Open to AI Engineer rolesBarcelona

I take LLM systems from a client's problem to production.

AI Engineer and Forward Deployed Engineer. At Twin AI I built LLM pipelines for enterprise clients and owned a client's product end to end. Before that, NLP and ML projects with research partners, two of them award-winning.

See the work Get in touch

mark@artichoke:~ — Konsole

[mark@artichoke ~]$

shipping LLM systems at an AI startup, promoted to FDE: 1.5 yrs; shipping LLM systems at an AI startup, promoted to FDE
BSc Computational Social Science, University of Amsterdam: 8 GPA; BSc Computational Social Science, University of Amsterdam
led on partner projects, both awarded: Teams of 5 and 4; led on partner projects, both awarded
based in Barcelona, no visa or sponsorship needed: EU citizen; based in Barcelona, no visa or sponsorship needed

Selected work

What I built, and how it runs

01Twin AI

Grant applications in two minutes, not days

AI Automation Engineer · Mar – Oct 2025

An LLM platform that drafts grant applications from the grant's rules and the client's own documents.

per application, down from days: 2 min; per application, down from days
enterprise clients live: 4; enterprise clients live
LLM-judge score on every section: 1 to 10; LLM-judge score on every section

PythonTypeScriptOpenAI / Claude / Gemini APIsAzure OpenAI (EU)RAGFull story →

02Twin AI

A pitch deck from partner data, on demand

Forward Deployed Engineer · Nov 2025 – Feb 2026

A generator that builds a partner's pitch deck from Salesforce, the product catalogue and Bynder, with a reason for every choice.

of manual work per deck, now on demand: 4–5 h; of manual work per deck, now on demand
systems wired in: 3; systems wired in
from discovery to weekly releases: Solo; from discovery to weekly releases

Next.jsTypeScriptSalesforce APIBynder APILLM pipelinesClient discoveryFull story →

Bynder
Approved shots
Product catalogue
SKUs and prices
Salesforce
Partner orders

Partner, products and ids invented for this picture.

03PestoSide project

One system for a whole restaurant, and AI for the typing

Full-stack developer · 2026 – now

Full story →

A restaurant platform that puts menu, till, kitchen and payments in one system. I build the AI that does the typing.

of the paper menu becomes a structured menu: 1 photo; of the paper menu becomes a structured menu
languages per menu, translated while you edit: 10; languages per menu, translated while you edit
waiting: parsing runs behind onboarding: 0 s; waiting: parsing runs behind onboarding

Next.js 16TypeScriptSupabase / PostgresGoogle VisionClaude APIPlaywright

Restaurants should be about food and guests.

Instead, they're buried in 43 tasks of software, compliance and paperwork.

Pesto puts them in one system, and lets AI do the typing.

Menu AI

From a photo of the paper menu, in any language. Guests order and pay by QR.

Menu from a photoTranslationsAllergens, nutritionQR ordering

Floor, kitchen, stock AI

Till, floor plan, kitchen screens, stock and reports in one place.

Tech cards draftedDelivery from a photoKitchen screensReports

Payments & receipts

Card payments at the table, and the fiscal receipt filed by itself.

Split bills by phoneReceipts filed

04QuotaClimat × University of AmsterdamCapstone

How 781 outlets frame the climate crisis

Technical Lead & ML Engineer · Feb – Jun 2026

The scraper, classifiers and dashboard behind a French media watchdog's evidence to the regulator. I led a team of five.

macro F1 in French: 0.82; macro F1 in French
articles classified: ~20k; articles classified
place, Environmental Award: 2nd; place, Environmental Award

PythonPyTorchXLM-RoBERTaFastAPINext.jsPostgreSQLDockerFull story →Code on GitHub →

Dashboard: coverage over time, frame distribution, editorial angles and sectors

One article with the words behind its frame highlighted, and its frame and angle probabilities

The reader: articles filtered by frame, angle, sector and outlet

Frame coverage ranked by outlet and by country of origin

Demo data. Outlet names are invented.

05TRAC FM × University of AmsterdamProof of concept

Radio call-ins in Ugandan languages, as themes on a dashboard

Pipeline design · 2026

Full story →

A pipeline design that turns radio call-ins in Ugandan languages into English themes, per question and region.

stages, each with a swappable model: 6; stages, each with a swappable model
per new model in the benchmark: 1 column; per new model in the benchmark
against the anthropologists' gold set: WER; against the anthropologists' gold set

WhisperFFmpegWebRTC VADEmbeddings + HDBSCANLLM translationReactDocker

Plays both in turn.Proof-of-concept design. Stations, calls and scores are illustrative.

Proof-of-concept design. Stations, calls and scores are illustrative.

06OceansX × University of AmsterdamPartner project

What well-being sounds like, in 600 hours of interviews

Machine Learning Engineer · Sep 2024 – Feb 2025

Full story →Code on GitHub →

Three model families on ~600 hours of interviews, mostly with sailors, and LIME and SHAP to show why they decided.

F1, BERT: 0.89; F1, BERT
of interviews transcribed: 600 h; of interviews transcribed
place, Well-Being track: 1st; place, Well-Being track

PyTorchHugging Facescikit-learnWhisperLIMESHAP

Beyond the degree

My BSc was light on maths and systems, so I added master's-level courses

Distributed & Parallel Programming
CUDA in C on the DAS-5 supercomputer's GPUs, plus pthreads, OpenMP, MPI across nodes and a Spark cluster in Docker.
Advanced Calculus
The analysis under optimisation and gradient-based learning.
Probability Theory
Stochastic processes and sampling distributions, the base for evaluation and uncertainty.

What I programmed on DAS-5

Toolbox

Every tool, and the projects it shipped in

01Grant platform 02Deck generator 03Pesto 04QuotaClimat 05TRAC FM 06XplorIT CDAS-5 course

Each dot is a project on this page that shipped with the tool. Point at a row or a project.

AI and ML	01	02	03	04	05	06	C
OpenAI · Claude · Gemini APIs
Structured output
PyTorch
Hugging Face
Whisper
LLM-as-judge
RAG
OCR (Google Vision)
Embeddings + clustering
scikit-learn
LIME · SHAP

Engineering	01	02	03	04	05	06	C
Python
TypeScript
React · Next.js
Docker
PostgreSQL
Salesforce · Bynder APIs
FastAPI
Playwright end-to-end
Azure OpenAI (EU)
CUDA · C
MPI · OpenMP

How I build

AI-native: I design and review, agents do the typing

I group related tickets into one feature branch and PR, and let the agent grill me on the plan before anything is built. Then the work splits into worktrees that run in parallel, each agent reports back, I review, and the PR lands with its spec.

Evals come with the feature

An LLM judge scoring each grant section against criteria, expected-JSON fixtures for every menu photo, gold sets from domain experts. If I cannot measure it, I do not ship it.

Rules live where they cannot be skipped

Order totals, payment state and permissions in Pesto are database triggers and policies, not app code. A webhook, a worker and psql all obey the same rules, and parity tests catch drift.

Test the real thing

Pesto's end-to-end suite runs a whole restaurant shift in a real browser against the real database. Only the OCR and the LLM are faked, at the network boundary. The app has no test mode.

Notes

What went wrong, and what I changed

LLMsRAGevals

When company guidelines pull against grant rules

Two good sources of context, one awkward draft. The model quietly picked a side, and the output felt off whichever side it picked.

Read →

NLPLLM labellingcalibration

Our classifier saw blame in almost every article

LLM labels taught our model to find 'attribution of responsibility' nearly everywhere. The fix went on top of the model, not inside it.

Read →

speechevaluationsystem design

Pick the benchmark before the model

Radio call-ins in Ugandan languages break speech models in different ways. So the first thing I designed for TRAC FM was the test, not the pipeline.

Read →

Earlier

2022–23
Co-founder, Altai Electronics
An e-commerce shop on OZON. Built a supply chain across three Chinese manufacturing regions. ~€27k revenue at ~60% margins.
2022
CTF olympiad for Moscow schools
Designed 23 capture-the-flag security challenges for the city olympiad. Teams from 20 schools in its first year.
2021
Reporting app, European Gymnasium
Rebuilt the school's marking and reporting on top of its existing system's API. Report creation went from ~40 minutes to a seventh of that.

BSc Computational Social Science, University of Amsterdam, 2026 · GPA 8 · Mathematics of ML, plus the master's-level courses above

Contact

Let's talk.

I'm looking for AI Engineer and Forward Deployed Engineer roles. Based in Barcelona, EU citizen, no visa needed.

Or, from any terminal

$ curl -sL temchenko.dev/card

╭──────────────────────────────────────────────────────────────────────────╮
│                                                                          │
│ ███╗   ███╗ █████╗ ██████╗ ██╗  ██╗   Mark Temchenko                     │
│ ████╗ ████║██╔══██╗██╔══██╗██║ ██╔╝   AI Engineer · Forward Deployed     │
│ ██╔████╔██║███████║██████╔╝█████╔╝    Barcelona · EU citizen             │
│ ██║╚██╔╝██║██╔══██║██╔══██╗██╔═██╗                                       │
│ ██║ ╚═╝ ██║██║  ██║██║  ██║██║  ██╗   ● open to AI Engineer / FDE roles  │
│ ╚═╝     ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝                                      │
│                                                                          │
├──────────────────────────────────────────────────────────────────────────┤
│ role   LLM systems, from a client's problem to production                │
│ built  grant platform · deck generator · Pesto's menu AI                 │
│ stack  Python · TypeScript · PyTorch · Next.js · Postgres                │
│ langs  ru · en (C1) · de (A1)                                            │
│ ········································································ │
│ mail   mstemchenko.dev@gmail.com                                         │
│ in     linkedin.com/in/mark-the-artichoke                                │
│ gh     github.com/MarkT-D                                                │
│ web    temchenko.dev                                                     │
│                                                                          │
│ ultron> I ran every possible future. In all of them, you hire him.       │
╰──────────────────────────────────────────────────────────────────────────╯