At Technische Universität München (TUM) & Ludwig-Maximilians-Universität München (LMU), students within Chair for Database Systems (TUM — HyPer/Umbra Columnar Engines) & Department of Statistics (LMU Munich) master the profound machinery of convex optimisation, high-dimensional covariance estimation, and stochastic analysis—foundational tools that industry too frequently relegates to toy academic exercises. Our remote part-time research pods bridge this divide, enabling you to integrate 10–25 hours per week of production R&D seamlessly between lectures, supervisions, and problem sheets. Rather than treating multi-agent AI as prompt craft, you will implement 100% of your university mathematics directly into live production code, treating ensembles of large language models as assets in a Markowitz mean-variance portfolio. Here, correlation is the new headcount: your mathematical derivations determine the exact capital and compute allocation across verifiers, deterministic tools, and neural reasoning agents.
Dedicate 10–25 hours per week on a genuinely flexible, asynchronous schedule engineered to harmonise with your lectures, lab practicals, supervisions, and weekly problem sheets. You determine your working hours without mandatory synchronous standups, ensuring your academic degree remains your absolute priority.
Earn $20–$45/hr USD in fully remote research stipends alongside our strict Exam-Safe Protocol: pause all commitments completely (0 hours/week) during university revision periods, mid-term assessments, and final Tripos or degree examinations, with your pod seat and research project unconditionally guaranteed upon your return.
Translate production research directly into your BSc/MSc dissertation, MMath/MPhil thesis, or seminar project. Co-author citable 2024–2026 working papers and gain unlimited access to our 39-hour Continuing Professional Education (CPE) curriculum and preprint repository at {preprints_url}.
| University Module / Tripos Paper | Mathematical Formulation | Research Pod | Production Architecture & Application |
|---|---|---|---|
| Convex Optimisation & Mathematical Programming | min_w w^T Σ w s.t. w^T μ ≥ μ*, 1^T w = 1, w ≥ 0 |
POD-01 | Formulating the Pareto frontier for compute allocation across heterogeneous LLMs and deterministic verifiers via Karush-Kuhn-Tucker (KKT) conditions and interior-point solvers. |
| Probability Theory & Correlated Random Sequences | N_eff = N / (1 + (N - 1)ρ̄) |
POD-01 | Quantifying ensemble redundancy: demonstrating why 50 correlated agents with average cross-correlation ρ̄ = 0.70 provide only N_eff ≈ 1.42 independent perspectives, and constructing orthogonal prompting topologies to maximise information entropy. |
| High-Dimensional Mathematical Statistics & Econometrics | Σ̂_LW = (1 - δ*)S + δ*F |
POD-02 | Resolving empirical covariance singularity when evaluating agents over short benchmark horizons (N > T) using Ledoit-Wolf optimal shrinkage and Marčenko-Pastur spectral clipping against random matrix noise. |
| Actuarial Risk & Extreme Value Theory | Σ⁻_ij = E[min(R_i-τ,0)min(R_j-τ,0)], CVaR_99%(L_w) ≤ M_audit |
POD-02 | Isolating downside epistemic failure via Markowitz (1959) target semivariance and enforcing Rockafellar-Uryasev Conditional Value-at-Risk bounds against catastrophic multi-agent cascade vulnerabilities. |
| Financial Economics & Stochastic Calculus | w_tan ∝ Σ⁻¹(μ - c·1), SR(w) = w^T(μ - c) / √(w^T Σ w) |
POD-03 | Extending Tobin's separation theorem: treating formal verification tools and deterministic compilers as risk-free baseline assets (σ_0 = 0) combined with the tangency portfolio of generative LLMs to maximise the operational Sharpe ratio. |
| Sequential Hypothesis Testing & Asymptotic Statistics | P(Type I Error) ≤ α (Neyman-Pearson FrugalGPT/RouteLLM Cascade) |
POD-04 | Designing sequential probability ratio tests and Neyman-Pearson cascades to route queries dynamically between lightweight specialist models and frontier frontier models, drawing on our 39-hour open CPE research archive at {preprints_url}. |
1. JobTeaser DE & Fachschaft Mathematik/Informatik Stellenbörse: Search for 'Markowitz Systems R&D Fellowship' on JobTeaser DE & Fachschaft Mathematik/Informatik Stellenbörse or submit your CV and transcript directly through your Technische Universität München (TUM) & Ludwig-Maximilians-Universität München (LMU) institutional email address for expedited review by our fellowship admissions committee. (RSS/XML Job Feed · iCal Open Office Hours · Google Classroom)
2. Python Seminar & Thesis Kit (/uni/de-tu-munich-lmu-munich/lab.py): A reproducible 5-minute Python seminar lab script for student workshops that computes effective degrees of freedom (N_eff) across correlated prompt outputs, implements Ledoit-Wolf analytical covariance shrinkage (Σ̂_LW), and visualises the Markowitz efficient frontier of multi-agent topologies.