Custom App & AI
Software and private AI, built in your cloud and kept running.
Web apps, integrations, AI assistants and open-source alternatives to paid tools. Built by a hands-on engineering lead, deployed on your servers or cloud, and run for you if you want.
StackReact · TypeScript · Python / FastAPI · Node.js · PostgreSQL · Docker · AWS, Azure, GCP or your own servers
| Your situation | What we do | How to start |
|---|---|---|
| The business runs on spreadsheets and macros nobody can maintain | A web app with roles, approvals and an audit trail | Discovery sprint, then build |
| Staff re-key the same data into several systems | APIs and queues that keep ERP, CRM, ServiceNow or SAP data in step | Scoped build |
| People cannot find answers in policies and documents | A private AI assistant that respects permissions and shows its sources | AI readiness sprint |
| An AI pilot works in a demo but cannot go live | Answer-quality tests, guardrails, masking of personal data, monitoring | Production audit |
| The back office is buried in invoices and forms | Document extraction with checks against your records and human review | Scoped build |
| You pay for software seats you could own | An open-source alternative, compared feature by feature, self-hosted and run for you | Cost review |
| You built something but cannot run it reliably | Containers, backups, monitoring and updates on your servers or cloud | Production audit |
| An ageing app whose developers have left | Assess, add tests, then modernise in steps | Assessment |
| You need engineers inside your team | Named engineers in your repository and sprints | Month-to-month engineer |
Your repository and cloud from day one
Code, infrastructure scripts and prompts live in your accounts. Nothing is locked to us.
Tests and automated releases
Every change runs through tests and a release pipeline before it reaches users.
Grounded, cited and tested
Answers come from your sources, show where they came from, and are checked against a test set before each release.
Permissions enforced on retrieval
An assistant only sees the documents the person asking is allowed to see.
A person approves AI actions
AI drafts and suggests. Anything that writes to a system waits for approval or a fixed rule.
We run what we build
Deployment, monitoring, backups and upgrades, or a full handover with runbooks.
Someone asks
"What is our approval limit for capital spend?"
The user signs in with your company login; their groups are passed with the question.
Only allowed documents are searched
Files the person cannot open are never searched.
Keyword and vector search over indexed documents, filtered by the user's permissions.
An answer with sources
A short answer, with links to the exact pages it came from.
The model answers only from retrieved passages; personal data is masked before it is sent.
It gets better
Thumbs up or down on each answer.
Feedback feeds a test set that runs before every release; usage and cost are logged.
Cost and features compared
Seats, features you use, and what you would give up.
Line-by-line feature match against the open-source option; hosting and support cost added in.
Set up on your servers
Same login as your other tools. Your data stays with you.
Containers on your cloud or server, single sign-on, TLS, automated backups.
Data moved over
Contacts, history and settings come across before the switch.
Export, transform and import scripts, with a trial run and a check of record counts.
Kept running
Updates and fixes handled. You keep the seat savings.
Pinned versions, uptime and disk monitoring, tested upgrades, restore drills.
Evaluated across OWASP API Security Top 10, golden-dataset RAG evaluation benchmarks (faithfulness, citation accuracy, prompt injection resistance), and CI/CD test coverage.
| ID | Area | Finding | Severity | Effort | Fix by |
|---|---|---|---|---|---|
| AI-01 | Security & RBAC | Vector search missing Entra ID group security trimming | High | M | Sprint 1 |
| AI-02 | Data Privacy | Customer NRIC / phone numbers unmasked in LLM prompt logs | High | S | Sprint 1 |
| AI-03 | Reliability | ERP order sync endpoint missing idempotency key & retry queue | High | M | Sprint 2 |
| AI-04 | AI Quality | 14% hallucination rate on multi-table policy PDFs (fixed via hybrid search) | Medium | M | Sprint 2 |
| AI-05 | Testing | Missing golden-dataset regression evals in GitHub Actions CI | Medium | S | Sprint 2 |
| AI-06 | Cost & Latency | Repeated policy queries uncached; semantic cache cuts token spend 40% | Low | S | Sprint 3 |
Publishing platform with automated distribution
A content system for a research publication: editor, automated publishing and automatic posting to social channels.
Multi-agent research assistant
Several AI agents that search, read and cross-check sources, then assemble a cited research brief.
Invoice reader
Reads supplier invoices in mixed formats and extracts supplier, line items and totals into structured data.
Agent portal over a self-hosted help desk
A secure portal for support agents with passwordless company login, client allocation, team-lead views and an audit log of admin actions.
Self-hosted business tools
Help desk, ERP, team chat and an app deployment platform, set up on private servers with backups and pinned versions.
Office assistant on Telegram and WhatsApp
An AI assistant staff message like a colleague for routine admin and look-ups.
Engineering lead
Architecture, build and code review across web, data, AI and hosting.
Full-stack and AI engineers
Added for the scope, working in your repository and sprints.
Your product owner and IT
Set priorities and approve releases. You get the code, runbooks and access.
Custom App & AI FAQ
Who owns the code?
You do. Code, prompts, test sets and infrastructure scripts live in your repository and cloud from the first day.
Which AI models do you use, and where does our data go?
Enterprise endpoints such as Azure OpenAI, AWS Bedrock or Google Vertex AI in your own account, or open models on your servers. Your data is not used to train models.
How do you stop wrong answers?
Answers must come from your sources and cite them. A test set of real questions runs before every release, and anything that writes to a system needs approval.
Can you replace a paid tool with open source?
Often. We compare features line by line first, and tell you where the paid tool is still the better choice.
Who runs it after launch?
We can, month to month, covering updates, monitoring and backups. Or we hand over with runbooks.
Can it connect to our ServiceNow or SAP?
Yes. Our ServiceNow and SAP leads work with us on those integrations.
Talk to our engineering lead
An app, an integration, a private AI assistant, or a cheaper open-source setup.