Jakob Grünwald
Full-Stack AI Engineering — Voice AI, agents & full-stack AI
Freelance AI engineer building production voice & chat agents, agentic systems, and full-stack AI products — shipped with tests, evals and CI, not demos.
I design and ship full-stack AI products and voice & chat agents end to end — from the model and prompt layer up to a polished UI and the infrastructure that keeps it running. My production voice and chat agents serve clients across the UAE, Saudi Arabia and Germany, on reusable infrastructure that carries roughly 90% from one engagement to the next — and I build full-stack AI products of my own, like a chat-first résumé agent backed by ~500 tests. Mathematics and computational neuroscience taught me to be precise where it matters — and to prove it.
An audit-first ladder — and you work directly with the engineer who builds it, not a project manager. We start small and low-risk, prove value, then build. Every engagement ships with tests, evals and CI, and you own all the code.
AI Audit & Roadmap
Teams with an AI idea or an existing system who want a clear, low-risk first step before committing budget.
- A written assessment of where AI fits, what is feasible, and what is not
- A prioritised roadmap with scoped phases, risks and effort ranges
- A working proof-of-concept on the riskiest part
You leave with a concrete plan and evidence — and the freedom to build it with me or anyone else.
Build Engagement
Clients who need a full-stack AI product or feature built and shipped — an agentic web app, a RAG system, or a production voice or chat agent — integrated with the systems they already run.
- A production AI product or feature — front end to infrastructure — integrated with your CRM, telephony or messaging stack
- An automated test suite, evals and CI/CD pipeline built from zero
- Documentation and a clean handover so your team can run and extend it
A working system in production with the safety net to change it confidently — and no vendor lock-in.
Handover & Enablement
Teams who want to own what gets built — with their own developers or IT able to run, maintain and extend it, instead of depending on one outside person.
- Complete documentation — architecture, runbooks, and the reasoning behind the design
- Hands-on training for your developers or IT to run, debug and extend the system themselves
- Eval and monitoring setup so model behaviour stays measurable after I step back
Your team owns the system and can maintain it without me — with the option to keep a senior hand on call if you want one.
Every project is scoped individually, but most engagements fall into one of three shapes. These are starting points for a conversation, not quotes.
There is no published day rate. Scope and price are fixed together in a short discovery call, so you know the number before any work begins.
NDA-bound on most client work — described by sector and region. References on request.
Real-estate & mortgage bot suite — UAE
Client engagement · Conversational & Voice AI
Four production bots: inbound WhatsApp and voice plus outbound calling, bilingual English/Arabic. Integrated with a 200+ field CRM and WhatsApp Business, with BANT lead qualification and regulatory do-not-call gating built in. A multi-model pipeline separates dialogue, QA, extraction and formatting — and 1,200+ automated tests hold the whole suite in place.
Dealership booking & qualification bots — Germany & Saudi Arabia
Client engagement · Conversational & Voice AI
Inbound service-appointment booking and outbound lead qualification for premium dealer networks, up to three languages per bot (DE/EN/FR and AR/EN). Live inventory and real-time slot-search APIs let callers actually book — not just get transferred — with CRM lead routing behind it and 180+ automated tests per deployment.
B2B telecom outbound qualification — Germany
Client engagement · Conversational & Voice AI
An outbound qualification bot covering 350+ target cities, extracting 27 structured fields per call. A PostgreSQL name-disambiguation service resolves contacts against a 34,000-name database, so each call reaches the right person and leaves clean, structured data behind.
ResumeTailor
Open-source · Full-stack AI product
A full-stack AI product I built end to end: a chat-first agent that helps people create and tailor résumés. The current version runs on Next.js, the Vercel AI SDK and Drizzle ORM with ~500 tests; an earlier full-stack v1 paired a React front end with a FastAPI / LangGraph backend and a self-built PDF/DOCX generation pipeline — and it is open source. v1 on GitHub →
Happy Coder — voice & UX extensions
Open-source · Voice AI & developer UX
An open-source fork with ~480 commits over roughly seven months. I built a full voice layer from scratch on LiveKit/Pipecat, redesigned the permission UX, and set up CI/CD from zero. React Native/Expo on the front, Fastify and Python behind. View on GitHub →
green-wald.de
Open-source · This very site
This very site: a bilingual, privacy-first static build on Astro 5 — no cookies, no analytics, no third-party requests, self-hosted fonts. Hardened with a strict Content-Security-Policy, a typed EN/DE content layer that keeps the two languages from drifting, and accessible progressive enhancement throughout. Open source, end to end. View on GitHub →
Developer tooling
Personal · how I keep delivery fast & reliable
Not a product I sell — the engine behind ~90% infrastructure reuse, and a window into how I work: custom MCP servers (one exposing 73 tools, backed by 452 tests), a 3-tier TypeScript testing library (902 self-tests), a ~8,800-LOC type-safe workflow framework, and a Claude Code plugin that encodes the full project lifecycle as 20 invokable skills.
I would rather be measured than impressive. These are commitments you can hold me to — not adjectives.
One senior engineer, start to finish
You work with me directly from scoping to handover — no account managers, no hand-offs, no overhead between you and the code.
Weekly written updates
Every week, a short written update on what shipped, what is next, and any open decisions — so you are never guessing about status.
Ships with tests and CI
Code arrives with an automated test suite and a working CI pipeline, so changes stay safe long after the engagement ends.
Reply within one business day
Emails get a substantive reply within one business day, even when the full answer takes longer.
You own all code and IP
You own the code — documented and handed over, free of lock-in, on your own infrastructure. I would rather train your team to run and extend it than leave you a black box only I can touch.
AI ships with evals, not vibes
Model behaviour is measured with evals and structured tests before it reaches production, so quality is a number you can track — not a feeling.
Research-grade rigor on the hard parts
When a problem is genuinely difficult, I apply the discipline behind my peer-reviewed research: define it precisely, test the assumptions, then build.
I came to AI engineering the long way around: a mathematics degree, then computational neuroscience research, then production engineering. That path taught me to sit with hard, ambiguous problems and test my way out of them — which is exactly what shipping reliable AI demands. I read the research, but I am paid to make things work in production.
- MSc, Modeling of Neuronal & Cognitive Systems — Université Côte d'Azur (16.3/20, magna cum laude)
- BSc, Mathematics — Freie Universität Berlin
- Microsoft Certified: Azure AI Engineer Associate
- Computational-neuroscience research with three papers — one published in the Journal of Vision (2023, co-author) and two first-author manuscripts in preparation (2026)
- Research roles spanning TU Berlin, Université Côte d'Azur, CNRS and McGill
Let's build something reliable
Have a full-stack AI product, a voice or chat agent, or a hard tooling problem you want done properly — with tests, evals and a clean handover? Tell me what you are working on and what a good outcome looks like, and I will reply within one business day.