled study, and speed says nothing about quality on its own. Uptime is for the production platform over its live period. Commercial performance is not claimed: the product is a live public beta with early users. Layout findings and budgets in the report are estimates, and the report shows what they rest on.
Learnings
Taken from points I have shared publicly on LinkedIn.
- Do not pass files to an AI agent as base64 inside the prompt. It inflates the payload, fills the context window and is billed as input tokens on every call. Upload the file once through a files API and pass a reference.
- Optimize prompts for caching. Stable content such as system prompts and reference data should not be re-billed as fresh tokens on every call.
- Model selection is a decision, not a default. Benchmark cheaper models on cost and quality for each task.
- Validation does not always need a human. A second agent can cross-check the first one's output for consistency at a fraction of the cost of human review on every action that is not irreversible.
- Track token usage and spend on a dashboard, or overruns show up on the invoice.
Want the next case study when it is out?
An occasional email about my work, such as new case studies and articles.
Check your inbox
I sent you a link to confirm your address. Nothing is added until you confirm. The link works for 7 days.
Building or scaling an AI product?
Tell me what it should achieve and how its quality is judged today.