01
Reprompt AI
Converts images into structured prompts for generative AI workflows.
Image input → Structured prompt
Package: org.reprompt.app
View Google Play listing ↗Selected work / Applied AI
Turning business problems
into working AI systems.
I design and build AI assistants, business automation and decision-support tools, connecting the technology to the people who use it.
Since 2021Software & AI project experience
50+Independent AI apps & automation projects
Business → deliveryDiscovery, design, build and rollout
Enterprise AIWorkflow automationHands-on engineeringStakeholder delivery
Selected projects / explore the workChoose a project to explore
How I deliverBusiness first. Engineering throughout.
01 UnderstandUsers, problems, constraints
02 DesignData, workflow, controls
03 BuildApplication and integrations
04 ValidateReview answers and exceptions
05 Roll outFeedback and iteration
Integration pattern / Model Context ProtocolConnect agents to business systems,
Illustrative demo · Fictional dataConnect agents to business systems,
with clear limits.
MCP is an open standard that lets an AI agent discover and use tools in other systems through one consistent interface. I use it to plug agents into CRM, email, calendars, documents and data, and I decide exactly what each one is allowed to do.
Agent“Prepare a follow-up for Acme Trading”
MCP clientDiscovers tools · sends calls
Tool calls
Run the task to see each MCP call.
Turn off the CRM server, then run the task again to see the agent stop instead of guessing.
How I design an MCP integrationLeast access by default
01List the jobs
What should the agent actually do? Start from the workflow, not the tools.
02Pick servers + tools
Expose only the tools those jobs need.
03Scope the access
Read, draft or write. Least access by default.
04Gate the writes
Anything that changes data or contacts a customer waits for approval.
05Log and review
Every tool call is visible, so the team can audit and improve it.
When to use what
Direct API integration
One app, one fixed workflow, and the steps never change. Simple and fast to build.
MCP server
Several agents or apps need the same tools, or the agent must choose which tool to use. Build once, reuse everywhere, with one place to control access.
MCP servers I use in my own AI workflows
Google DriveFind and read files
Microsoft 365Mail, calendar and documents
FirecrawlWeb research and scraping
Chrome browserBrowse and act on web pages
Desktop CommanderLocal files and terminal
01 / HR Policy AssistantA clear answer, with a source.
Agent spec ↓Illustrative interactive demoA clear answer, with a source.
A human when needed.
Employees ask in everyday language. The assistant finds relevant policy passages before answering, and escalates when the documents do not support an answer.
Employee conversation
Evidence & routing
Choose a scenario, then run the walkthrough.
Answers tied to policy evidenceUnclear cases go to HRHuman review before policy decisions
Agent spec / how I designed itWorkflow · guardrails · success metrics
Employee questionEveryday language
Policy searchApproved documents only
Confidence checkIs there a source?GUARDRAIL
Answer + sourceWhen the policy supports it
HR teamGets unclear casesHUMAN
Goal
Give employees a clear policy answer with its source, and route anything unclear to HR.
Tools & access
Gemini · Vertex AI embeddings · Firestore vector search over approved policy documents.
Guardrails
- Answers only from approved policy text
- No source found, no answer: it routes to HR
- Policy decisions stay with HR
Success metrics
- Questions answered with a cited source
- Escalation rate to HR
- Routine questions HR no longer handles by hand
Rollout
Start with one policy area (leave), review answers with HR, then add more policies.
MMD & AIMX / Conversational sales systemsFrom an inventory question
Agent spec ↓Recreated interfaces · Sample dataFrom an inventory question
to a qualified sales conversation.
WhatsApp and voice agents answer using approved business and inventory data, capture buyer requirements and organise the conversation in a mini CRM so the sales team can continue with context.
Business problem
Customers ask about availability, products and buying options across channels. Sales needs the enquiry, qualification details and conversation history in one place.
What I built
For MMD and AIMX, built inventory-connected conversational systems, WhatsApp lead capture, voice sales assistants and a protected mini CRM with transcripts and customer details.
Business workflow
Agents answer and qualify enquiries, handle initial sales conversations and support the next step. Salespeople review the history and follow up through WhatsApp or email.
01 / ChannelsWhatsApp or phone enquiry
02 / GroundingApproved inventory & information
03 / QualificationNeeds, budget, timeline & contact
04 / HandoverMini CRM → Sales follow-up
Inventory-grounded availabilityWhatsApp + voice enquiriesStructured qualificationTranscripts & customer historySales follow-up with context
Delivered components: Next.js · Gemini · WhatsApp Cloud API · ElevenLabs Conversational AI · Firebase Auth · Firestore · Notifications · Protected lead dashboard. The demo recreates the workflow; it does not access real stock, place calls or message customers.
Agent spec / how I designed itWorkflow · guardrails · success metrics
WhatsApp / callCustomer enquiry
Sales agentAnswers + qualifies
Approved inventoryStock and product data
Mini CRMLead + full transcript
SalespersonFollows up with contextHUMAN
Goal
Answer stock questions on WhatsApp and phone, qualify the buyer and hand the lead to sales with full context.
Tools & access
WhatsApp Cloud API · ElevenLabs voice · Gemini · approved inventory data · Firebase / Firestore mini CRM.
Guardrails
- Availability comes only from approved inventory data
- Final allocation and price are confirmed by sales
- The agent supports the sale; it does not close contracts or take payments
Success metrics
- First response time to an enquiry
- Enquiries qualified (needs, budget, timeline)
- Leads handed over with a full transcript
- Time sales spends re-asking questions
Rollout
WhatsApp first, then voice calls, then CRM follow-up tools for the sales team.
03 / Market Intelligence & ForecastingExplore the scenario.
Agent spec ↓Illustrative interactive demoExplore the scenario.
Understand the trade-off.
The calculation engine produces the numbers. AI helps explain what those numbers mean, so a manager can compare options and decide what to investigate.
Illustrative monthly unit forecast
100 unitsScenario monthly demand
1.50 monthsStock cover at 150 units
── Baseline┄ Scenario
Management brief preview
Adjust demand, then compare. Results are illustrative, not measured business outcomes.
Calculation produces the numbersAI explains the evidenceManagement owns the decision
Demo rule: demand = 100 × (1 + adjustment). Stock cover = 150 ÷ demand. Briefs are prewritten templates; no live AI is called.
Agent spec / how I designed itWorkflow · guardrails · success metrics
Business data+ external indicators
Calculation engineProduces the numbers
ScenarioChange one assumption
AI briefExplains, never calculates
ManagerOwns the decisionHUMAN
Goal
Help managers compare scenarios and understand the trade-off in plain language.
Tools & access
Python · FastAPI · PostgreSQL · Redis · Next.js · OpenAI and Gemini for the written brief.
Guardrails
- The calculation engine produces every number
- The language model only explains them
- Managers own the decision
Success metrics
- Time to prepare a scenario brief
- Forecast error against actuals
- Scenarios reviewed in management meetings
Rollout
One product line in the monthly review first, then wider planning.
AI Agent Workspace / Tool-connected applicationsOne request. The right tools.
Agent spec ↓Recreated agent workflowOne request. The right tools.
A clear path to action.
An agent workspace that keeps task conversations separate and connects them to business tools. The user describes the job; the agent checks information, prepares actions and asks for approval where needed.
Business need
People switch between conversations, records and business apps to complete one task. They need an assistant that can work with those systems.
My contribution
Built agent applications with separate conversations and connected tools. This mockup demonstrates that application pattern using fictional customer records.
Delivery approach
Understand the user’s workflow, choose the required tools, define permissions and validate the action path with the business team.
Customer follow-up
Switch chats to see separate task context. Tool selections are independent for each conversation.
A request becomes a visible sequence of information checks and proposed actions.
Separate conversation contextSelectable tool accessApproval before external actionVisible task activity
Agent spec / how I designed itWorkflow · guardrails · success metrics
User request“Follow up with Acme”
AgentOwn task context
ToolsCRM · Docs · Email · Calendar
Approval gateNothing sent yetGUARDRAIL
User approvesThen the action runsHUMAN
Goal
Turn one request into a checked, ready-to-approve action, without switching between apps.
Tools & access
CRM records (read) · company documents (read) · email (draft only) · calendar (propose a slot). Access is chosen per conversation.
Guardrails
- Nothing leaves the system without approval
- If a tool is switched off, it says so instead of guessing
- Each conversation keeps its own context
Success metrics
- Time from request to approved action
- Drafts approved without edits
- Tasks stopped for missing access
Rollout
One team, one task type (customer follow-up). Add new task types after reviewing the first ones.
Business applications / Two delivered platformsConnect the sales workflow,
Agent spec ↓Synthetic operational recordsConnect the sales workflow,
from vehicle stock to quotation.
These are the operational systems behind useful automation: structured records, role access, repeatable workflows and a clear interface for the business team.
Agent spec / how I designed itWorkflow · guardrails · success metrics
Vehicle stockOne record per unit
Pipeline stagesPayments + status
Role accessBy division and roleGUARDRAIL
QuotationConsistent template
ReportsOne operating view
Goal
Put vehicle stock, pipeline and quotations in one system the sales team trusts.
Tools & access
Next.js · TypeScript · Firebase Auth + Firestore (CarFlow) · React + Supabase (quotation portal).
Guardrails
- Users see only their division and role
- Quotations follow one approved template
- Stock comes from one source of truth
Success metrics
- Time to produce a quotation
- Records kept in one place instead of spreadsheets
- Time to answer “where is this vehicle?”
Rollout
One division first, then roles and reports for the wider team.
HR productivity / Two delivered applicationsUseful first drafts.
Agent spec ↓Scripted sample outputsUseful first drafts.
Human review built into the flow.
AI helps HR teams turn inputs into structured learning content and clearer objectives. The output stays reviewable by the people responsible for it.
Agent spec / how I designed itWorkflow · guardrails · success metrics
HR inputTopic, role or goals
DraftCourse or objectives
Clarity checkScored against a rubricGUARDRAIL
HR reviewEdits and approvesHUMAN
ExportPDF / XLSX
Goal
Turn HR input into a structured first draft that HR can edit and approve.
Tools & access
React · Vite · Gemini API · PDF and XLSX export.
Guardrails
- Outputs are drafts, never final
- HR owns the final content
- A clarity score is not a performance rating
Success metrics
- Time to a usable first draft
- Edits needed before approval
- Drafts approved per review cycle
Rollout
One programme or review cycle first, then other departments.
Product thinking,
from idea to public release.
Alongside enterprise work, I have built 50+ independent AI applications and automation projects. These three Google Play applications are examples from the portfolio.
02
SnipAI
AI-powered screenshot analysis and assistance for everyday tasks.
Screenshot → AI assistance
Package: com.snipai.app
View Google Play listing ↗03
PDF CV Builder
AI-assisted résumé creation with a finished PDF export workflow.
Career input → Résumé → PDF
Package: com.pdfcvbuilder.app
View Google Play listing ↗Experience across employed roles & independent work
2021 onwardIndependent software & AI projects
2023–2025AI development, consulting and project delivery
2025AIMX Motors · AI & full-stack development
2025–presentDAL Group · AI Specialist
My contribution spans user needs, application design, backend integration, AI workflows and deployment. Independent projects are separate from employer systems and data.