TPL TPL · AI ADOPTION BRIEFING

Act One.
The Transformation

Built over eight years. Operated by AI today.
How a 27-person company became AI-native.
27
People run
the whole company
One roof in Patras
125+
Countries
we ship to
From Patras and 11 EU branches
90%
International
revenue
Of ~€4M turnover — Greece is 10% of the business
2006
Roots in
Telematica
Two decades of commerce infrastructure
Knowledge. Second Brain. Private AI. Build. Product.
Spyros Kotsalidis · CEO
01COVER
2026
02POSITIONING
AT FIRST GLANCE
Look closer

At first glance:
What does TPL look like?

01
At first glance, we look like a logistics company.
02
We take products from vendors, list them on marketplaces, manage orders, handle fulfillment, and serve the end customer.
03
Some call us e-commerce, distribution, or 3PL.
04
But that is only the surface.
TPL · 2026
03POSITIONING
THE PARADOX
The Paradox

125 countries.

0140,000 orders
0290% international
03€4M revenue
0427 people
0530% YoY growth in 2026
This should not be possible from Patras.
TPL · 2026
04THE REVEAL
AI NATIVE COMPANY
This is what flying low, hitting hard looks like.
The New Positioning

TPL is a logistics company.

TPL
evolved to an AI-native
commerce operator.

Key Point
We are the operating system for brands that want to be global — without becoming global.
The AI-powered cross-border commerce platform that helps brands promote, sell and deliver products internationally.
TPL · 2026
05PROOF
TECHNOLOGY
Proof

The platform is already built.
Now AI runs it.

Integrated technology. Connected operations.
Built for modern commerce.
01
Cloud-based platform with marketplace integrations
Amazon · eBay · Etsy · Google · Meta · Pinterest · Bing
02
Product Data Flows & PIM Capabilities
Centralized, accurate product information across all channels.
03
Digital Marketing Connectivity
Seamless integration with leading advertising and analytics tools.
04
Logistics & Fulfillment Operations
End-to-end shipment management and real-time tracking.
05
Customer Service Workflows
Omnichannel support with efficient case management.
06
Our Team
Software developers, systems engineers, e-commerce specialists, operations & support.
Built over 8+ years. Operated by AI today.
TPL · 2026
06EVOLUTION
AI
Evolution

From manual complexity
to AI-orchestrated operations.

TPL operates in extreme complexity — hundreds of products, multiple vendors, dozens of marketplaces, multiple countries. AI doesn't add capability. It re-orders the chaos.
YESTERDAY — MANUAL COMPLEXITY
Hundreds
of Products
Multiple
Vendors
Dozens of
Marketplaces
Multiple
Countries
Thousands of
Pricing Variables
Logistics
Decisions
Customer Service
Events
TOMORROW — AI-POWERED INTELLIGENCE
Predict DemandSmarter forecasts. Better decisions.
Optimize PricesDynamic pricing that maximizes profit.
Automate ListingsFaster, accurate, and scalable.
Calculate Real MarginsTrue profitability across all variables.
Detect Logistics IssuesProactive alerts. Fewer disruptions.
End-to-End
Intelligence
From data to decisions.
Higher
Accuracy
Fewer errors. More confidence.
Operational
Efficiency
Automate. Scale. Grow.
Sustainable
Growth
Profitable today. Stronger tomorrow.
TPL · 2026
The Thesis

AI doesn't add capability.
It re-orders the chaos.

AI transformation is not a procurement exercise. It is re-structuring how the company knows, builds, and decides — in that order. Tools come last.

[TBD: one-line contrast — "what most companies do" vs "what we did"]
07THESIS
THE THESIS
TPL · 2026
TPL TPL · AI ADOPTION BRIEFING

Act Two.
The Playbook

Not tools first. We built the plumbing, then the brain, then let real use cases prove it.
How we designed the transition.
01
Infrastructure
first
The north star, private AI tiers, build-with-AI pipeline
02
Then the
second brain
Knowledge federation, Onyx retrieval, one home per fact
03
Then real
use cases
Shown live, on our own sites, in production
Infrastructure. Second Brain. Real Use Cases.
08ACT 2
THE TRANSITION
09THE CATALYST
MODELS & PARTNERSHIPS
Growth Accelerator

Models & Partnerships.

Powerful frontier intelligence combined with secure local execution.
PARTNER NETWORK
Anthrop\c
LLM Provider
Official Partnership with Anthropic — leveraging state-of-the-art Claude models and enterprise infrastructure.
Anthropic
LOCAL SECURE LLM
NVIDIA Spark DGX
Local Secure Execution
Local execution powered by NVIDIA — leveraging the NVIDIA Spark program for highly optimized on-premise model performance.
NVIDIA
CLOUD LLM
Amazon Bedrock
AWS Infrastructure
TPL joined AWS startup AI accelerator program
AWS
Frontier Intelligence. Secure Execution. Connected Knowledge.
TPL · 2026
10THE FOUNDATION
INFRASTRUCTURE
Own The Metal

Infrastructure as code.

Two production clusters, fully documented in git — every decision, runbook, and config is a versioned file the AI can read.
PRIVATE CLOUD
infra-okd
OKD / Kubernetes Cluster
The AI-era platform. OKD 4.18 (Kubernetes) on 4× Dell M630 blades in a VRTX chassis, plus an R720 services host — runs Onyx RAG, the developer platform, monitoring, and the private AI workloads. Dual-Starlink WAN, domain okd.tpl.one.
DocsArchitecture, master plan, 50+ ADRs, numbered runbooks
PatternSingle source of truth — repo drives the hardware
services.okd.tpl.one →
PRODUCTION VIRTUALIZATION
infra-hyperv
Hyper-V / Windows Cluster
The legacy estate, documented the same way. Microsoft Hyper-V on 4× Dell M640 blades in a second VRTX chassis — hosts the business-critical VMs, including the production and dev K3s clusters. Live system: read-only diagnostics first, a written production-safety policy before any change.
DocsADRs, runbooks, read-only PowerShell diagnostics toolkit
PatternSame repo structure replicated from infra-okd
services.hyperv.tpl.one →
Everything In Git. Runbooks & ADRs. AI-Readable Infrastructure.
TPL · 2026
11VISIBILITY
MONITORING
Eyes On Everything

Watching the metal.

If a human has to SSH in to ask "is it healthy?", the monitoring isn't done yet — every layer reports somewhere.
METRICS
Grafana
Dashboards
Host + NTP metrics on the r720 services host, alongside OKD's own in-cluster Prometheus/Thanos.
infra-okd · live
Grafana
grafana.okd.tpl.one →
TSDB
Prometheus
Time-Series Store
chrony_exporter + node_exporter scraping on r720; a second Prometheus runs inside the OKD cluster itself.
infra-okd · live
Prometheus
prometheus.okd.tpl.one →
SNMP / NMS
Zabbix
Hardware & Incidents
Polls the Dell management plane — VRTX CMC, iDRACs, the internal switch, the PDU, UniFi — with vendor templates and alerting.
infra-okd · live
Zabbix
zabbix.okd.tpl.one →
AI USAGE
SigNoz
OTel APM & Cost
Org-wide Claude Code usage — per-developer tokens, cost, sessions, traces, and logs, OTel-native into ClickHouse.
infra-okd · live
SigNoz
signoz.okd.tpl.one →
Prometheus & Grafana. Zabbix SNMP. SigNoz AI Usage. infra-hyperv: read-only diagnostics toolkit, full stack pending.
TPL · 2026
12THE SUBSTRATE
PLATFORMS
What It Runs On

The platforms underneath.

Three layers of hosting — a Kubernetes platform for the new, hardened virtualization for the critical, and containers everywhere in between.
CONTAINER PLATFORM
OKD
Kubernetes / OpenShift
The AI-era platform. OKD 4.18 (the community OpenShift distribution) on 4× Dell M630 blades — runs Onyx RAG, the developer platform, monitoring, and the private AI workloads, managed through the web console.
infra-okd · live
OKD
console-openshift-console.apps.okd.tpl.one →
VIRTUALIZATION
Hyper-V
Windows Server VMs
The production estate. Microsoft Hyper-V on 4× Dell M640 blades hosts the business-critical VMs — including the production and dev K3s clusters — with a written production-safety policy on every change.
infra-hyperv · live
Hyper-V
services.hyperv.tpl.one →
CONTAINER RUNTIME
Docker
Containers & Compose
The common unit. Docker containers package the services that run inside the Hyper-V VMs and on the services host — the same image runs on a laptop, a VM, and the cluster.
both clusters · live
Docker
OKD / Kubernetes. Hyper-V VMs. Docker Containers.
TPL · 2026
The Destination

Where we want to get.

The north star: an AI-native company whose second brain doesn't just answer questions — it runs processes autonomously. We are not there yet. This talk is the honest map of the steps we've already taken.
WHERE WE ARE·2026
AI-assisted, human-approved
Second brain answers
390 docs indexed, everyone asks it first — but it only answers
Agents draft, humans approve
Session logs, ADRs, nightly findings — nothing auto-commits
Non-devs ship apps
Guard-railed pipeline, review gates on every deploy
First agents in Shadow/Supervised
Nightly knowledge agents, QA analyst — watched, measured
→
THE NORTH STAR
AI-native, autonomously operated
Second brain runs processes
Order desk, listing QA, returns triage — end to end, autonomous
Agents earn autonomy
Shadow → Supervised → Autonomous, rung by rung, per process
Humans set direction & exceptions
People curate, decide, handle the edge — the grind is gone
Every process measured
Eval harness + cost ledger before any rung is climbed
We are not there yet — and anyone who tells you they are is selling something. What follows are the steps we have actually taken.
13DESTINATION
WHERE WE'RE GOING
TPL · 2026
Step 1 · Knowledge First

Before any model: every fact
gets exactly one home.

What we built
01
A federation of knowledge vaults — one router, ~10 thematic vaults
02
Single source of truth — no fact lives in two places
03
Typed, machine-readable knowledge (structured frontmatter, wikilinks)
Who does the bookkeeping
04
The AI agent — not the humans — writes session logs & handoffs
05
Decisions become ADRs, incidents become runbooks, on request
06
The human curates; the agent files
Explore the live federation graph graph.okd.tpl.one →
14STEP 1
KNOWLEDGE FIRST
TPL · 2026
Step 2 · The Second Brain

The company's internal Google —
self-hosted, always current.

What it is
01
Enterprise RAG (Onyx) over the entire knowledge federation
02
390 documents indexed · re-synced seconds after every save
03
Multilingual retrieval — Greek and English, one index
How people use it
04
Public assistant — frontier cloud model over shareable knowledge
05
Confidential assistant — local model, data never leaves the building
06
Nightly agents that find gaps, propose links, challenge stale docs
Ask the second brain onyx.apps.okd.tpl.one →
15STEP 2
SECOND BRAIN
TPL · 2026
Step 3 · Private AI

One question decides the model:
can this data leave the building?

The tiers
T0–T1
Cloud frontier — Anthropic Claude where confidentiality allows
T2
Local GPU — our own DGX Spark box. Confidential data never egresses
T3
Never ingested at all — some things don't belong in any index
The posture
01
Model-agnostic orchestration — cloud, on-prem, AWS Bedrock behind one layer
02
No provider lock-in — frontier where it matters, local where it must
03
Fail-closed: unlabeled data defaults to confidential
[TBD: tier table as a visual · local model tok/s numbers if we want hardware cred]
16STEP 3
PRIVATE AI
TPL · 2026
Step 4 · Build With AI

Non-developers ship real apps.
Developers ship faster.

Creators
01
Semi-technical staff build internal tools entirely with Claude Code
02
Guard-railed pipeline: validate → security → build → staging → prod
03
"You do not need to know git or the terminal" — the platform's own onboarding doc
Developers
04
Dual AI dev pods per developer — subscription + Bedrock
05
AI writes the docs too: context files, changelogs, decision records
06
Human review stays in the loop — AI drafts, people approve
17STEP 4
BUILD WITH AI
TPL · 2026
The Proof

Built with AI —
not by the dev team.

One hackathon day and the projects it unlocked. Real tools, real users, real production.
HACKATHON · 2026-05-15
TPL Claude Challenge
One day · VIOPA Cowork
The Content team — zero developers — built a 10-agent product-content swarm in Claude Desktop. Live demo on a real SKU, 45–90s per product. Next morning: production project.
01 · LIVE PRODUCTION
Mission Control
Sales · financial · marketing analytics
Real-time dashboard over the ERP — 16 views, P&L, three ad-platform integrations. The first creator-platform app, built with Claude Code.
mission-control.tpl.one →
02 · HACKATHON → PRODUCTION
AI-Native CMS
10-agent content swarm
Excel + images in → enriched product content in 6 languages, per channel (Amazon, Google, eBay, Shopify). Born at the hackathon, now the PIM/CMS reference architecture.
03 · IN BUILD
FeedForge
Marketing self-service
Marketers map product data to marketplace XML feeds themselves — a deterministic engine ships them on schedule. Replaces the developer-per-supplier bottleneck. Built by a creator.
feedforge.okd.tpl.one →
04 · EARLY
BD Advisor
Business development
AI-assisted advisor for the business-development workflow — started June 2026, early scaffold.
bd-supplier-advisor.tpl.one →
05 · CLIENT ENGAGEMENT
Vanos Project
The playbook, sold
A 12-month AI-first transformation program for a client: their own second brain, a Mission Control site, a wiki-chat assistant. Same method you're watching now.
vanos.tpl.one →
The dev team is 7 people. The builders on this slide are content, marketing, and leadership.
18PROOF
BUILT WITHOUT DEVELOPERS
TPL · 2026
The Guardrails

Freedom with structure —
autonomy is earned in stages.

The rules
01
One written AI policy for everyone: what goes to cloud, what stays on-prem, what never gets ingested
02
No customer PII reaches external models without review
03
Fail closed: unlabeled data defaults to confidential
The ladder
S1
Shadow — the agent watches and drafts; output goes nowhere
S2
Supervised — the agent acts; a human approves every step
S3
Autonomous — earned per process, never granted by default. No agent skips a rung
19GOVERNANCE
THE GUARDRAILS
TPL · 2026
The Discipline

If it isn't measured,
it isn't deployed.

01
Every AI feature ships with an eval harness and a cost dashboard — no exceptions
02
Org-wide observability: tokens, cost, sessions per developer and per model
03
A 32-question nightly eval as the thermometer of the second brain
04
Content stays redacted — we measure usage, never prompts
[TBD: one real chart from the usage dashboard (redacted) as the money shot]
20MEASUREMENT
MEASURE EVERYTHING
TPL · 2026
The Scars

What broke.

A fail-open default once exposed 153 infrastructure documents to the wrong confidentiality tier. We caught it, closed it, and made fail-closed the system invariant. If a transformation story has no scars, it hasn't shipped.

[TBD: pick 2–3 more honest lessons — Greek retrieval fix, stale handoffs, what we'd skip next time]
21HONESTY
WHAT BROKE
TPL · 2026
The Takeaway

Five steps. In order.

01
KnowledgeFoundation
One home per fact. The agent does the bookkeeping — session logs, decisions, handoffs.
02
Second BrainRetrieval
RAG over everything, current within seconds. Everyone asks it before asking a colleague.
03
Private AITrust
Tiered models — frontier cloud where allowed, local GPU where it must. Confidential data never leaves.
04
BuildVelocity
Everyone ships with AI inside guardrails — content, marketing, leadership. Not just the devs.
05
ProductBusiness
AI moves from how you work to what you sell — each feature gated by evals and a cost ledger.
+
Wrapped inAlways on
Governance (Shadow → Supervised → Autonomous) and measurement (evals, cost, usage) around every step. We also run this playbook as engagements for other companies.
22PLAYBOOK
THE PLAYBOOK
TPL · 2026
Our Road Through AI Adoption

The road is
repeatable.

We didn't buy a transformation. We built one — and wrote down every step.

Presented by
Spyros Kotsalidis
Role
CEO
Contact
kotsalidis@tpl.gr
Web
tpl.gr
23CLOSE
THANK YOU
TPL · 2026