KLEM OS powered by end-to-end AI
KLEM group has designed an ambitious operating system that covers every stage of its value chain — from client intake to creative production to finance. Paravium brings that blueprint to life with a synthetic workforce of six specialised AI agents, coordinated by a single superagent, that live natively inside every layer of KLEM OS and work in parallel with your human teams.
Paravium's architecture rests on applied research developed in partnership with one of Spain's most respected higher-education institutions: the Universidad Politécnica de Madrid — Spain's leading engineering and technology school, consistently ranked among Europe's top institutions for artificial intelligence and computer science. Every agent, every embedding model and every synthetic data pipeline deployed at KLEM rests on peer-reviewed foundations, not on off-the-shelf wrappers around third-party APIs.
The KLEM OS you designed
Five business layers and twenty-one capabilities that describe, with remarkable clarity, what a next-generation agency group needs to operate end-to-end.
Each piece of the OS numbered in order to better map each capability in Paravium's agentic approach.
The twenty-one capabilities, numbered
Client intelligence portals
- 1AI insights — creative & media performance
- 2AI briefing & approvals
- 3Event production centre
- 4Intelligence centre — brand & market
Human+ teams
- 5AI-enabled strategy & planning
- 6Ideation & growth
- 7Growth roles — account & sales
- 8Budget & pitch builder
AI automated delivery
- 9Rapid creative & concepting
- 10Scaled content development
- 11Media planner & optimiser
- 12Influencer AI engine
- 13Event coordinator
Central brain
- 14CRM
- 15RAG
- 16DAM
- Capabilities 14 · 15 · 16 are not modules to build — they are the knowledge layer itself.
AI foundation infrastructure
- 17Financial analysis
- 18Billing & invoicing
- 19AI project coordination
- 20KLEM AI assistant
- Box 20 is the superagent itself.
- 21AI training & development
The synthetic workforce that runs it
Every request — from a human team member or a client — enters through a single superagent. The superagent understands context; the router dispatches work to the right specialists; six agents handle execution. Everything rests on a three-component knowledge layer: multimodal embeddings for semantic search, an entity network (an evolved CRM in the form of a knowledge graph), and object buckets for raw asset storage.
This architecture is our recommended starting point — not a fixed product. The number of specialised agents may grow beyond six, the module catalogue may evolve, and the underlying models will be replaced as better ones become available. Innovation in this field is continuous and dynamic, and the system Paravium delivers is built to pivot with both technological evolution and the learnings we accumulate together.
Paravium research lines that power it all
Digital Ocean GPU infrastructure
The hardware foundation where the agentic system lives. Paravium is a Digital Ocean partner. Data residency chosen at kickoff to match KLEM's regulatory needs; multi-region deployment available from day one.
- H100 × 8 GPU — 640 GB VRAM, 160 vCPU, 1920 GB RAM, 2 TB + 40 TB NVMe
- Full agentic stack: superagent, router, six agents, three-component knowledge layer
- EU-based inference option (SOC2, AES, data transfer compliance)
- Upgradeable to performance tier at any time, no rebuild required
- H200 × 8 GPU — 1.1 TB VRAM, 192 vCPU, 1920 GB RAM, 2 TB + 40 TB NVMe
- More VRAM unlocks larger models, multimodal workloads, fine-tuning without aggressive quantization
- Per-jurisdiction compliance and data residency across regions
- Recommended once KLEM expands beyond a single region or runs heavier models
Dual-compliance by design — Morocco + EU
Deployed from day one against two frameworks simultaneously: Morocco Law 09-08 and the EU AI Act. One compliance posture, broad regional footprint.
Morocco · Law 09-08
Moroccan personal data protection law, aligned with GDPR principles. Data residency in Morocco by default.
European Union · AI Act
Risk-based classification, transparency, human oversight, data governance. All modules shipped EU AI Act-ready.
Regional coverage unlocked
Same system valid across Morocco, EMEA and GCC. United States would require separate HIPAA deployment — out of scope here.
How we inject the KLEM personality without losing best practice
Bias is the mechanism by which the system learns KLEM's character. Calibrated in explicit, measurable stages — from neutral baseline to KLEM-personality-with-best-practice.
Built on synthetic data. No KLEM nor client-specific bias. Establishes the best-practice floor.
KLEM data fine-tuned in. System aligns 70% with KLEM style, questions the 30% against best-practice.
Balance settles closer to 50/50. KLEM steers the dial at any time.
How teams and clients interact with the system
Chat and voice are the default. Structured dashboards layered on where needed. Every workflow drawn jointly at the KLEM–Paravium touchpoints.
Chat and voice
Default across every module. Speak or type — in any language KLEM prefers.
Structured dashboards
Where reporting is needed, a dashboard ships with the module. Layered on chat, never replacing it.
External channels
WhatsApp Business, Telegram, email, Slack — reachable from the channels KLEM already uses. New channels created on demand.
Role-based access
Different users see different views. Access rules per module and per user, adjustable at any time.
Five commitments that set us apart
No black box
No sealed systems. Full knowledge transfer.
Sprints, not consulting
Short sprints, small teams, building the real thing from day one.
Model-agnostic by design
Never married to one provider. The best model wins, every time.
Proven in complex organisations
Hospitals, elite football clubs, health institutions.
Open to external integration
External tools plug in natively via APIs and MCP. Nothing rebuilt if KLEM owns it, nothing locked out.
What Paravium needs from KLEM
The system Paravium builds is not dropped into an organisation — it grows with it. That means we need specific people on KLEM's side, committed to specific roles, for specific windows of time.
Formal training
From week 3 to week 27: 2.5h/week of structured training for KLEM's project management team. Ad-hoc after.
Time investment from KLEM
On top: 4–7h/week time investment · data gathering, normalisation, fine-tuning feedback, knowledge transfer. Not meetings — building the knowledge the modules will run on.
Roles needed from KLEM side
- 1 executive sponsor with decision-making authority — unblocks strategic decisions within 48h, owns the engagement at the top of KLEM. Typically a founding partner or C-level.
- 1 dedicated project lead — single point of contact between KLEM and Paravium. Coordinates, prioritises and translates between worlds.
- 1 data referent per agency (4 total) — surfaces, organises and hands over each agency's historical data (briefs, campaigns, client assets). Heavy engagement months 1–3, lighter after.
- Module-specific power users — one or two per module as they come online, chosen by KLEM leadership. They are the internal champions who make adoption happen.
Risks & how we handle them
No engagement is risk-free. We prefer to put risks on the table from day one.
Ten actionable products, six agents, twenty-one contexts
Twenty-one capabilities become ten actionable modules. Five are absorbed into the platform. Modules 1 and 2 are the first to go live — in production well before November 2026, and exhibited on stage at the 50th anniversary evening. Catalogue non-exhaustive — more modules can be added as needs evolve.
The ten modules below are Paravium's recommended starting point. Every module — scope, inputs, outputs, KPIs, priority — is subject to joint definition at the KLEM–Paravium touchpoints. Nothing ships before alignment.
Note on creative modules. Modules that generate or transform creative content typically require 4–8 additional weeks of fine-tuning vs. analytics or ops modules. Brand voice and creative judgement need more calibration cycles — this is reflected in our timelines.
Voice Brief Studio
- Multilingual voice capture — Whisper + ElevenLabs, tuned for Moroccan Darija
- Structured brief auto-generated: objectives, audience, constraints, budget, timeline
- Playback for confirmation — agent reads back, client confirms or corrects
- Handoff to account team with history and previous briefs already linked
- Embedded in KLEM's client portals or accessible via WhatsApp Business
Indicative ROI
Indicative benefits
- 30–40% faster brief cycle — voice input cuts back-and-forth
- 2–3x brief completeness — structured capture prevents missing fields
- Differentiation vs regional competitors — unique in Morocco today
Self-Service Creative Studio
- Client-facing portal — KLEM-branded, per-client login, scoped to their approved library
- 50+ platform templates — Instagram, TikTok, LinkedIn, display, print, OOH, email
- Reformatting only — does not create new concepts (that's Module 5)
- Brand-guideline enforcement — colors, fonts, logos, tone locked per client
- Pay-per-use revenue — new revenue stream
Indicative ROI
Indicative benefits
- 80% faster reformatting — from days to minutes
- Creative team frees up — 15–25% reclaimed for new work
- New recurring revenue — clients pay for speed; KLEM monetises existing IP
Brand & Market Radar
- Daily brand pulse — sentiment, share of voice, emerging topics, per market
- Creative performance benchmarking across platforms
- Market intelligence — competitor moves, category trends, macro signals
Indicative ROI
Indicative benefits
- 25–35% faster strategic decisions — always-fresh data vs monthly reports
- Upsell lever — insights themselves become a paid client deliverable
Media Command Center
- Cross-channel planning across every paid channel KLEM runs, including influencer allocation
- Budget optimiser with scenario simulation
- Influencer AI engine — match, brief, track, reconcile
- Post-campaign reconciliation auto-generated
Indicative ROI
Indicative benefits
- 12–18% media efficiency gain — better allocation, better attribution
- 50% faster reconciliation — weeks to days
Content Intelligence Lab
- Rapid concepting — idea generation from brief to pitch-ready concepts
- New creative production — visual, copy, video, in KLEM's brand voice
- Trend signal integration — TikTok, X, Instagram, cultural moments
- Brand-guideline enforcement — same engine as Module 2, applied to new creative
- Scope boundary: new campaign creative only — not reformatting
Indicative ROI
Indicative benefits
- 3–5x creative output at constant team size
- Faster time-to-market — from weeks to days on concept rounds
Win Room
- Strategy drafts — hypothesis, audience, positioning, channels
- Ideation engine — concept territories with supporting rationale
- Budget & pitch builder — priced scenarios, exportable to KLEM's templates
- Approval workflow — internal review, client-side sign-off, audit trail
Indicative ROI
Indicative benefits
- 40–50% faster pitch turnaround — win rate up
- Strategy quality uplift — every pitch backed by Radar data, not gut feel
Account Growth Cockpit
- Client health scoring — engagement, sentiment, usage, payment
- Upsell signal detection — cross-agency opportunities flagged
- Meeting prep auto-assembled from CRM, past briefs, latest campaign data
- Churn risk alerts with suggested saves
Indicative ROI
Indicative benefits
- 15–25% upsell rate increase — signals surfaced systematically
- Churn reduced — early warning on at-risk accounts
Event Production OS
- Run-of-show builder with auto-generated backup scenarios
- Vendor matching from historical suppliers, quality, price
- Live event tracker — mobile-first for on-site teams
- Post-event reconciliation — budget, lessons, assets archived
Indicative ROI
Indicative benefits
- 20–30% faster event production cycles
- Lower overruns — early warning on budget and timeline risks
Revenue Operations Hub
- Auto-invoicing from project milestones
- Revenue recognition per campaign, per agency, consolidated
- Financial analysis — margin per client, per account director
- Cash-flow forecasting at group level
Indicative ROI
Indicative benefits
- 60–70% admin time saved — finance team redeployed to analysis
- Real-time margin visibility per client
Project Control Tower
- Project status auto-summarised from Slack, email, docs
- Blocker detection and dependency tracking
- Deliverable reminders routed to the right team
- Continuous training — completed projects enrich the knowledge layer automatically
Indicative ROI
Indicative benefits
- 20–30% fewer status meetings
- Knowledge compounds — every project makes the next one better
The ranges above are indicative, drawn from public benchmarks published in 2025 and 2026. Sources: Gartner, WPP, McKinsey, CrewAI.
Catalogue is non-exhaustive and non-binding
New modules can be designed, split, merged or replaced at any point. The KLEM–Paravium touchpoints is where specs, priority, data dependencies and KPIs are agreed together. Nothing committed to production before that alignment.
Fifty years of words that moved a country. One evening to make words move an AI.
Genius is to last. In 1976 Hamid Kadiri founded KLEM with that conviction. In November 2026 KLEM group celebrates fifty years — announcing a new institutional shareholder, a pan-African expansion, and the launch of KLEM OS.
Everything described in this tab is a proposal, not a plan. It is our current thinking on how the evening could land. Every element is to be refined jointly with KLEM leadership, the production team and the family in the months leading up to November. Consider this a creative brief, not a script.
Three acts. One stage. Fifty years of KLEM — told by its founder, its clients and its future.
A three-act choreographed piece. Modules 1 and 2 used live, on stage, by real clients. At the centre, an AI avatar of Hamid Kadiri opens with the Goethe quote he chose as KLEM's first slogan.
The Origin Voice
1976 → 2026A younger Hamid Kadiri — photoreal AI avatar — walks on stage and speaks the Goethe line that became KLEM's first slogan. He tells the founding story in his own voice, in Arabic or French.
- Photoreal avatar · voice cloned from archives
- Multi-lingual · French, Arabic, English…
- The كلام philosophy explained on stage
The Live Demonstration
2026 · nowAhmed invites a real client on stage. The client speaks a brief into Voice Brief Studio. Self-Service Creative Studio then reformats an iconic KLEM campaign into five platform-ready executions in under sixty seconds — live.
- Real client · real brief · live on stage
- Production modules clients already use
- Iconic KLEM work reformatted in <60 s
The Future Voice
2026 → 2076Ahmed and Hamid take the stage together to announce the new institutional shareholder, pan-African ambitions, and KLEM OS going live across all four agencies. Possibility of an AI avatar of Ahmed — present self or future self — answering audience questions about the vision for the next fifty years.
- New institutional shareholder announced
- Pan-African expansion plans unveiled
- KLEM OS live across the four agencies
- Possibility of an Ahmed avatar taking live audience Q&A on the next fifty years
Two AI avatars — one looking back, one looking forward.
Act I · Hamid avatar narrating the past. Act III · possibility of an Ahmed avatar — present or future self — taking live audience Q&A on the next fifty years. Both built with the people they represent and approved by the family.
What we need from KLEM side
- Interview sessions with Hamid and Ahmed — voice, stories, founding and future narratives
- Archive access — photos, videos, press, fifty years of campaigns
- Creative approval from the family — scripts, tone, boundaries
- Scoping of the Q&A domain — what the Ahmed avatar can and cannot be asked about
What Paravium brings
- Voice cloning — production-grade ElevenLabs, multi-lingual fidelity
- Photoreal facial animation — age variants, real-time rendering
- Live Q&A infrastructure — Ahmed avatar wired into the superagent and knowledge layer, with a safety net of pre-approved answers
- Demo choreography & rehearsals — full rehearsals before the evening
Modules 1 and 2 are not for the evening. The evening is the moment they go public.
By November, both first modules have been in production for months. Dozens of KLEM clients use them weekly. The demo on stage is not a prototype — it is ordinary Tuesday work, performed in front of two thousand people. That is the wow effect.
A roadmap for Year 1
Year 1 has three beats: kickoff · module-by-module build · milestone evening in November.
Kickoff, first two modules in production, then cadence
Kickoff & foundations
- 1-day kickoff in Casablanca · KLEM project team + Paravium team. Module prioritisation, data map, bias targets.
- Infrastructure deployment · Digital Ocean GPU stack provisioned.
- Synthetic dataset generation · purpose-built, for the prioritised modules only.
- Launch and Layer 1 profiling of the agentic stack · superagent and agents for prioritised modules.
- KLEM starts data-gathering · briefs, proposals, case studies, client assets.
Module 1 takes shape
- Module 1 training on synthetic dataset · first three weeks.
- Module 1 training integrating real KLEM dataset · from week 4 onwards.
- Module 2 training on synthetic dataset · in parallel.
- Weekly project team training sessions kick-off.
🎯 Module 1 live · Module 2 takes shape
- Module 1 · fine-tuning, bias calibration and launched to production.
- Module 2 training integrating real KLEM dataset · from week 4 onwards.
The cadence kicks in
Rest of the year follows a module-by-module cadence. Each new module: 8–15 weeks end-to-end depending on complexity.
Identity and order of Modules 3–10 agreed at kickoff, re-sequenceable as KLEM learns. We commit to the rhythm, not to a calendar we would have to revise.
Four phases converging on the 50-year anniversary evening
The anniversary as Year 1 milestone and showcase to capitalise on first months of work.
Foundation phase May – August 2026
Priority: Modules 1 and 2 ready for the showcase. Not demo products — real products that real clients are using in production on the night.
Pre-event phase September – October 2026
- Avatar development · voice cloning, facial animation, Hamid renders for Act I, optional Ahmed render for Act III, 3-act scripts, approval cycle with the family.
- Archive digitisation · 50 years of KLEM indexed, embedded, available as a knowledge layer for the avatar. Stays in the system after the event.
- Dress rehearsal · 1 full technical rehearsal.
Sep–Nov takes real avatar-production bandwidth and may slow one additional module that period. An idea worth surfacing: the avatar layer itself could be productised as a new module in the KLEM OS catalogue, offered to KLEM's clients for their own founders, anniversaries or launches. Same architecture, new revenue stream.
The evening November 2026
Full detail in the 50 Years · WOW tab. What follows is the shape of the evening as currently envisioned — specifics to be defined jointly with KLEM leadership.
- Real clients demonstrating Modules 1 and 2 live on stage
- Founder avatar activated as the central experience of the evening
- Keynote by Ahmed with the superagent as co-presenter
- Paravium engineers on site
Post-event phase December 2026
- Joint debrief KLEM + Paravium
- The evening's learnings feed the next round of modules
- Assets generated are permanently integrated into the system
An applied AI research lab — not a consulting firm
Paravium is a tech-based company spun out of the Universidad Politécnica de Madrid, with more than twenty years of work in agentic AI.
Four people directly accountable to KLEM
Nacho Larriba
Engineer from the Universidad Politécnica de Madrid · 22 years building AI for sectors that cannot afford to fail. ACTÚAUPM 2025 Trajectory Award. Nominated by MIT in 2011 as one of the best engineers in Spain. Professor of Applied AI in Business at UPM International MBA.
Micky Turci
Journalist and digital media specialist. Led the launch of Marca.com, then digital strategy at El Mundo and OkDiario. Leads audience intelligence, segmentation, commercial activation and digital strategy at Paravium.
María Curtichs
Strategy lead on the KLEM engagement. Ex-Bain & Company. Entrepreneur across AI and education ventures.
Carlota Turci
Experience and events specialist with years running large-scale brand production at Volkswagen Group España. At Paravium she closes the loop between the technology and the client-facing experience.
4–5 engineers & ops in the day-to-day
A dedicated support team of 4–5 engineers and ops specialists works the KLEM engagement every day: monitoring, routine maintenance, module-specific fine-tuning and direct escalation.
24/7 platform monitoring
The platform runs 24/7 — AI agents never sleep, and neither does the system that keeps them healthy. Incidents, anomalies and performance drift are caught in real time.
Trusted by organisations across critical verticals
UD Las Palmas — agentic AI deployed across the club
“It is not the big who eats the small, it is the fast who eats the big.”
The same Paravium agentic architecture deployed across every department of the club — a structural parallel with KLEM. Nacho Larriba leads the engagement.
Madrid HQ, partnerships across four continents
Paravium operates from Madrid with global scale, through partnerships in Atlanta, Toronto, the UK and Boston — including a 31-person team integrated with Vijil.ai.
Five operating principles
Technical extension of your team
We integrate as part of your team.
Modular multi-agentic infrastructure
One superagent orchestrator decides which specialists to activate.
Distilled proprietary models
Proprietary distillations of leading foundation models.
Train in North America, infer in Europe
Train in the US and Canada, infer in Europe under SOC2, AES and data transfer compliance.
Total transparency
If we cannot explain it, we do not recommend it.