10
In progress / availableAI Tourism Orchestration — Dynamic workflows on client log analysis
The right message, to the right visitor, at the right time. Dynamic workflows triggered by deep client log analysis.
AI orchestration layer for national or regional tourism ecosystems. Connects destination CRMs, hotel Property Management Systems, airline reservation systems, and analytics. Goes beyond communication orchestration: deep client log analysis (past stays, preferences, behaviors, signals) and dynamic workflow triggering (complementary booking, upsell, room transfer, local recommendation). Delivers the right contextual action at each step of the journey.
Key features
- Integration layer between destination CRMs, hotel PMSs, airline GDSs, analytics
- Deep client log analysis (stay history, preferences, behaviors, intent signals)
- Dynamic workflows triggered by contextual analysis (not just static rules)
- AI orchestration for personalization and smart upsell
- Real-time silo connection (booking, loyalty, preferences, cultural intentions)
Technologies
MCP (Model Context Protocol), LLM, log analysis and contextual scoring engine, CRM, PMS, GDS, data warehouse integrations.
HITL Framework — Human validation loop for AI
Keep humans in the decision on critical cases, at scale.
Industrial Human-in-the-Loop framework: human validation interface on AI outputs, case queue to arbitrate, confidence scoring, human/AI agreement metrics, continuous learning from corrections.
Key features
- Validation interface designed for production rhythm
- Case queue prioritized by criticality and AI uncertainty
- Confidence scoring and configurable thresholds
- Continuous learning from human corrections
- Human/AI agreement metrics and quality reporting
Technologies
React frontend, queue management, LLM integration, real-time metrics, EU AI Act compliance.
18
In progress / availableAI Banking Orchestration — Dynamic workflows on client log
Deep banking client log analysis, contextual workflow triggering: product reco, fraud alert, credit opportunity, complaint management.
AI orchestration layer for banking players: deep client log analysis (transactions, interactions, life events, risk signals) and dynamic contextualized workflow triggering. Product recommendation (savings, credit, insurance) at the right time, fraud alert based on behavior deviation, contextual complaint management, upsell or cross-sell opportunity at the opportune moment. Designed for strict regulatory compliance (KYC, GDPR, banking secrecy, AML).
Key features
- Deep client log analysis (transactions, interactions, behaviors, signals)
- Dynamic workflows triggered by contextual analysis (not just static rules)
- Contextual financial product recommendations (savings, credit, insurance, investment)
- Real-time anomaly and fraud signal detection
- Contextual complaint management with full client history access
Technologies
MCP (Model Context Protocol), LLM, log analysis and contextual scoring engine, core banking, CRM, fraud detection, data warehouse integrations.
Agentic workflows — Multi-agent AI orchestration in production
Run multiple specialized AI agents together, on real business task chains — without breakage and without rogue agents.
Design and operation of agentic workflow chains: several specialized AI agents (research, writing, validation, publishing) coordinated around explicit orchestration (n8n, Airflow, custom). The promise is not multi-agent magic: it is the **discipline of a chain that runs in production every day**, with human checkpoints, decision logging, source allowlists, and inference-cost governance. Validated in production on a real Middle-East economic think-tank case since late 2025.
Key features
- Explicit agentic chain design (no opaque auto-orchestration): every step declared, every LLM call traced, every transition documented
- n8n stack (recommended), Airflow, or custom orchestration on Cloudflare Workers / Node depending on client constraints
- Task-specialized agents (sourced web research, long-form writing, factual validation, CMS publishing, Gamma slide generation) rather than autonomous generalist agent
- Configurable human checkpoints (HITL) on risky steps: final publish, spend > threshold, source allowlist exit
- Source allowlist and complete decision logging: we know why an agent chose this source, this phrasing, this action
Technologies
n8n self-hosted or cloud (default recommendation), alternatively Airflow or custom orchestration. Mixed LLM per step: Claude (long-form writing, reasoning), OpenAI GPT-4o (multi-modality, speed), open-source models via Groq for non-critical steps. WordPress, Gamma, Notion, Google Workspace connectors, business APIs.
25
In progress / availableCustomer dunning and cash forecasting — Graduated reminders drafted by AI, approved by you
Get paid sooner without straining the relationship: a solution built for you at scoping, where AI drafts each reminder and you decide whether it goes out.
Chasing overdue customers is the job nobody wants: it takes method, tact and time, and it always comes last. This solution reads the customer accounts in your ERP or invoicing software, keeps the ageing balance current, proposes for each overdue invoice the reminder that fits its age and the customer's history, and hands it to the account owner: **nothing is sent without their approval**. Expected receipts feed a cash forecast that corrects itself with every payment. The solution is designed and built at scoping, from building blocks we already run internally: reminder preparation inside a CRM (a person sends) and overdue alerts in our invoicing tool.
Key features
- A solution built for you at scoping, from building blocks we run internally: it does not exist off the shelf
- Ageing balance fed from the receivables ledger of your ERP, by age bucket, customer and account owner, refreshed at every receipt
- Graduated dunning scenarios defined with you at scoping: courtesy notice before due date, first reminder, second reminder, formal notice prepared for signature
- Each reminder drafted by AI from the facts of the file: invoices concerned, amounts, due dates, partial payments, payment promises already made
- Approval queue: every proposed reminder is reviewed, edited or rejected by the account owner, then sent from your usual mailbox
Technologies
AI layer on top of the existing ERP or invoicing software (reads customer accounts and due dates, writes back dunning statuses). n8n or equivalent orchestration, language model for drafting, sending from the client's own mailbox (Microsoft 365 or Google Workspace) after approval, ageing-balance and cash-forecast dashboard. Hosting is set at scoping.
Email assistant — Triage, daily digest and suggested replies
Start the day with a mailbox already read, sorted and summarised, and send only what you have reviewed.
An assistant that reads the inbox and the sent folder every day, sets aside automated notifications (while keeping security, billing and deadline alerts), analyses the rest in batches and produces a daily digest, tasks for what is pressing, calendar entries for confirmed meetings and suggested replies. **It never sends anything and drops no draft into the mailbox**: messages awaiting a reply are listed in the digest with a suggested answer, and the person replies from their own mailbox, inside the thread. Used internally every day on an executive mailbox; a deployment on your Microsoft 365 tenant can be scoped with you.
Key features
- Daily reading of the inbox and sent items through the Microsoft Graph API, orchestrated in n8n
- Filtering of automated mail (notifications, newsletters, receipts) with explicit exceptions: security, billing, deadlines
- Batched AI analysis: what needs an action, what is waiting for a reply, what can be filed
- Tasks created in Microsoft To Do for urgent messages, confirmed meetings added to the calendar with conflict checking
- Daily digest in HTML and as an audio version, to read or to listen to on the move
Technologies
n8n (orchestration), Microsoft Graph (mailbox, calendar and To Do access), cloud-hosted language model for the analysis (a local-model variant can be scoped), HTML digest and text-to-speech.