The problem
Customer relationships rarely fail for lack of effort. They fail because context is scattered. Leads sit in a spreadsheet, the last conversation is buried in someone's inbox, quotes are built in a separate tool and invoices live with finance. When a salesperson moves on, their pipeline goes with them.
Facilis CRM is a product of Fensso Tech LLP, TechOvelpo's parent company, built by TechOvelpo engineers. It holds the whole commercial cycle in one place: capturing a lead, opening a deal, quoting, closing, invoicing and post-sale support. Contacts and companies carry their full interaction history, and calls, emails and meetings are logged against the right record.
The engineering question was how to make a wide product coherent. A CRM touches sales, marketing, finance and support, and each group wants something different from the same record. The risk is a pile of modules that do not agree with each other.
Constraints
- One record, many users. A contact has to make sense to a sales rep, a marketer and a support agent at once.
- Permissions that match real organisations. Reps should see their own records, managers their team's, and sensitive actions should leave a trail.
- Workflow across stages. A won deal should produce an invoice without anyone re-keying data, and inventory should reflect what was sold.
- Data in and out. Teams arrive with existing customer data and want to be able to leave with it, so imports and exports are requirements, not extras.
- AI as an assistant. The system recommends follow-ups, segments customers and forecasts, but the team decides what to do.
Options considered
| Decision | Options | What we chose and why |
|---|---|---|
| Database | Document store, or relational | MySQL through Sequelize. A CRM is relational by nature: contacts belong to companies, deals to contacts, quotes to deals, invoices to quotes. Foreign keys and joins express that directly. |
| Backend structure | Fat route handlers, or layered controllers and services | Thin controllers that delegate to a services layer, with Express.js, so business rules live in one place and can be reused. |
| Front-end state | One global store, or small stores per feature | Zustand with a lightweight store per feature. It keeps modules independent while the product grows. |
| Authentication | Server sessions, or tokens | JWT with refresh tokens, and bcrypt for password hashing. |
| Validation | Ad hoc checks, or a schema library | Joi schemas on requests, so bad input is rejected at the edge. |
We picked the relational option over the one we used for Krayuv deliberately. Krayuv's centre of gravity is live bidding; Facilis's is connected records and reporting. Choosing the same database for both would have been convenient and wrong.
Architecture
React front end (feature stores, protected routes, charts)
| HTTP, Axios with auth interceptors
v
Express.js API
middleware: authentication, role-based access, audit logging,
request validation, error handling, rate limiting
controllers -> services (business logic) -> Sequelize models
|
v
MySQL
| Layer | Responsibility |
|---|---|
| Front end | Pipeline board, customer records, quote and invoice screens, dashboards built with Recharts |
| Middleware | Authenticate, authorise by role, record audits, validate, handle errors |
| Services | Business rules for deals, quotes, invoices, automation and forecasting |
| Models | Relational schema with associations between customers, deals, quotes and invoices |
The module list gives a sense of scope: leads, contacts and companies, deals and pipeline, quotes, invoices, inventory and stores, tasks and calendar, documents, email campaigns and templates, workflow automation, forecasting, reports and dashboards, a knowledge base, and imports and data export.
Notable engineering decisions
Thin controllers, real services
Controllers handle the HTTP shape and nothing else. Business rules, such as what happens when a deal is marked won, sit in services. When the same rule is needed from a workflow automation as from a button click, it is called, not copied.
Authorisation and auditing as middleware
Role-based access and audit logging are applied in the request pipeline, not sprinkled through handlers. The advantage is that a new endpoint inherits the protection by default, and a missing check is easier to spot in review.
Quote to cash without re-keying
Quotes are generated from the deal, invoices from the won deal, and inventory stays in step with what has been sold. We treated these as one flow with shared data instead of three modules linked by exports. The benefit shows up as fewer reconciliation errors between sales and finance.
Automation between stages
Workflow automation handles routine steps, such as assigning records, creating tasks and sending templated emails. It is built on the same services the interface uses, so automated and manual actions follow the same rules.
AI that recommends and does not decide
Predictive analytics across the pipeline, follow-up recommendations and customer segmentation from interaction history all feed suggestions. The person still chooses the action. That boundary is a product decision, and it keeps accountability clear.
What was hard, and what we would do differently
Breadth is the real cost. Each CRM module is simple alone. The difficulty is consistency among fourteen of them: shared concepts of owner, status and history. We would define those shared concepts first and build modules against them, instead of letting each module invent its own.
Permissions are a data problem. Role-based rules that read cleanly on paper get complicated when records have owners, teams and shared views. Our lesson is to settle the ownership model early, because adding ownership to an existing schema means backfilling data.
Imports are where reality arrives. Customer data comes in with duplicates, missing fields and inconsistent formats. Import tooling deserves the same design effort as the screens, including previews and clear error reporting.
Dashboards depend on definitions. Pipeline value, conversion rate, sales cycle length and retention are only trustworthy if the whole team agrees on what they mean. We would write down each definition alongside the code that computes it.
Be cautious with AI claims. Forecasts are only as good as the data behind them. A recommendation feature should be introduced as an aid and measured against how teams actually use it.
What this means if you are building something similar
If you are considering a custom CRM, the first question is whether you need one. Off-the-shelf tools cover many teams well, and building makes sense when your sales process, data model or integrations are unusual enough that configuration cannot reach them. If that describes you, the decisions above are a reasonable starting point: a relational model, layered services, authorisation and auditing in middleware, and a clear line between automation and human judgement.
Our full-stack developers work across React and Node.js, which is the combination used here, and staff augmentation is a sensible way to add that capacity to a team that already has a product owner. For a product built from scratch, see how a dedicated development team works. To discuss your own CRM or internal platform, get in touch. The sister platform for procurement is covered in the Krayuv case study.
