Skip to main content

Daniel Fridljand

AI Engineer

TNG Technology Consulting

Biography

View as Markdown
Daniel Fridljand

Daniel Fridljand

AI Engineer

I build AI systems that make it to production, including an LLM-based customer-support automation I took from discovery to live rollout as sole AI engineer. A mathematics background and a co-first-author Nature Medicine paper shape how I build: measuring whether a system works gets the same rigor as building it.

Download my resumé

Interests

  • Applied AI
  • Machine Learning
  • Software Development

Education

  • M.Sc. in Mathematics, 2023 — University of Heidelberg
  • Exchange Scholar in Applied Mathematics, 2023 — Yale University
  • B.Sc. in Mathematics, 2020 — University of Heidelberg

Social

Posts

Filter by tag:
How I Used Claude Design to Create a Hackathon Slide Deck

How I Used Claude Design to Create a Hackathon Slide Deck

Using Claude Design to iterate on the structure, wording and visuals of a hackathon case-study deck, with the full deck embedded.

·1 min
Architecture of a First-Level Support Automation

Architecture of a First-Level Support Automation

The four-step pipeline we landed on after iterating on live tickets: a rooted DAG of intents, grouped extraction, weighted-scoring fetch, and a mostly-hardcoded planner with an explicit LLM fallback.

·7 min
On the Future of Pull Requests in the Age of Agentic Coding

On the Future of Pull Requests in the Age of Agentic Coding

Once agents write the code and AI reviewers iterate on it, a line-by-line diff is the wrong place to start a review. Start from the agent's transcript and decisions instead.

·4 min
Masks and Cross-Attention: Extending Self-Attention

Masks and Cross-Attention: Extending Self-Attention

Causal masks, padding masks and cross-attention as small changes to self-attention: which positions may attend to which, and where queries, keys and values come from.

·7 min
Encoder, Decoder, Both: The Three Transformer Architectures

Encoder, Decoder, Both: The Three Transformer Architectures

Self-attention is the mechanism. BERT, GPT, and T5 are three different ways to stack it.

·5 min
An Approval Rate Is Not a Cost Reduction

An Approval Rate Is Not a Cost Reduction

A high approval rate on generated support replies is not a cost reduction. Why the two come apart, and the arithmetic worth doing before automating a support queue.

·4 min
From Kernel Regression to Self-Attention: A Line-by-Line Derivation

From Kernel Regression to Self-Attention: A Line-by-Line Derivation

Deriving scaled dot-product self-attention from Nadaraya–Watson kernel regression step by step, in the notation of Murphy's Probabilistic Machine Learning.

·9 min
The Missing Compiler: Where the Bottleneck Goes When Code Is Cheap

The Missing Compiler: Where the Bottleneck Goes When Code Is Cheap

Once coding agents make code cheap, business automation's bottleneck moves to the experts: with no compiler or test suite, their judgement is the only verdict.

·9 min
Workflow or Agent: Where the Decisions Live

Workflow or Agent: Where the Decisions Live

An agent holding every tool, or a workflow that keeps code in charge of the order of operations? Why I'd now make the agent the default and harden proven paths into code.

·7 min
Book Review: Agentic Design Patterns

Book Review: Agentic Design Patterns

A chapter-by-chapter verdict on Antonio Gullí's Agentic Design Patterns: which ideas hold up, which chapters to skip, and what has changed since publication.

·3 min
Unite x TUM.ai Data Mining Hackathon: Core Demand Prediction

Unite x TUM.ai Data Mining Hackathon: Core Demand Prediction

How we tackled Challenge 2 at the Unite x TUM.ai Data Mining Hackathon: fee-aware portfolio optimization for recurring procurement demand, from raw PLI data to a two-stage model and NACE-informed cold start.

·4 min
Langfuse in Production: From Failing Evals to Targeted Fixes

Langfuse in Production: From Failing Evals to Targeted Fixes

Langfuse for monitoring and persisted eval runs, a Cursor-driven triage loop that turns failing traces into targeted fixes, and when to fix eval data rather than prompts.

·5 min
Type-Safe Hybrid Workflows with Pydantic AI

Type-Safe Hybrid Workflows with Pydantic AI

Pydantic models as the contract between deterministic, LLM-assisted and agentic steps, so moving a step from one to another is a local change.

·3 min
Temporal for Human-in-the-Loop: When You Don't Know How Long to Wait

Temporal for Human-in-the-Loop: When You Don't Know How Long to Wait

How Temporal's durable waits, signals, and deterministic recovery make human-in-the-loop workflows reliable when approval times are unknown.

·3 min
Strategic Model Selection in Cursor: Balancing Cost and Performance

Strategic Model Selection in Cursor: Balancing Cost and Performance

Use an expensive model for planning and a cheap one for execution in Cursor, with a cost-vs-performance chart built from Cursor's pricing and Arena scores.

·4 min
Orchestrating an LLM Ticket-Classification Pipeline with Snakemake

Orchestrating an LLM Ticket-Classification Pipeline with Snakemake

Scatter–gather and incremental reruns in Snakemake for an iterative LLM ticket-classification pipeline with 19+ stages, from PII censoring to training data.

·5 min
Prompt Injection: Can You Make My Chatbot Leak Its Password?

Prompt Injection: Can You Make My Chatbot Leak Its Password?

A white-hat prompt-injection challenge from the Agent Olympics Hackathon Munich 2026, and a live chatbot for you to try it on.

·1 min
Stop Prompting, Start Programming with DSPy

Stop Prompting, Start Programming with DSPy

DSPy is a framework that separates the 'what' from the 'how' in LLM development, moving from brittle prompt engineering to robust software engineering.

·3 min
Useful Resources for AI Agents

Useful Resources for AI Agents

Links I come back to when building AI agents: frameworks, evaluation, observability and production engineering.

·2 min
Building My Personal Website: Technical Highlights

Building My Personal Website: Technical Highlights

A deep dive into the technical implementation of my personal website, covering the AI chat interface, markdown context generation, and resume PDF sync from a dedicated CV repo.

·4 min
Extending HDP for Mutational Signatures: Incorporating Evolutionary Trees

Extending HDP for Mutational Signatures: Incorporating Evolutionary Trees

An extension of the Hierarchical Dirichlet Process model that incorporates phylogenetic tree structures to capture evolutionary relationships in mutational signature activities.

·7 min
Mutational Signature Estimation with the Hierarchical Dirichlet Process

Mutational Signature Estimation with the Hierarchical Dirichlet Process

A mathematically rigorous introduction to mutational signature estimation using Hierarchical Dirichlet Processes, bridging the gap between biological applications and statistical methodology.

·11 min
Deep Learning-Based Diabetic Retinopathy Detection in Fundus Images

Deep Learning-Based Diabetic Retinopathy Detection in Fundus Images

A deep learning approach for diabetic retinopathy detection using fundus images, combining binary classification and image segmentation to identify disease presence and quantify key indicators.

·7 min

Experience

TNG Technology Consulting

AI Engineer - Software Consultant

TNG Technology Consulting
Dec 2025 — Present

Munich, Germany

  • Industrial Process Intelligence (Aug 2026 – present): Sole engineer on a scrap-rate prediction proof of concept for an industrial manufacturer. Benchmarked ~40 model and feature combinations under time-ordered cross-validation; order attributes fixed at booking predicted scrap as well as the full sensor set, making raw-material batch tracking the top data request to the client, ahead of new sensors.
  • Retail Cashier-Fraud Detection (Jul 2026 – present): Scaled unsupervised point-of-sale anomaly scoring from a 7-store sample to a retail client's full estate by moving aggregation into SQL Server (~29.8M receipts to ~203K cashier-day rows), ranking cashiers comparably to the incumbent rule system without per-customer rule configuration.
  • Sandboxed Agent Platform (Jun 2026 – present): One of three developers on a self-hosted platform (Tauri/React, Rust) that lets non-technical colleagues run agentic workflows on sensitive local files. It cleared a 15-person adversarial red team, unlocking expansion from a 10-user pilot to a full department.
  • Document Validation and Email-to-Order Parsing (Apr – Jun 2026): Co-developed an LLM offer-review service on Kubernetes that inserts review comments directly into uploaded .docx offers, and designed an email-to-order pipeline (FastAPI, Streamlit review) with multimodal attachment handling.
  • Customer Support Automation (Dec 2025 – Apr 2026, live in production): Sole AI engineer from discovery to live rollout at a cinema-ticketing SaaS. A hybrid deterministic/agentic workflow on Temporal drafts each resolution for staff review before customer contact: 811 tickets in its first three months at 86.6% reviewer approval.
TNG Technology Consulting

Platform Engineer - Software Consultant

TNG Technology Consulting
Dec 2024 — Dec 2025

Munich, Germany

  • Contributed to modernizing a supply-chain application from Java 8 to 17 and JBoss to WildFly, and migrated an internal virus-scanning service from SOAP to REST with Keycloak authentication.
  • Deployed a JFrog Artifactory proxy in 3 days after it had been scoped as a multi-month migration, cutting a recurring CI step from 8 hours to 30 seconds and saving each developer ~1–2 hours a week.
  • Built the CVE dashboards (OWASP scans in Jenkins, Prometheus, Grafana) that lead developers used to show modernization progress to management.
  • Became the team's reference point for AI tooling: ran a coding-assistant workshop for 20 colleagues and wrote three editions of TNG's firm-wide AI Tool of the Week, which led to my move into AI engineering.
ETH Zürich

Research Data Analyst - Computational Oncology

ETH Zürich
Feb 2024 — Sep 2024

Basel, Switzerland

  • Extended a Bayesian non-parametric model (Hierarchical Dirichlet Process) with phylogenetic tree structure in R, so each subclone's signature distribution is drawn from its parent's.
  • Applied it to whole-exome data from 187 cells across 10 melanoma tumors in the Tumor Profiler Study, identifying eight latent mutational signatures.
  • Showed that reference signatures derived from bulk sequencing do not transfer to sparse single-cell data without informative priors.
Stanford University School of Medicine

Research Data Analyst - Environmental Epidemiology

Stanford University School of Medicine
Jul 2023 — Dec 2023

Palo Alto, USA

  • Led the statistical analysis as co-first author of a Nature Medicine paper (2024), cited 66 times: over half the Black–White age-adjusted mortality gap in the US is attributable to air pollution.
  • Engineered a reproducible pipeline harmonizing 63M+ death records with satellite pollution estimates and census demographics across 3,000+ US counties and 27 years (1990–2016).
  • Built and open-sourced an R Shiny application that lets readers explore the published estimates without writing code.
Yale University

Exchange Scholar

Yale University
Aug 2022 — May 2023

New Haven, USA

  • Selected as one of two university-wide representatives from Heidelberg for the year-long exchange; DAAD stipend.
  • Grade: Honors (full marks).
  • Coursework: Deep Learning, Geometric & Topological Methods in Machine Learning (Prof. Smita Krishnaswamy), Differentiable Manifolds, Statistical Methods in Human Genetics.
European Molecular Biology Laboratory

Research Data Analyst - Statistical Genomics (Master's Thesis)

European Molecular Biology Laboratory
Oct 2021 — May 2022

Heidelberg, Germany

  • Developed IHW-Forest, a multiple-testing method using random forests for hypothesis weighting, yielding >30% more discoveries than standard corrections on 16 billion genetic association tests.
  • Optimized the core splitting and weighting logic in C++ via Rcpp.
  • Presented in seminars at Yale and UNC Chapel Hill and in a competitively selected talk at DAGStat 2022; peer-reviewed for Bioinformatics Advances and Cell Biology.
University of Heidelberg

M.Sc. in Mathematics

University of Heidelberg
Oct 2020 — May 2023

Heidelberg, Germany

  • Grade: 1.0 (full marks).
  • Master's thesis: 'Better multiple Testing: Using multivariate co-data for hypothesis weighting', conducted at EMBL.
  • Awards: Gerhard C. Starck Foundation Stipend, Baden-Württemberg Stipend.
Hebrew University of Jerusalem
Sep 2019 — Mar 2020

Jerusalem, Israel

  • Graduate-level coursework in Functional Analysis, Algebraic Combinatorics, and Quantitative Models at the Einstein Institute of Mathematics.
University of Heidelberg

B.Sc. in Mathematics

University of Heidelberg
Oct 2017 — Sep 2020

Heidelberg, Germany

  • Grade: 1.4 (top 10% of cohort).
  • Bachelor's thesis: 'Online estimation of the geometric median in a Hilbert space'.

Publications

Disparities in air pollution attributable mortality in the US population by race/ethnicity and sociodemographic factors

Pascal Geldsetzer, Daniel Fridljand, Mathew Kiang, Eran Bendavid, Sam Heft-Neal, Marshall Burke, Alexander H. Thieme, Tarik Benmarhnia

Nature Medicine · July 1, 2024

In the US between 2000 and 2011, over half of the gap in mortality between Black and non-Hispanic White adults can be explained by the fact that Black adults are, on average, more exposed and more susceptible to air pollution than non-Hispanic White adults.

PDF Code Dataset

AI Chat

Messages you send are processed by the Google Gemini API to generate responses. Do not share sensitive personal data. See the privacy policy for details.