Mehrab

Forward-deployed AI engineering studio

Your AI pilot works.Getting it into production is the hard part.

We join your team, wire AI and automation into your real systems, get it signed off, and hand it over running.

Book a free 30-min callOr grab a virtual coffee
  • Founder-led
  • Senior engineers only
  • Remote or on site
Run · invoice-intakeRunning
  1. Invoice received

    email · INV-20931.pdf

    Trigger
  2. Fields extracted

    14 fields · confidence 0.97

    AI
  3. Matched to shipment

    ERP · SHP-55120

    System
  4. Waiting for reviewer

    amount above auto-approve limit

    Human
  5. Booked in ERP

    same day · no retyping

    Done
Owner
Finance ops
Scope
Carrier invoices < €10k
Cost limit
€0.00 / €0.10 per invoice
Audit trail
0 / 5 events logged

Example run · illustrative data

Problems

Where AI projects stall.

Most AI projects don't fail on the model. They stall on one of these six things, somewhere between the demo and production.

  1. Pilot

  2. 01 · Ownership

    “The demo worked. Then nothing happened.”

    Nobody owns getting the pilot into production, so it sits there.

    What we do

    We own one workflow end to end, until it is live.

  3. 02 · Capacity

    “My team has no time for this.”

    Your engineers are on the roadmap, and hiring senior AI engineers takes months.

    What we do

    Senior engineers in your team from week one, no hiring process.

  4. 03 · Integration

    “Our data and systems are a mess.”

    The model is the easy part. Connecting it to how your business actually runs isn't.

    What we do

    We build in your real systems, APIs and data, not in a sandbox.

  5. 04 · Sign-off

    “Legal won't sign off.”

    No clear answer on approvals, audit trails, GDPR or the EU AI Act, so it cannot launch.

    What we do

    Controls designed in from day one, plus a sign-off pack for legal.

  6. 05 · Quality

    “We can't tell if it's working.”

    No evaluations, no monitoring, no baseline. Quality is a feeling.

    What we do

    Evaluations and a quality baseline before launch, alerts after.

  7. 06 · Cost

    “The bill keeps growing.”

    Model and infra costs climb with no limits and no clear cost per task.

    What we do

    Cost per task tracked, with hard limits that stop the spend.

  8. In production

What we fix

Problems we get hired for.

AI is how we solve them, not the pitch. Pick one to see how it runs today, what we put in its place, and what typically changes.

The problem

Most questions are already answered in your docs, but every one still waits in the queue for a person.

How it runs today

  • Ticket waits in a shared queue
  • Agent searches docs and old tickets
  • Agent writes the reply from scratch
  • Hard cases wait behind easy ones

~14 min of agent time per ticket

What we build

  1. TriggerTicket arrives in Zendesk or Intercom
  2. AIDrafts a reply with links to the sources it used
  3. SystemPulls order and account data through your APIs
  4. HumanAgent reviews, edits and sends
  5. ResultHard cases routed straight to senior agents

~4 min to review and send

Typical change after go-live

First response
0.8 h-87%

was 6 h

Handle time
4 min-71%

was 14 min

Tickets per agent per day
71+122%

was 32

Measured against a baseline we agree with you in week one.

Have one of these? Book a call

How we work

One team across the whole job.

No hand-offs. Product, engineering, data and compliance stay together from discovery to support.

Product
Interviews & shadowing
Engineering
Systems, APIs, constraints
Data & privacy
Data inventory & owners
Governance & risk
EU AI Act class, policies
People & adoption
Users, approvers, concerns

You get · Discover

Brief: workflow, data, risks, success

Start with a free call

Built to pass review

Safe enough for legal to sign off.

Enterprise AI fails review more often than it fails technically. We design the controls in from day one, so security, privacy and legal can approve it.

Every request passes seven layers

Layer 01 / 07

Identity & access

Every request is tied to a real person and role.

  • Single sign-on with your identity provider
  • Role-based access, least privilege
  • Scoped permissions per tool and dataset
Entra ID / OktaRBACSecrets manager
› user=a.devries · role=finance-reviewer · ok

What your reviewers get

EU AI Act

  • Risk classification of the use case
  • Transparency notices for users
  • Human oversight designed in
  • Logging and record-keeping

GDPR

  • Lawful basis and purpose limitation
  • Data minimisation by default
  • DPIA input and data-flow maps
  • Access, correction and deletion requests

Data & contracts

  • DPAs with every AI vendor
  • EU data residency where required
  • Sub-processor and transfer review
  • Retention and deletion schedules

Security

  • SSO and least privilege
  • Secrets and key management
  • Threat model and red-team results
  • Mapped to your ISO 27001 / SOC 2 controls

We're not your lawyers. We work with your DPO, security and legal teams, and give them the evidence they need: data flows, DPIA input, risk classification and audit logs.

Work

What changes when it ships.

Baseline in week one. Numbers at go-live. Client names left out.

European freight forwarder · 1,400 staffSprint · 8 weeks

Invoice intake without retyping

38k carrier PDFs a month. Nine people retyping into the ERP.

Minutes per invoice

11 min2 min-82%

Against the baseline

BeforeAfter
Go-liveW1W7W14

Also moved

Invoices touched by a person
100%14%
Entry errors
4.2%0.6%
Days to book an invoice
6 days1 days

What we did

  • Shadowed AP; mapped 14 layouts
  • Extraction with confidence vs ERP shipments
  • Exception queue with finance approval limits

Stack

PythonClaudeAmazon BedrockPostgreSQLSAP APIAWS

Stack

Every layer of a production AI system.

Tools change every quarter. Knowing where each one belongs doesn't. We're model- and cloud-agnostic and build on what your team already runs.

  1. L1

    Interface

    Where people actually use it

    • ReactProduct UIs and internal tools your team ships
    • Next.jsFull stack web apps with fast iteration
    • TypeScriptTyped frontends and shared API contracts
    • SlackAgent alerts and workflows where teams already work
    • Microsoft TeamsCopilot-style bots inside enterprise chat
    • ZendeskSupport queues wired to triage agents

  2. L2

    Agents & orchestration

    Agents, tool calls and multi-step workflows

    • LangGraphStateful agents with explicit control flow
    • LangChainTool wiring and prompt pipelines
    • Model Context ProtocolStandard tool access for agent runtimes
    • PydanticStructured inputs and outputs on APIs
    • FastAPIAgent services behind your gateways
    • PythonAgent backends and data libraries

  3. L3

    Models

    Chosen per task on quality, cost and where data may go

    • ClaudeAgents and extraction where reasoning quality matters
    • OpenAIGeneral LLM calls and embeddings at scale
    • GeminiMultimodal and Google Cloud native tasks
    • MistralOpen weights where you host models
    • OllamaLocal models for dev and air gapped runs
    • Amazon BedrockManaged models inside your AWS account
    • Azure OpenAIOpenAI APIs with Azure networking controls
    • Hugging FaceOpen models and hosted inference

  4. L4

    Retrieval & data

    Your data, searchable, fresh and governed

    • pgvectorVector search inside the Postgres you already run
    • PostgreSQLTransactional store and RAG metadata
    • ElasticsearchHybrid keyword and vector search
    • LlamaIndexConnectors and indexing over document stores
    • KafkaStreaming ingestion for fresh retrieval
    • AirflowScheduled sync jobs into search indexes
    • dbtCurated tables agents query with confidence
    • SparkLarge batch feature and document pipelines
    • SnowflakeWarehouse side analytics for agents
    • BigQueryGCP warehouse queries in tool calls
    • MongoDBFlexible documents behind agent tools
    • RedisCaches and rate limits on hot paths

  5. L5

    Cloud & platform

    Production infrastructure, as code

    • KubernetesRun agent services with your SRE patterns
    • DockerReproducible images for every deploy
    • TerraformIaC for queues, stores and model access
    • AWSBedrock, Lambda and data in your VPC
    • Google CloudVertex, BigQuery and GKE workloads
    • AzureOpenAI, Entra and your existing subscriptions
    • GitLab CIPipelines that gate prompts and evals
    • GitHub ActionsCI for agents alongside app code
    • KeycloakSSO and roles in front of agent APIs
    • GoFast gateways and platform sidecars
    • Node.jsBFFs and realtime interfaces to agents

Evaluation & guardrails

Watches every layer, before and after launch

  • LangSmithTraces, datasets and regression on prompts
  • OpenTelemetryStandard traces across agent spans
  • GrafanaDashboards on latency and eval scores
  • PrometheusMetrics and alerts on agent SLOs
  • DatadogAPM on production agent traffic
  • SentryError capture on tool and model failures
  • Golden sets
  • LLM-as-judge
  • PII redaction
  • Cost limits
  • Audit logs

Running something else? We build on your stack, not ours.

Why Mehrab

Senior help. No hiring cycle.

Hire full-time
Typical agency
In-house team
Mehrab
Starts in
months
weeks
whenever there's time
this month
Who builds it
new hire
rotating team
existing team
senior engineers, founder-led
Owns production
eventually
rarely
yes
yes, until handover
Knowledge stays
yes
often leaves
yes
yes, documented and paired
Commitment
permanent
contract per project
ongoing
scoped, flexible

Engagements

Ways to work together.

Start small or embed for months. Every option is led by the founder and ends with something running in production, not a slide deck.

Early or stuck projects

Advisory days

A few daysA year
Length
1–5 days
Team
The founder
How it starts
A call and read access to your docs
  • Architecture and risk review of what you have
  • A written plan with costs and build-vs-buy calls
  • A clear go or no-go for production
Most common

One workflow that has to ship

Fixed-scope sprint

A few daysA year
Length
4–8 weeks
Team
Founder + 1–2 senior engineers
How it starts
A discovery week, then a fixed price
  • That workflow live in production
  • Evaluations, monitoring and cost limits in place
  • Sign-off pack for legal, runbooks for your team

Building an AI capability

Embedded team

A few daysA year
Length
3–12 months
Team
Senior engineers inside your team
How it starts
Usually after a sprint
  • Several use cases shipped on shared foundations
  • Your engineers paired with ours, so the knowledge stays
  • Help hiring the people who take it over

Agencies and platform teams

Contracting

A few daysA year
Length
Daily or hourly
Team
One senior engineer
How it starts
Whenever you need hands
  • Hands-on delivery in your process and tools
  • AI, platform or data engineering
  • Flexible scope, no minimum term

How most clients start

Advisory daysSprintEmbedded team

Sprints are fixed price after the discovery week. Advisory, embedded and contracting are day rates. We work out the numbers on the call.

Book a call
Mohammad Hossein Malek

About

Founder-led. Senior only.

Mehrab is led by Mohammad Hossein Malek, a forward-deployed AI engineer with 12 years in software. For the last three he has worked inside enterprise teams: scoping the problem with them, building agents and retrieval on their data, and taking it through rollout. He has shipped across supply chain, telecom, aviation, healthcare, mobility and payments. He leads every engagement personally and brings in vetted senior engineers when the scope needs more hands. No juniors learning on your project, no hand-offs to people you haven't met.

The name means “good”: good work, done properly, and still running after we leave.

LinkedIn →

Amsterdam · Dubai · LondonRemote or on site across EuropeEnglish12 years engineering

Stack

PythonTypeScriptLangGraphClaudeOpenAIAmazon BedrockpgvectorKafkaKubernetesTerraformAWSGoogle Cloud

Questions.

Tell us where it's stuck.

A free 30-minute call. You'll leave with a clear next step, whether or not we work together.

Book a free 30-min callOr grab a virtual coffee
+31 6 1868 6189WhatsAppmhos.malek@gmail.com