AWS Summit · Talk by Ed Dulharu

From Solo Founder
to Full Company.

Building with agentic AI on AWS. One person. Twelve AWS services. Six departments running in production. This is the stack I actually use every day.

1 Founder · 0 Employees · Full Company Ed Dulharu · Founder & CTO, vExpertAI GmbH
Not a sales pitch. Just what actually works on AWS.
// 02 — Who's talking

Context, not credentials.

I run every function of a startup using AI on AWS. Marketing. Sales. Engineering. Finance. Content. Operations. Same person. Same AWS account.

  • 20+
    Years in network & security architecture AT&T · Infosys · Kyndryl
  • EDU
    Lecturer at TU Munich & Politehnica Bucharest Teaching what the industry is doing now, not five years ago
  • 📚
    Author of a 4-volume book series "Generative AI for Networking Engineers"
  • 🚀
    Founded vExpertAI mid-2024 Self-funded · Munich-based · Built entirely on AWS

The point

I run every function of a startup using AI on AWS.

Marketing. Sales. Engineering. Finance. Content. Operations.

Bedrock · SageMaker · Lambda · Step Functions · Neptune · S3.

This is my daily stack.

// 03 — The reality

The numbers behind a one-person company.

1
Founder
No employees, no interns.
6
Departments
All AI-augmented on AWS.
12+
AWS services
In daily production use.
10×
Output multiplier
vs. doing it manually.

// My AWS stack

Bedrock SageMaker Lambda Step Functions Neptune S3 EventBridge SES CloudWatch IAM DynamoDB ECS Fargate
// 04 — Agentic architecture on AWS

A 5-layer guild — fully AWS-native.

Each layer has one job, one set of AWS services, and one clean interface. The same pattern scales from a single agent to a coordinated multi-agent system.

LAYER 1

Orchestration

Multi-agent coordination, routing, state management.

Step Functions · Bedrock Agents
LAYER 2

Foundation Models

Reasoning, generation, tool use via Action Groups.

Bedrock · Claude · Mistral
LAYER 3

Knowledge

Graph relationships, RAG retrieval, document store.

Neptune · Bedrock KB · S3
LAYER 4

Execution

Containerised agent tasks, serverless compute.

Lambda · ECS Fargate
LAYER 5

Observability

Monitoring, cost tracking, responsible-AI controls.

CloudWatch · Bedrock Guardrails
// 05 — Six departments, one person

How each function actually runs.

Every team in a normal company maps to a small set of AWS services plus a few well-trained agents. Here's what's behind each door.

📣 Marketing & Content

solo founder content machine

Outbound automation

Bedrock generates personalised sequences. Lambda orchestrates scheduling. EventBridge triggers follow-ups automatically.

Bedrock + Lambda + EventBridge + SES

Technical content

Bedrock (Claude) drafts from my outline. S3 stores templates and brand voice. 3× faster than writing from scratch.

Bedrock + S3 + Bedrock Knowledge Bases

LinkedIn (DACH)

Bedrock calibrates tone for the German market. 50+ prompt iterations stored in S3. Cultural nuance, not robotic output.

Bedrock + S3 prompt library

4-volume book series

"GenAI for Networking Engineers." Bedrock-assisted research & structuring. SageMaker for domain fine-tuning.

Bedrock + SageMaker + S3

💼 Sales & Outreach

AI doesn't close deals — it gets me to the table 10× faster

DealMind — real-time sales coach

Bedrock Agents feed context, objection handlers, and next-best-action during live calls.

Bedrock Agents + Lambda + DynamoDB

Pipeline automation

AI-powered outbound at scale. Step Functions orchestrate multi-step prospecting workflows end-to-end.

Step Functions + Bedrock + SES + S3

Proposal generation

988-paragraph Airbus DS curriculum generated with Bedrock. Days, not weeks. Knowledge Bases feed the domain context.

Bedrock + Bedrock Knowledge Bases + S3

Pipeline intelligence

Bedrock analyses conversations, suggests priorities, drafts follow-ups. Neptune maps relationship graphs.

Bedrock + Neptune + CloudWatch

🛠️ Engineering & R&D

AI writes production code · I architect & review

5-layer agentic guild

The full pattern: Step Functions for flow, Bedrock Agents with Action Groups calling AWS services, Neptune for relationships.

Step Functions + Bedrock Agents + Neptune

NeT/RED — hybrid cyber platform

Hybrid red/blue-team cybersecurity platform. Network digital twin running on Containerlab on EC2.

EC2 + Containerlab + Bedrock

Domain fine-tuning

Mistral-7B on SageMaker for NOC/SOC domain data. Where prompts hit their ceiling, fine-tuning takes over.

SageMaker + Mistral-7B

Infra-as-code & cost control

CDK deploys everything — reproducible, versioned. Cost Explorer + budgets keep the full stack under $500/mo.

CDK + Cost Explorer + Budgets

⚙️ Operations & Knowledge

the unsexy stuff that makes everything else possible

Knowledge pipeline

599 R&D conversations classified and searchable via Bedrock Knowledge Bases. S3 as the document store, auto-indexed.

Bedrock KB + S3 + OpenSearch

Finance & admin

Bedrock drafts contracts, proposals, invoices. Lambda automates formatting. I review and sign.

Bedrock + Lambda + S3

Daily automation

EventBridge triggers morning briefings. Bedrock summarises overnight emails. Lambda prepares priority lists.

EventBridge + Lambda + Bedrock + SES

Guardrails

Bedrock Guardrails enforce content policies. CloudWatch monitors cost & latency. IAM least-privilege across all services.

Bedrock Guardrails + CloudWatch + IAM
// 06 — A day on AWS

Every hour, an AWS service does the heavy lifting.

This is a real Tuesday. Same shape on a Friday. Most of the work is finished before I open my laptop.

06:00

EventBridge triggers Lambda

Bedrock summarises overnight emails into a single morning brief.

EventBridge → Lambda → Bedrock
07:00

Daily research drop

Knowledge Bases surface the 3 papers that actually matter, out of 20+ new uploads from arXiv & co.

S3 → Bedrock Knowledge Bases
09:00

Customer prep

Bedrock Agents prep briefing docs from the Neptune relationship graph — every meeting walks in pre-loaded.

Bedrock Agents → Neptune → S3
12:00

Content drafting

Bedrock drafts the Substack article from a bullet-point outline stored in S3.

Bedrock → S3 prompt templates
15:00

NOC/SOC inference

SageMaker fine-tuned model runs domain-specific inference for network & security analysis.

SageMaker Endpoints → Lambda
18:00

End-of-day workflow

Step Functions run follow-ups, CRM updates, and tomorrow's priority list. Then I close the laptop.

Step Functions → Bedrock → SES
// Key insight

AWS didn't replace my work. It replaced the work that kept me from doing my real work.

// 07 — Talks & deep-dives I can deliver

Topics to inspire your customers on practical AI.

If you're hosting a customer day, an internal kickoff, or an enablement session — these are the four conversations I run most often. Each one is grounded in a working system, not a slide.

01 / AUTONOMOUS R&D

The daily AI research agent

A custom-built agent that monitors arXiv, Hacker News, and YouTube every morning — and lands a 3-paper summary in your inbox before coffee. Live walkthrough of the pipeline, prompts, and failure modes.

arXiv Hacker News YouTube Bedrock + Lambda
02 / MULTI-AGENT OPERATIONS

5-agent system for network & security ops

How a small fleet of specialised agents handles complex network and security operations — including autonomous fault detection, triage, and remediation. The orchestration patterns, the guardrails, and what breaks first.

Step Functions Bedrock Agents Auto-remediation NOC / SOC
03 / SAFE AI EXECUTION

Digital twins for pre-validating AI actions

Why letting agents touch production directly is the wrong question. Using network digital twins (Containerlab on EC2) to pre-validate every AI-generated action before it reaches a real device. The blast-radius math, end-to-end.

Digital twins Containerlab Blast radius Pre-validation
04 / COMPLIANCE & ON-PREM AI

Local AI for DORA, NIS2 & the EU AI Act

How to deploy AI platforms locally — meeting strict DORA, NIS2, and EU AI Act requirements — without sending a single byte of customer data to an external service. Architecture, trade-offs, and the parts that surprise everyone.

DORA NIS2 EU AI Act On-prem inference
Discuss a session →
// 08 — What actually works

Honest lessons. No hype.

// Do this on AWS

  • Start with Bedrock — don't build infra first.
  • Use Knowledge Bases before reaching for fine-tuning.
  • Lambda for glue, Step Functions for flow.
  • Bedrock Guardrails from day one — not day 90.
  • Check Cost Explorer weekly. Surprises compound.
  • Human reviews everything customer-facing.

// Don't do this

  • Don't fine-tune when prompting still works.
  • Don't build agents before you've nailed the prompts.
  • Don't skip IAM least-privilege. Ever.
  • Don't ignore Bedrock Guardrails.
  • Don't confuse AI speed with AI accuracy.
  • Don't wait for the "perfect" service launch.
// 09 — Cost reality

What this actually costs a bootstrapped startup.

Six departments. Production workloads. Real customers. Numbers, not vibes.

// Bedrock
~$150 /mo
Claude + Mistral inference, Knowledge Bases.
// SageMaker
~$80 /mo
Fine-tuning + inference endpoints (spot).
// Lambda
~$5 /mo
Millions of invocations, mostly free tier.
// Neptune
~$120 /mo
Serverless, scales to zero when idle.
// S3 + others
~$30 /mo
Storage, SES, EventBridge, CloudWatch.
// Misc / glue
~$0 /mo
Most of it is free-tier or pennies.
Full AI stack total
~$385
/ month · six departments · production

That's less than one freelancer's lunch budget — for an entire company's AI infrastructure.

// 10 — The takeaway

AI on AWS won't replace your company.
But a company using AI on AWS
will replace one that doesn't.

You don't need a 50-person AI team. You need one person, one AWS account, and the willingness to start tomorrow.

Ed Dulharu Founder & CTO · vExpertAI GmbH ed@vexpertai.com