AI Integration
AI that does the work, not the talking
We build AI into businesses that already run — storefronts, operations, support desks. Scoped against a real workflow, measured before it ships, and kept working after it does.
What we build
Four things we ship — scoped, priced and measurable
AI operations agent
An agent that runs inside the tools your team already uses — WhatsApp, Slack or Teams — and handles the daily operational load instead of adding another dashboard to check.
- Morning briefings on revenue, orders and blockers
- Order and delivery tracking with follow-ups
- Stock alerts and drafted reorders
- Team check-ins and status roll-ups
Support & sales agent
A customer-facing agent grounded in your own catalogue, policies and documentation — so it answers from your business, not from the open internet.
- Retrieval over your docs, catalogue and past tickets
- Clean handoff to a human, with full context
- Lead qualification and booking
- Deployed to web, WhatsApp or email
Commerce AI
The AI layer for a storefront you already run. We build on Shopify and WooCommerce, so this sits on top of a stack we ship in production.
- Semantic product search and discovery
- Size, fit and product guidance
- Product copy and multi-language localisation
- Review summarisation and returns deflection
Document & workflow automation
The unglamorous work that quietly costs the most: getting information out of documents and into the systems that need it.
- Invoice, contract and form extraction
- CRM enrichment and email triage
- Internal knowledge assistant over Notion, Drive or Confluence
- Scheduled reporting and monitoring
Our engine
We did not wrap someone else's API — we built the engine
dekhval
An AI operations manager for small businesses, built and operated by us. It works inside WhatsApp, Slack and Teams — sending morning briefings, tracking orders, raising stock alerts, running team check-ins and answering customer queries proactively, rather than waiting to be asked. Bilingual in English and Bangla, live across e-commerce, SaaS, professional services, healthcare and RMG.
- We run it in production, so we know its failure modes — not just its demo
- Client integrations build on the same engine, which is why scoping is fast
- Deployed into your accounts and your infrastructure, with documentation
- No vendor lock-in: you get the code, the config and an exit path
Client work
The commerce we build on
How we work
Audit first — because most AI projects fail at scoping, not at the model
01
Audit
We map how your business actually runs, then pick the two or three automations with the clearest payback. You get a written scope, a cost and defined success criteria — before anyone writes code. Fixed price, two weeks.
02
Prototype
We build against your real data and measure it against the criteria we agreed. If a prototype cannot clear the bar, we tell you and we stop. That is cheaper for both of us than shipping something nobody trusts.
03
Integrate
It goes into the tools your team already opens — no new app, no training programme. Every automation ships with guardrails, a human handoff path and an audit trail of what it did and why.
04
Operate
Models drift, catalogues change and edge cases surface in month three. We monitor quality, run evaluations against real traffic and keep tuning. AI is not a launch, it is an operating system.
Stack
What we build on — and why it is not a secret
We pick per problem rather than per vendor, and everything runs in accounts you own. If a self-hosted open model is the right answer for your data, that is what we will recommend.
Models
- Claude
- GPT
- Gemini
- Llama & Mistral (self-hosted)
Agents & orchestration
- Claude Agent SDK
- LangGraph
- Vercel AI SDK
- MCP
Retrieval
- Qdrant
- Weaviate
- Postgres pgvector
Automation
- n8n
- Temporal
- Make
Channels
- WhatsApp Business
- Slack
- Microsoft Teams
- Email & web
Commerce & CRM
- Shopify
- WooCommerce
- HubSpot
- Notion & Airtable
Evaluation & monitoring
- Langfuse
- LangSmith
- Custom eval suites
Infrastructure
- Vercel
- Cloudflare
- AWS Bedrock
- Azure OpenAI (EU regions)
Data & compliance
Your data does not become our training set
For European and Australian clients this is usually the first question, and it should be. Here is our position before you have to ask for it.
- No training on client data — contractually, not as a preference
- A signed DPA before integration work begins
- EU data residency available, or fully self-hosted open models
- Human handoff and an audit trail on every automated action
- AI disclosure built in, per EU AI Act transparency obligations
- Risk classification assessed during the audit, not after launch
Full detail on how we handle personal data is in our GDPR compliance statement.
FAQs
The questions worth asking first
Do we have to replace the tools we already use?
No, and we would push back if you asked us to. Every integration we build goes into your existing stack — your storefront, your CRM, your Slack or WhatsApp. A system that requires your team to adopt a new dashboard is a system your team will quietly abandon by month two.
What happens to our data?
It stays yours. We do not train models on your business data, we sign a DPA before any integration work begins, and where the data is sensitive we deploy to EU regions or self-host open models entirely on infrastructure you control. We will tell you exactly which third parties touch which data before you commit.
You have your own product. Does that lock us in?
No. Building on our own engine is why we can scope and ship quickly — the hard operational parts already exist and are running in production. But your integration is deployed into your accounts, on your infrastructure, and you get the code, the configuration and documentation. If you want to take it in-house or move on, there is a path out. An agency that can only keep you by trapping you is not one worth hiring.
How is this compliant with the EU AI Act?
It depends on what the system does, and that assessment is part of the audit. Most operational and support automation is limited-risk, which means transparency obligations — people must know they are talking to an AI. Some use cases, employment screening in particular, are classified high-risk and carry real documentation and oversight requirements. We will tell you which category you are in rather than leaving you to find out later.
What if the AI gets something wrong?
It will, occasionally — anyone who tells you otherwise is selling something. That is why we scope guardrails first: confidence thresholds, a defined human handoff, and an audit trail of every action taken. We design for the failure case before we design for the demo.
How quickly can something be live?
The audit takes two weeks. A first working integration typically follows within two to four weeks after that, depending on how clean the data and the surrounding systems are. We would rather give you one automation that genuinely works than six that need constant supervision.
Can you work alongside our existing developers?
Yes. We frequently build the AI layer while an in-house team owns the core product. You get documented code, the infrastructure in your own accounts, and a handover — not a dependency on us.
Start with the audit , not with the build
Two weeks, fixed price. You walk away with a written scope, a cost and a clear answer on whether AI is worth it for your business — and that answer is genuinely sometimes no. If you go ahead, the audit fee comes off the build.