Philipp Bühler

Software · AI · Quantitative systems

I design and builddecision systems —software, models,businesses.

Master’s student and founder of Philipp Bühler Digital Solutions. I build web platforms and AI-assisted workflows for real businesses, and I am writing a thesis on how a reinforcement-learning agent compares to classical optimization when the objective is risk-adjusted return.

Selected workResearchContact
Diagram: one path through five connected domainsFive stacked horizontal layers labelled Engineering, AI & Data, Product & Business, Entrepreneurship, Research & Finance, threaded by a single continuous line that passes through one node on each layer.ENGINEERINGAI & DATAPRODUCT & BUSINESSENTREPRENEURSHIPRESEARCH & FINANCE
One path — not five separate skills
02Profile

Short version

The through-line across everything below is narrower than the list of fields suggests: I am interested in systems that have to decide something under uncertainty, and in the engineering that makes those decisions dependable enough to hand to someone else.

That covers a website production pipeline where every stage has explicit gates and the actor that builds is never the actor that reviews. It covers automations that remove a manual step from a small business without removing the ability to see what happened. And it covers a reinforcement-learning agent allocating a portfolio, which is the same problem with a harsher scorekeeper.

The commercial side runs as Philipp Bühler Digital Solutions. This page is not that company — it is the person behind it.

A boundary worth stating

This is a personal site. Philipp Bühler Digital Solutions is a separate company with its own website — mentioned here because it is part of the work, not presented here as an offer.

Based in
Germany
Currently
Master’s thesis · running a one-person practice
Working on
Web platforms · AI workflows · portfolio optimization
Background
KPMG · Mercedes-Benz · self-employed
03Selected work

What I have built

Four entries: one internal system, one company, one piece of research, and one slot deliberately left open.

01In continuous use

WebHustle

Website production pipeline · internal system

Context
Website projects for small businesses repeat the same work — research, design decisions, implementation, review, launch. Repetition is where quality quietly drifts: a step gets skipped because it went fine last time, and the person who built something reviews their own work.
Approach
A structured production workspace where every stage has explicit inputs, gates, and evidence. A reusable design system feeds all projects; per-project state is a tracked artifact rather than memory; and specialised AI agents have separated write scopes, so building and reviewing are enforced as different actors rather than requested as a habit.
Role
Sole author — architecture, operating rules, workflows, implementation

Stack

  • Astro
  • TypeScript
  • Tailwind CSS
  • Claude Code agents & skills
  • Git-based governance
02Operating

Philipp Bühler Digital Solutions

Company

Context
Small and local businesses usually need one working digital thing — a site that earns its keep, a process that stops being manual, a workflow that survives a busy week. What they are typically offered is an agency retainer.
Approach
A one-person practice scoped per project rather than sold as packages: web development and design, digitalisation, AI-assisted solutions, automation of business processes, search visibility, and ongoing maintenance where it is actually agreed.
Role
Founder

Stack

  • Web development
  • Web design
  • AI solutions
  • Automation
  • SEO
  • Maintenance

Company website URL

03In progress — no results yet

DRL Portfolio Optimization

Master’s thesis · research

Context
Mean-variance optimization solves a static problem at every rebalancing date and leans hard on an estimated covariance structure. It is elegant, it is the benchmark for good reasons, and it assumes a stability that markets do not offer.
Approach
Train a PPO agent to allocate across assets under three risk profiles, then compare it against mean-variance optimization and a market portfolio on identical data and identical constraints. Explainable-AI methods are scoped as an optional extension, so the agent’s allocations can be interrogated rather than only scored.
Role
Author — thesis in progress

Stack

  • Python
  • PPO / reinforcement learning
  • Portfolio theory
  • Backtesting
04Reserved — no client project published yet

Client project

Reserved slot

Context
This card is a placeholder, kept visible on purpose. A real client project belongs here — with its actual brief, constraints, and outcome.
Approach
It stays empty until there is something true to put in it. Inventing a case study to fill a grid is the one thing a portfolio must not do, and an obviously reserved slot is more honest than a plausible fiction.
Role
To be filled from a real engagement

Stack

  • To be determined by the project
04Research · master’s thesis

Reinforcement learning for portfolio allocation

The question is not whether a learned policy can beat a benchmark on one backtest — that result is cheap and rarely survives. It is whether a policy trained under an explicit risk profile allocates differently from a classical optimizer in a way that holds up when both are given the same data, the same universe, and the same constraints.

Schematic risk–return comparison, not measured resultsA risk versus return chart showing the comparison being set up. A curve marks the mean-variance efficient frontier and a single point marks the market portfolio. At each of the three risk profiles — Conservative, Balanced, Aggressive — the PPO agent is drawn as a vertical bracket with a question mark, because its return is not yet known. No result is shown.CONSERVATIVEBALANCEDAGGRESSIVEMVOMARKET???RISK →RETURN →
  • Mean-variance frontier
  • Market portfolio
  • PPO — not yet measured
Schematic, not measured. The axes show the comparison being set up — risk against return, three strategies across three risk profiles. Positions are illustrative and carry no claim about outcomes.

Working title

Deep-reinforcement-learning-based portfolio optimization using PPO, compared against mean-variance optimization and a market portfolio.

Strategies compared

  • PPOLearned allocation policy, trained per risk profile
  • Mean-VarianceClassical optimizer — the benchmark
  • Market portfolioPassive reference

Scope

  • Proximal Policy Optimization (PPO)
  • Mean-variance optimization (MVO)
  • Market portfolio as a passive reference
  • Three risk profiles
  • Risk / return trade-off
  • Explainable AI (optional extension)
05Experience

Where the work happened

Roles, periods and specifics are deliberately left as placeholders rather than approximated. They are marked below and are the first thing to fill in.

  1. Start dateCurrent

    Philipp Bühler Digital Solutions

    Founder

    Web development and design, digitalisation, AI-assisted solutions, automation, search visibility, and maintenance for small and local businesses.

  2. Thesis datesCurrent

    Master’s thesis · research

    Author — deep reinforcement learning for portfolio optimization

    PPO-based portfolio allocation across three risk profiles, benchmarked against mean-variance optimization and a market portfolio. In progress; no results yet.

  3. Study periodCurrent

    Master’s studies

    Programme + university

    Ongoing master’s programme; the thesis above is part of it.

  4. Period

    KPMG

    Role title

    Professional experience at KPMG. Role, period, department and focus are to be supplied — nothing about this position is inferred here.

  5. Period

    Mercedes-Benz

    Role title

    Professional experience at Mercedes-Benz. Role, period, department and focus are to be supplied — nothing about this position is inferred here.

06Capabilities

Four kinds of problem

Grouped by the kind of problem rather than by tool. Areas of active work — not a certification list, and deliberately not a wall of logos.

01

Engineering

Building the thing so it survives contact with a second developer.

  • TypeScript
  • Astro & static site generation
  • Component architecture
  • Tailwind CSS & design tokens
  • Accessibility and performance budgets
  • Git-based workflows
02

AI & Data

Getting useful, checkable work out of models.

  • LLM-assisted engineering workflows
  • Agent orchestration & separated responsibilities
  • Prompt and context design
  • Reinforcement learning (PPO)
  • Python data tooling
03

Product & Business

Deciding what is worth building, and for whom.

  • Scoping and requirement clarification
  • Offer and package design
  • Business process automation
  • Search visibility (SEO)
  • Client communication
04

Research & Finance

The quantitative side, and the honesty it demands.

  • Portfolio theory
  • Mean-variance optimization
  • Risk / return analysis
  • Backtesting methodology
  • Explainable AI
07Development

Build activity

Repository and contribution activity. The figures below are sample data, clearly marked — no GitHub account is wired up yet, and inventing public metrics would be the fastest way to make everything else on this page untrustworthy.

SAMPLE DATA — no GitHub account connected. Every figure below is placeholder.
24
Public repositories
1180
Contributions, 12 months
6
Active projects
5
Primary languages

Contribution pattern · 26 weeks

LessMore

Language distribution

  • TypeScript
    38%
  • Python
    27%
  • Astro
    18%
  • CSS
    11%
  • Other
    6%

Repositories

  • sample-repository-oneTypeScript

    Placeholder entry — replace with a real repository.

  • sample-repository-twoPython

    Placeholder entry — replace with a real repository.

  • sample-repository-threeAstro

    Placeholder entry — replace with a real repository.

08Current focus

Open right now

What is actually open on the desk right now.

  • Ongoing

    Building Philipp Bühler Digital Solutions

    Turning a practice into something repeatable: offer structure, delivery standards, and the maintenance side that most one-person operations under-plan.

  • Active

    Developing the WebHustle pipeline

    Extending the production system — delivery skills, independent quality review, and the separation of authorities that keeps a fast pipeline honest.

  • In progress

    Master’s thesis — DRL & portfolio optimization

    Setting up the PPO agent, the risk profiles, and a comparison against mean-variance optimization and a market portfolio that is fair by construction.

  • Continuous

    AI-assisted engineering workflows

    Working out where an agent genuinely reduces effort and where it just moves the review burden somewhere less visible.

09Contact

Get in touch

The fastest route is email. Everything below is a placeholder until the real values are supplied — no address on this page is a guess.

  • Emailmailto linkPreferred
  • GitHubGitHub profile URL
  • LinkedInLinkedIn profile URL
  • CompanyCompany website URLSeparate site