Lionel R. Panlap
Get in touch
Trajectory Work Capabilities Services Background Get in touch
Open to new roles and freelance projects

I build AI systems that reach production — and I measure whether they actually work.

AI Engineer and full-stack developer based in Bochum, Germany. I've shipped LLM systems into production, evaluated 27 language models for my published research, and I build complete products end to end.

See the work Start a conversation
Lionel Richy Panlap
27

Language models evaluated in my master's research, published on arXiv.

2M

Transaction records analysed at ALDI SÜD to find what drives customer behaviour.

70%

Processing time removed from a data-quality pipeline through content fingerprinting.

300+

Businesses whose data quality and web presence run on software I designed.

How the stack was built

Each capability arrived because a project needed it. Nothing here is theoretical — the bar starts the year I first shipped something with it.

201720192021202320252026
Systems & infrastructureNetworks, Windows Server, Active Directory
2017
Data & analyticsPython, Pandas, statistical modelling, Power BI
2022
Deep learning & researchPyTorch, model evaluation, published study
2022
LLMs in productionFastAPI, structured outputs, Temporal, observability
2025
Agents & orchestrationLangChain, MCP, multi-agent pipelines, n8n
2026
Product & architectureNext.js, Supabase, security hardening, CI/CD
2026

Solid fill marks depth of use; the lighter band marks that the skill stays in active rotation.

Selected work

Three products where I owned the architecture, the AI layer and the shipping. Two are client-facing, one runs inside an organisation.

Live SaaS 2026

NuptialFlow

A wedding platform for couples who plan across families and continents. Guests, budget, committees, seating and schedule live in one place — and each committee only ever sees its own data.

  • Access control enforced in PostgreSQL itself, not just in the interface — a catering lead cannot read the budget, by design.
  • AI seating suggestions from Google Gemini, always proposed to a human before anything is written.
  • Payments confirmed through Stripe webhooks server-side; no key ever reaches the browser.
Next.js 14TypeScriptSupabasePostgreSQL RLSGeminiStripeVercel
NuptialFlow architecture diagram
System architecture — client to services, with the trust boundary marked.
In development Multi-agent AI 2026

Opporterra

An intelligence platform for people investing or building in Africa. It finds real opportunities by country, sector and capital, then tests them against risk and competition.

  • A pipeline of specialised agents: one extracts sources, one scores the opportunity, one weighs risk, one writes the report.
  • Three model providers with fallback, plus live web search, so an answer is never left to a single system.
  • Every claim carries its citation — the point is traceable analysis, not confident-sounding text.
Next.jsGeminiAnthropicOpenAIBrave SearchSupabaseTrigger.dev
Opporterra product and technical architecture
Product and technical architecture — from user profile to exportable report.
Internal tool Data quality 2026

DataCare

Built for Mwemba, a directory of African restaurants, barbershops and grocers across Europe. It keeps the data on 300+ businesses trustworthy, and gives the ones without a website a real one.

  • Fifteen categories of data problems caught automatically — missing fields, near-duplicates found by fuzzy name and location matching, broken encoding, malformed phone numbers and URLs.
  • Content is fingerprinted, so unchanged records are never re-analysed. That alone removed about 70% of the processing time on repeat runs.
  • The language model is called only for genuinely ambiguous cases. Everything else stays deterministic, which keeps results reproducible and costs flat.
  • A four-stage generator writes each business a complete site: model-written copy, the shop's own validated photos, a colour palette derived from its name, assembled through templating.
  • Hardened against cross-site scripting and server-side request forgery, and deployed through the existing Jenkins and Kubernetes setup.
PythonGroqGPT-4.1-miniJinja2JenkinsKubernetes

Open source

Earlier ground-up work, kept public. These are where the fundamentals were learned rather than assembled.

AI-Agent

Autonomous agents built with LangChain to take over repeat work in finance, HR and marketing.

Python

FirstNeuron_DeepLearning

A neural network written from scratch in PyTorch — forward pass, backpropagation, training loop.

Jupyter Notebook

Python_REST_API_With-Flask

A REST API on Flask and SQLAlchemy, with the database layer kept separate from the routes.

Python

Sklearn_MNIST

Classic classifiers on MNIST — KNN, SVM and SGD — compared on accuracy and training cost.

Jupyter Notebook

ReactUndFlask

A React front end talking to a Flask back end, wired the long way to understand the contract.

JavaScript · Python

28 repositories

The rest of the work, from web servers to machine learning experiments.

View on GitHub

What I actually do

Four areas, each backed by work that shipped rather than a course I once took.

Getting language models into production

The hard part is never the demo. It's what happens under load, when the model returns something unexpected, or when the bill arrives. I design for those.

Structured outputsRAGPrompt engineeringFine-tuningLangChainMCP

Knowing whether the AI is any good

I evaluated 27 models for my thesis and learned that automatic scores alone will mislead you. I set up measurement that reflects what users actually experience.

BLEUROUGEBERTScoreG-EvalUser studiesLangfuse

Backends that hold up

APIs, data pipelines and the unglamorous parts around them — retries, timeouts, migrations, tests that catch things before users do.

PythonFastAPISQLAlchemyTemporalPytestDockerSQL

Products, front to back

Interface, data model, payments, access control, deployment. I've shipped the whole chain, which makes me a better engineer on any one part of it.

Next.jsTypeScriptReactSupabaseStripePostgreSQL RLS

Working together

Available for freelance projects alongside permanent roles. Every engagement starts with a call about the problem, not the technology.

Put AI into your product

You know where AI could help but not how to make it dependable. I take it from idea to something your users can rely on.

  • Assistants and chat grounded in your own documents
  • Agents that carry out multi-step work
  • Output validation so answers stay inside the rails
Scoped per project

Build the application

A web product built properly — the interface people see and the architecture that keeps it standing.

  • Web applications and internal tools
  • APIs, databases and integrations
  • Sites and portfolios that load fast
Scoped per project

Make sense of your data

Cleaning, analysis and reporting, so decisions rest on something firmer than instinct.

  • Data quality and automated cleanup
  • Analysis and machine learning models
  • Dashboards people will actually open
Scoped per project

Background

I started in server rooms in Cameroon and moved toward AI as the questions got more interesting.

  • Generative AI Developer, Retail AI GmbHFrankfurt. Shipped LLM-driven analysis workflows, owned the testing strategy.
  • M.Sc. Computer Science, Hochschule Ruhr WestDeep learning focus. Thesis graded 1.3 and published on arXiv.
  • Data Analyst intern, ALDI SÜDCustomer Knowledge Hub. Market basket analysis, segmentation, regional KPI modelling.
  • IT Systems Administrator, CameroonNetworks, servers and Active Directory in live business environments.
  • Microsoft Azure AI Fundamentals & Azure FundamentalsAI-102 in preparation. Also certified in MCP and PCEP Python.
  • DAAD scholarshipAwarded for the final stage of study in Germany.

Tell me what you're building

Whether it's a role you're hiring for or a project that needs a hand — write to me and I'll come back quickly.

panlaprichy@gmail.comEmail — best for detailed enquiries +49 1573 4735727WhatsApp or Telegram linkedin.com/in/lionel-panlapFull career history github.com/lionel-richyCode and side projects