Malte Poppensieker
DE Book a call
AI Product Builder · London · Remote

Most AI projects fail
at the question,
not the code.
I start at the question.

I help companies turn AI opportunities into finished products. Unlike a traditional consultant who stops at the roadmap or a developer who starts with a specification, I can own the full cycle: from identifying the right problem and defining the product through building, deploying, and iterating the solution.

Tell me the problem See the track record
Spec sheet
i start at
the question
scope
strategy → production
you get
analysis, spec, product
you can stop
after any step
guarantee
90 days of fixes, free
headcount
one
Malte Poppensieker
01 - Who I am

I'm the person who normally sits in three chairs on an AI project. I sit in all of them.

I'm Malte Poppensieker. I consulted at Capgemini, managed product at Amazon, ran the international business at Nerdy - the largest online tutoring platform in the US - and then founded an AI learning platform. Master's in computer science, MBA from Cambridge.

That combination is the whole argument. This work usually needs a consultant, a product manager and an engineer. Hire three and the strategy never quite meets the code. I sit in your process workshop in the morning and in the system architecture in the afternoon, so the business case, the specification and the system stay the same thing.

So you don't need to be technical, and you don't need to assemble a team of specialists. Bring an idea, a workflow that's eating your team alive, or a customer problem you suspect AI could solve. I work out whether it can, we decide together what the product should actually be, then I architect it, build it and ship it. And if it breaks in the first 90 days, I fix it.

  • Consultant
  • Product manager
  • AI engineer
02 - How I work

Six steps. The first one is me trying to talk myself out of the job.

Consultant, Product Manager and Engineer. In that order, on one contract. Remote from London, in your standups, in front of your board when it needs to be.

  1. The honest call

    Thirty minutes, video or phone. I'm checking two things: whether AI genuinely helps here, and whether I'm the right person. Often the honest answer is a rules engine, a better form, or fixing the process first. I'll tell you that before you spend money on a model.

  2. Assessment and workshop

    We map the process or the idea together, separate where AI adds value from where it doesn’t, and agree what to build. Most AI projects fail because nobody decided what "done" looks like. You leave owning the specification - readable by people who don’t code, buildable by anyone.

  3. A prototype you can click

    Not a mockup in a deck. An interactive prototype we review together and iterate until it unmistakably reflects the thing you meant.

  4. Architecture, thoroughly done

    Model choice, data flow, evaluation, guardrails, cost. Token economics, latency, failure modes, the price of a human in the loop - decided before the build, not discovered after launch. This is the step whose absence kills projects in month two.

  5. The build, in the open

    Regular demos so we can confirm we're on the right track, and iteration on your feedback throughout. No six-week silence ending in a surprise.

  6. Ship, deploy, stand behind it

    When you're happy, I handle launch and deployment. Find an issue in the 90 days after go-live and I fix it, free. Then handover to your team, or I keep owning it - your call.

You can stop after any step. Plenty of clients take the specification and build it themselves - that’s a fine outcome. And ninety days after launch, bugs are still my problem.
90-day fix guarantee
03 - Track record

I've been the Engineer, the Consultant, the Product manager, and the CEO. Usually not in that order.

Two decades of other people's hard problems: enterprise consulting, big-tech product, two marketplaces built out internationally, and a startup I founded and coded myself.

Brainie

Founder

An AI-backed learning platform for maths. I defined the pedagogy with subject experts, designed the entire system, built significant parts of it myself, and managed a group of external developers.

Amazon

Product Manager

Product management inside one of the most demanding product organisations there is. It is where I learned to define products that stand the test of reality - and delight users.

Capgemini

Consultant

Enterprise consulting: getting inside other companies’ processes, then translating them into something that can actually be built. This is why you don’t have to speak technical with me.

Nerdy

General Manager

The largest online tutoring platform in the US. I was fully responsible for the business outside the United States.

Gett

Global Head of Gett Together

I built the worldwide ride-sharing business - new markets, new operations, from the ground up.

Universität Trier
Master Corporate Computer Science
so the code is built to last
University of Cambridge
MBA
so the business case holds up
04 - Projects

Most of these exist because someone described them out loud to me.

Recent AI work. In every one of them the hard part was deciding what the product should be - the build followed from that decision. Details anonymised where clients asked.

Learning

Maths tutor that refuses to answer

The strategic call was refusing to answer: a tutor that hands over solutions teaches nothing and competes with free chatbots. So it teaches Socratically, works out which prerequisite concept is actually missing, and adapts what comes next. A live subscription product, not a demo.

  • Product definition
  • Adaptive learning
  • Python
  • React
Legal

Precedent finder for court cases

Scans a large corpus of completed cases and surfaces the authorities bearing on one proposition in seconds. A ranked list of documents would have been easy and useless, so every result carries its weight and its subsequent treatment - whether it is still good law.

  • Requirements
  • Retrieval at scale
  • Legal corpus
Generative

Character animation from one prompt

Written intent to finished motion, no manual keyframing. The hard part was never generating movement - it was keeping a character recognisably itself across takes. Consistency and reproducibility were the product, not motion quality.

  • Product decision
  • Generative pipeline
  • Animation
Monitoring

A news scanner with an opinion

Reads 140+ sources continuously and surfaces only what its owner actually cares about. Watch areas are described in plain English rather than keywords, and nothing is surfaced bare: every item carries why it matters and the passage it came from.

  • Product definition
  • Monitoring
  • Filtering
Content ops

Human review, at machine speed

Generation and publication had to be separate systems, not two settings of one: nothing reaches a learner unreviewed. A deliberate cap on throughput, and the reason the content can be sold with a guarantee attached.

  • Workflow design
  • Human in the loop
  • Content ops
No visual shown, at the client’s request.
Finance agent

Reconciliation nobody has to do

Fetches the invoices itself - portals, inboxes, file storage - matches them against the bank line and posts the result. The design work was the confidence threshold: a false auto-match costs far more than an unnecessary escalation. Twenty hours a week down to two.

  • Process analysis
  • AI agent
  • Finance ops
Media pipeline

2,400 explainer videos, no animation crew

The decision that made it work was not using a video model: the visuals are rendered programmatically from the mathematics itself, so they are correct by construction rather than plausible-looking. AI where it is reliable, code where correctness matters.

  • Architecture
  • Generation pipeline
  • Automation
Learning

RoleAI Engineer, Product Manager & Consultant

Brainie - an AI maths tutor for German school years 7 to 10, aligned to the state curriculum. A live subscription product with paying subscribers, not a demo.

I founded it, defined the pedagogy with subject experts, architected the system and built much of it.

Problem to product

An AI tutor that refuses to answer

Problem

Private tutoring works, but it costs around €40 an hour and it is not available at 9pm the night before an exam. Recorded content is affordable and always available, and largely ignored - because it cannot respond to the specific student in front of it.

Why AI

The gap is not content, it is a patient one-to-one dialogue that adapts in real time. That is a genuine language-model capability, and it is the reason a video library can never close it.

Product decision

The system never gives the answer. That single constraint shaped everything: it is what separates the product from a general chatbot with a subject prompt, what makes the pedagogy defensible to parents, and what the pricing and the grade guarantee rest on. Getting there meant deciding what a session is, not just which model to call.

What I built

Interactive explanations across fractions, equations, functions and geometry. An adaptive path per student. Exercises where learners choose their own route to the solution and their own pace. Maths tooling for formulae, geometry and graphs that talks directly to the model underneath. The tutor always gives useful feedback, and never gives the answer.

Architecture

Python/Django and PostgreSQL on the backend, React on the front end. LLM-driven assessment and content generation, with human review before anything reaches a learner. Video pipeline on Celery and Manim.

Result

A live product with paying subscribers. At signup the system projects a target grade and a date, and that projection backs a money-back guarantee - the modelling has to be right, because the business model depends on it.

Skills and deliverables
  • Artificial Intelligence
  • Python
  • React
  • Machine Learning
  • LLM Prompt Engineering
Legal

RoleAI Engineer & Product Manager

A retrieval system that scans a large corpus of completed court cases and surfaces the authorities bearing on one specific proposition - in seconds rather than hours.

I worked with a non-technical founding team from requirements through to a working system.

Problem to product

Research you can actually cite

Problem

Finding the authorities that bear on a specific proposition is slow, and the real exposure is not missing a case - it is citing one that has since been doubted or overruled.

Why AI

Keyword search fails here for a structural reason: the same legal principle is expressed in different words across decades of judgments. Matching meaning rather than strings is exactly what embeddings and retrieval are for. This was a case where AI was clearly the right tool, and I said so early.

Product decision

A ranked list of documents would have been easy and useless. Lawyers need to know how much weight an authority carries and what happened to it afterwards. So every result shows its weight on the proposition, its subsequent treatment, and the passage actually relied on. Nothing has to be taken on trust.

What I built

Retrieval across 1,200+ authorities with jurisdiction, court and date filters. Ranked results with treatment history. Passage-level citation. A case reader that puts the ratio, the relied-on passage and the later treatment on one screen.

Architecture

Python, Django and AWS Lambda, with LangChain orchestration over a vector index of the corpus. React front end.

Result

Question to citable position in a single pass. Taken from prototype to working system with a founding team that had no technical background - I owned the requirements, the product design and the code.

Skills and deliverables
  • Retrieval Augmented Generation
  • Python
  • LangChain
  • Django
  • Artificial Intelligence
Generative

RoleAI Engineer

A fully automated character animation process driven by one written prompt - from plain-English intent to finished motion, with no manual keyframing.

The hard part was never generating movement; it was keeping a character recognisably itself across takes. Locked reference sheets and seeded, reproducible takes solved that.

Problem to product

Consistency, not novelty

Problem

Character animation is expensive because it is hand-posed. Every second of motion is somebody’s afternoon, which puts it out of reach for anyone producing at volume.

Why AI

Generative video can produce plausible motion from a description, which removes the keyframing. That capability is real, and recent enough to be worth building on.

Product decision

The naive version - prompt in, clip out - is a toy, because the character looks subtly different in every take and nothing can be iterated. Consistency and reproducibility were the product, not motion quality. So each character gets a locked reference sheet, every take is seeded, and artists steer with dials - motion amount, reference adherence, duration - rather than by rewriting prompts and hoping.

What I built

A prompt-to-motion pipeline with per-character reference sheets, a style layer applied at render time, motion dials, and take management with seeds, so any result can be reproduced or rerolled deliberately.

Architecture

Generative video models behind a Python orchestration layer, with a render queue and a studio interface over the top.

Result

A brief like "four springy steps toward camera, antenna wobbling on each landing" becomes finished motion without a keyframe being set - and the same character still looks like itself in the next take.

Skills and deliverables
  • Artificial Intelligence
  • AI Development
  • Generative Design
  • AI Video Generation
Monitoring

RoleAI Engineer & Product Manager

A monitoring system that reads 140+ news sources continuously and surfaces only the content its owner actually cares about.

Watch areas are described in plain English rather than as keywords. The system reads each article, judges relevance against those areas, and explains why an item matters with the source passage attached.

Problem to product

Signal, not volume

Problem

Keyword alerts fail in both directions at once. They miss the article that matters because it used different words, and they bury the reader in articles that merely mention the term. The result is a feed nobody trusts and therefore nobody reads.

Why AI

Relevance to an interest like "regulatory pressure on private credit" is a judgement about meaning, not a string match. Making that judgement over a thousand articles a day is exactly the kind of work that only became possible recently.

Product decision

Two calls decided the product. First, watch areas are described in ordinary language, so the user states an interest instead of guessing which words a journalist will use. Second, nothing is surfaced bare: every item carries a short why-it-matters and the passage it was drawn from. An unexplained AI verdict gets ignored on the first false positive - showing the reasoning is what makes it usable.

What I built

Continuous ingestion across 140+ sources on a schedule. Per-article relevance judgement against the user’s watch areas, sentiment scoring, and a written rationale with the source passage. A review interface with tiering, alerts and coverage tracking.

Architecture

Python, LLM classification and summarisation, vector search over ingested content, scheduled ingestion with deduplication.

Result

Over a thousand articles a day reduced to a handful worth reading, each one auditable back to the sentence that triggered it.

Skills and deliverables
  • Artificial Intelligence
  • Python
  • Retrieval Augmented Generation
  • LLM Prompt Engineering
  • AI Agent Development
Content ops

RoleAI Engineer & Product Manager

A workflow that pre-generates complex learning content and routes it through structured human review, so quality stays in human hands while volume stops being the bottleneck.

Subject experts approve, edit or reject in purpose-built tooling with live LaTeX, geometry and graph preview.

Problem to product

Quality stays human

Problem

Covering a curriculum needs far more questions, explanations and worked examples than subject experts can author by hand. But in education a wrong worked example is worse than a missing one - it teaches the mistake, confidently.

Why AI

Volume. Generation gets a draft of every item in place at a rate no team of authors can match, which turns an authoring problem into a reviewing problem.

Product decision

Generation and publication had to be separate systems, not two settings of one. Everything generated lands in a review queue and only becomes active when a human says so. That is a deliberate cap on throughput, and it is the reason the content can be sold with a guarantee attached. Reviewers also needed real tooling - reading raw model output is slow and error-prone, so the builder renders live LaTeX.

What I built

Bulk generation of explanations, questions and solution steps - including interactive geometry and graph elements, images and complex formulae. A review queue with explicit state. Purpose-built reviewer tooling with live LaTeX preview, question and explanation builders, and a full audit trail.

Architecture

Python and Django, React front end, LLM generation, and a workflow engine with state and audit logging.

Result

Quality stays in human hands and volume stops being the constraint. I defined the workflow with the review team and built the generation layer, the queue and the tooling.

Skills and deliverables
  • Artificial Intelligence
  • Python
  • Django
  • React
  • AI Product Management
Finance agent
No visual shown, at the client’s request.

RoleAI Engineer & Product Manager

An agent that runs reconciliation end to end. It fetches the invoices on its own - from supplier portals, email inboxes and internal file storage - matches each one against the bank transaction, and posts the result straight into the accounting system.

The design work was the confidence threshold: a false auto-match costs far more than an unnecessary escalation, so the agent clears only what is unambiguous.

Problem to product

Fetch, match, post

Problem

Reconciliation is mostly confirming what was already obvious, with the occasional genuine judgement call buried in the pile. Worse, most tools assume someone has already gathered the invoices - and that gathering, across supplier portals, shared inboxes and file shares, is the actual work.

Why AI

Two things here resist rules. Reading an arbitrary supplier PDF and pulling out amount, date and reference is a document-understanding problem. Deciding that "INV-4417 / Müller GmbH" corresponds to a bank line with a truncated reference is fuzzy matching over messy text. Both are squarely AI. The posting itself is not, and stayed conventional code.

Product decision

Where to put the confidence threshold - and it is not symmetric. A wrong auto-posting corrupts the ledger and costs hours to unpick; an unnecessary escalation costs someone thirty seconds. So the agent is deliberately cautious, and every escalation arrives with the candidate matches and the agent’s reasoning already attached. The person makes a decision instead of starting an investigation.

What I built

Retrieval from supplier portals, email inboxes and internal file storage. Matching on amount, date, reference and counterparty. Automatic posting to the accounting system for confident matches. An escalation queue with reasoning for everything else.

Architecture

An Openclaw-based agent for extraction and matching, a rules engine for the deterministic checks, and an integration layer into the accounting system.

Result

The unambiguous majority clears without anyone touching it, and the finance team’s remaining work is decisions rather than data gathering. Weekly hours for invoice gathering and reconciliation went from twenty to two.

Skills and deliverables
  • AI Agent Development
  • Artificial Intelligence
  • Automation
  • LLM Prompt Engineering
  • Accounting
Media pipeline

RoleAI Engineer & Product Manager

A fully automated pipeline that produces explainer videos end to end - script, narration, animated visuals, render - with no animation team. More than 2,400 videos produced through it.

I designed the architecture and built it.

Problem to product

Volume without a studio

Problem

Producing explainer videos conventionally means a script writer, a voice artist and an animator for every single one. Output is capped at whatever a small team can hand-make in a week, and covering a full curriculum that way was never going to happen.

Why AI

Script writing and narration are generation problems that language and speech models handle well. Sequencing and pacing are orchestration. Those parts genuinely needed AI.

Product decision

The visuals did not. Generative video is impressive and wrong too often - and in a maths explainer, a plausible-looking but incorrect diagram is worse than no video at all. So visuals are rendered deterministically from the underlying mathematics with Manim. AI where it is reliable, code where correctness matters. That split is the reason the pipeline could be trusted at volume.

What I built

Topic in, finished video out: LLM script generation with review, programmatic animation, synthesised narration, render and publish. A job queue, so throughput scales with compute.

Architecture

Python throughout. LLM generation, Manim for animation, ElevenLabs for voice, Celery and n8n for orchestration and retries.

Result

More than 2,400 videos produced. Output limited by compute rather than headcount - and the quality bar held, because the part that had to be exact was never left to a model.

Skills and deliverables
  • Artificial Intelligence
  • Python
  • LLM Prompt Engineering
  • AI Development
  • ElevenLabs
Space here for yours. Bring the problem - I'll bring the other three job titles.
Start a project
05 - What I do

Half of this job happens before anyone opens an editor.

  • AI feasibility assessment and strategy
  • Business process analysis and automation design
  • Product definition and requirements engineering
  • Solution architecture for AI-native systems
  • Full delivery: build, integration, deployment
  • Handover to your team, or continued ownership
What I build

Seven things, over and over. Yours is probably one of them.

  • AI agents that do real work in existing systems
  • Content generation and processing pipelines
  • Content distillers and scanners
  • Chatbots
  • Customer service automation
  • Knowledge and retrieval systems (RAG)
  • Complex AI-native applications
Technology

Tools are the least interesting part. Here they are anyway.

I pick the boring option unless there's a reason not to - your team has to maintain this after I've gone.

  • Python
  • Django
  • React
  • React Native
  • Serverless
  • Anthropic/OpenAI/Gemini/Deepseek/Moonshot
  • Elevenlabs
  • LangChain
  • LangGraph
  • Vector Databases
  • Openclaw
  • AWS
  • Google Cloud
06 - Four ways to start

The first one is free, and it's the one that might send you away.

You can stop after any of them. Plenty of clients take the assessment or the specification and build it themselves - that’s a fine outcome.

Option one

Exploration call

Thirty minutes, video or phone, no deck. You describe the problem; I tell you honestly whether AI helps here and whether I'm the right person. If I'm not, I say so.

Free
Book a call
Option two · Where most start

Process and opportunity analysis

We map how your business actually works: what happens, where the time goes, what it costs. You leave with a written analysis of where AI adds value and where it doesn’t - each opportunity sized by effort, running cost and risk, including the ones I’d advise against.

£2,600 / $3,600
Option three

Requirements workshop

One working session, alone or with your whole team. We settle what the product actually needs to do, and you leave with a detailed Product Requirements Document. Yours to keep, whoever builds it.

£1,300 / $1,800
Option four

Full development engagement

Prototype, architecture, build, launch - with regular demos throughout and 90 days of free fixes after go-live. One person accountable from first line to deployment.

Fixed price, quoted after the workshop

You've had the idea
for a while. Let’s find out if it’s real.

One call, thirty minutes, no deck. Describe it in plain words and you'll leave knowing whether AI is the right answer and what it would take - whether or not I'm the one who builds it.

malte [at] aiproductconsultant.com Book a 30-minute call
based
London, United Kingdom
languages
English · German
availability
Available as needed - open to offers

Mini-Malte

Malte's answers, on tap

I'm Mini-Malte — Malte's stand-in, so ask me anything you'd ask him. His background, how a project actually runs, what it costs. Or just describe the thing you want built.