# Master’s thesis shortlist — Kristjan Variksoo

**Programme:** University of Tartu, Data Science and Artificial Intelligence  
**Target:** June 2027 defence  
**Purpose:** short attachment for a prospective supervisor

The Institute of Computer Science requires the supervisor and preliminary title for a spring/summer 2027 defence to be registered by **2 November 2026**. The thesis is due **12 May 2027**, with master’s defences scheduled for **1–8 June 2027**.[1] The options below are framed as bounded empirical research with baselines, measurable evaluation and a defensible negative-result fallback, consistent with the ICS thesis guideline.[2]

The shortlist was recalibrated against **133 defended Data Science-curriculum theses from 2022–2026**, selected from the registry by explicit curriculum metadata.[4]

## Recommended primary topic — 97.5/100

### Lead-Time-Aware Calibration of Binary Precipitation Forecasts in Estonia

**Question.** How much does post-hoc probability calibration improve rain/no-rain forecast skill across predeclared lead-time bands when every provider and baseline is evaluated on the identical valid-time cohort?

**Method.** Freeze forecast/observation timestamp semantics and one rain threshold; compare each provider, persistence, rolling climatology and equal-weight ensemble; fit one static logistic/isotonic calibration model and a lead-time-aware variant only if event counts support it; use rolling-origin validation with no valid-time leakage; predeclare one primary proper score and report reliability, false alarms/misses, missingness and provider availability.

**Scope cuts.** Precipitation amount, weather-regime discovery, dynamic location/provider weighting and multiple decision-loss applications are secondary or excluded.

**Existing evidence.** The immutable archive covers **27 locations and 6 providers**. On 19 October 2026 it contained **24,615 capture cycles**, **147,376 source snapshots**, **26,110,710 forecast intervals** and **131,117 observation intervals**. Common-cohort parsers and preliminary skill metrics already exist.

**Minimum defensible result.** A reproducible Estonia-specific benchmark, calibration and error/availability study, even if lead-time-aware calibration does not beat the best provider or static calibration.

**Main limitation.** The archive has less than a full annual cycle. If event counts are sparse, pool locations, remove regime claims and keep conclusions seasonal.

**Relevant ICS areas.** Data Science, Machine Learning and Data Systems.[7]

## Strong backup A — 89.5/100

### An Evaluation Framework for Verified Completion in Tool-Using Agent Sessions

**Question.** Can a fixed event schema and independently audited labels predict whether a tool-using agent session ends in verified completion rather than plausible but unverified output?

**Method.** Freeze one runtime/schema; define obligation, execution, artifact, verification, delivery and correction labels; double-annotate a stratified gold sample; compare current rules and logistic regression, with one richer model only after label audit; use session-level temporal holdout and report agreement, prevalence, AUPRC, calibration and error analysis.

**Scope cuts.** Recovery-policy design, broad fault injection, checkpoint implementation and cross-runtime generalization are optional. The framework and label-validity study are the thesis.

**Existing evidence.** The retained default Hermes profile contains **12,278 sessions and 522,761 message/event rows** from 28 April through 19 October 2026, with thread metadata, tool events, operational artifacts and incident cases.

**Minimum defensible result.** A privacy-safe completion/failure taxonomy, annotation protocol, agreement study and baseline benchmark—even if no recovery intervention is included.

**Main limitation.** Raw log volume is not labelled outcome readiness. If verified failures are too rare, stop at taxonomy, label validity and reliability analysis.

**Relevant ICS areas.** Software Engineering Analytics, Software Engineering & Information Systems, Machine Learning, and Natural & Artificial Intelligence.[7]

## Strong backup B — 88.5/100

### Calibrated Abstention for Receipt-to-Transaction Matching

**Question.** What precision–coverage trade-off can calibrated abstention achieve for receipt-to-transaction matching under temporal and merchant-disjoint evaluation?

**Method.** Freeze extraction, normalization and candidate generation; audit existing positive links and add realistic hard negatives, no-match and ambiguous cases; compare amount/date rules, the current heuristic and text-similarity or logistic pair classification; measure candidate recall, precision at automatic-link coverage, selective risk, recall@k and calibration.

**Scope cuts.** General multimodal understanding, explanation-quality experiments and an LLM reranker remain optional until the labelled set justifies them.

**Existing evidence.** The local ledger contains **11,031 transactions from 325 source files** (2017–2026), **732 source documents**, **2,141 line items** and **184 document–transaction links**, including **172 user-reviewed links**.

**Minimum defensible result.** A labelled benchmark and calibrated rule-versus-compact-ranker comparison showing when the system should abstain.

**Main limitation.** The positive set is small and private. Raw data should remain local; publication should use aggregate results, schemas, synthetic examples or permitted derived features.

**Relevant ICS areas.** Data Systems, Information Systems, Machine Learning and Natural Language Processing.[7]

## Other viable directions

- **Provenance-Preserving Reconstruction of Public-Office Tenures from Estonian Official Records — 87.5/100.** Freeze one source universe and evaluate final tenure reconstruction.
- **Reconstruction Fidelity and Storage Cost of Semantic Preservation for Raspberry Pi OS Images — 86.0/100.** Strong systems benchmark; needs comparative data-driven framing.
- **Source-Aware Time-to-Removal Modelling for Estonian Online Vacancies — 85.5/100.** Clear survival target; requires enough completed lifetimes by the data cut.
- **Event-Anchored Backtesting of Early-Warning Rules in Resolved Prediction Markets — 84.5/100.** Strong public-data study; selection and timestamp semantics are critical.
- **Personalized Versus Pooled Activity Recognition from Smartphone Inertial Data — 83.0/100.** Requires a consented, multi-session participant dataset.
- **Material-Change Classification for Monitored Web-Page Diffs — 82.0/100.** Requires independent gold labels and source-held-out evaluation.

## Suggested first email

**Subject:** Supervision request — precipitation forecast calibration thesis for June 2027

Dear Professor [Surname],

I am a University of Tartu Data Science and Artificial Intelligence master’s student planning to defend in June 2027. I would like to ask whether you might supervise a thesis on **lead-time-aware calibration of binary precipitation forecasts in Estonia**.

I already operate a prospective archive covering 27 locations and six providers. As of 19 October it contained 24,615 capture cycles, 147,376 source snapshots, 26.1 million forecast intervals and 131,117 observation intervals. My proposed research question is whether post-hoc probability calibration, first static and then lead-time-aware if event counts support it, improves rain/no-rain forecasts over each provider, persistence, rolling climatology and an equal-weight ensemble on the identical valid-time cohort.

The minimum defensible result would be a reproducible Estonia-specific forecast benchmark, calibration and error/availability analysis, even if the lead-time-aware model does not beat the best provider or static calibration. I would predeclare one rain threshold and primary proper score, use rolling-origin validation and keep precipitation amount, regime modelling and dynamic provider selection outside the core scope.

Because the ICS deadline for registering a supervisor and preliminary title is 2 November, would you be available for a 20–30 minute discussion this week? I can adapt the precise question and scope to your recommendations.

Best regards,  
Kristjan Variksoo

No professor has been contacted and no topic has been registered.

## Sources

[1] https://cs.ut.ee/en/content/thesis-deadlines-and-guidelines — Thesis deadlines and guidelines | University of Tartu
[2] https://cs.ut.ee/sites/default/files/2026-08/2026_02_03_CSI_Guidelines_for_preparing_and_grading_of_graduation_theses.docx.pdf — Guidelines for preparing and grading graduation theses
[4] https://thesis.cs.ut.ee — UT ICS Graduation Theses Registry
[7] https://cs.ut.ee/en/content/research — UT Institute of Computer Science research groups