# Candidate master’s-thesis directions for supervisor discussion

**Student:** Kristjan Variksoo  
**Programme:** University of Tartu, Data Science and Artificial Intelligence  
**Target:** thesis submission in May 2027 and defence in June 2027

## Time-critical process fact

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 published thesis submission date is **12 May 2027**, and the master’s defence window is **1–8 June 2027**.[1]

The current ICS guideline treats the master’s thesis as research: it should produce new results, justify methods and technological choices, compare with prior solutions, and evaluate how well the result works using measurable evidence.[2] The directions below were recalibrated against **133 defended Data Science-curriculum theses from 2022–2026** selected by the registry’s explicit curriculum field, not by title keywords.[4]

The supervision-fit labels use the Institute of Computer Science’s official research-group names.[7]

## Recommended hierarchy

1. **Primary — Lead-Time-Aware Calibration of Binary Precipitation Forecasts in Estonia (97.5/100).** Best combination of ready data, clean baselines, leakage-safe temporal evaluation and useful negative-result fallback.
2. **Backup — An Evaluation Framework for Verified Completion in Tool-Using Agent Sessions (89.5/100).** Timely AI topic with rich logs; success depends on a credible gold-label audit.
3. **Backup-of-backup — Calibrated Abstention for Receipt-to-Transaction Matching (88.5/100).** Concrete and measurable; keep multimodal/LLM work optional until the labelled set grows.
4. **Strong alternative — Provenance-Preserving Reconstruction of Public-Office Tenures from Estonian Official Records (87.5/100).** Excellent validated source assets; must freeze one source universe and one primary pipeline outcome.
5. **Strong systems alternative — Reconstruction Fidelity and Storage Cost of Semantic Preservation for Raspberry Pi OS Images (86.0/100).** Exceptional artifact and benchmark readiness; needs a systems/data supervisor comfortable with the programme-fit argument.
6. **Strong alternative — Source-Aware Time-to-Removal Modelling for Estonian Online Vacancies (85.5/100).** Clear survival-analysis question with a descriptive fallback; dependent on enough completed lifetimes by the data cut.

The first email should contain at most **two related directions**, not all six. No professor has been contacted and no topic has been registered.

---

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

**One-sentence pitch.** Determine whether a simple lead-time-aware calibration model improves rain/no-rain probabilities over each provider, persistence, climatology and an equal-weight ensemble on the same Estonian valid-time cohort.

**Primary research question.** How much does post-hoc probability calibration improve binary precipitation forecast skill across predeclared lead-time bands when every method is evaluated on the identical valid-time cohort?

**Scope cuts.** Precipitation amount, weather-regime discovery, location-specific dynamic provider selection and multiple decision-loss applications are secondary or excluded. They must not become co-equal contributions.

**Methods and baselines**

- Freeze issue-time, valid-time, observation and rain-threshold semantics before fitting.
- Baselines: each provider, persistence, rolling/seasonal climatology and equal-weight ensemble.
- Candidate methods: one static logistic or isotonic calibration model, followed by a lead-time-aware variant if event counts support it.
- Rolling-origin temporal validation; forecasts for one valid time may not cross train/test boundaries.
- Primary metric: one proper score such as Brier score. Secondary analysis: reliability, false alarms/misses and provider availability.
- Predeclare one primary rain threshold and a frozen cohort/data-quality ledger.

**Existing assets, observed 19 October 2026**

- An immutable archive spanning **27 points**, **6 forecast providers** and multiple observation sources.
- **24,615 capture cycles**, of which **24,298 were complete** and **317 partial**.
- **147,376 source snapshots**, **26,110,710 forecast intervals**, and **131,117 observation intervals**.
- Existing common-cohort parsers and preliminary skill calculations.

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

**Closest defended precedents.** *Short-Term Harbor Water Flow Forecasting Using Meteorological and In-situ Measurements* (2026), *In-Depth Analysis of Miscalibration in Binary Classification* (2025), and *Hierarchical Forecasting Methods in Day-Ahead Electricity Consumption Forecasting* (2024).[4]

**Main boundary.** If the archive lacks enough events per lead-time stratum, pool locations, remove regime claims and state a seasonal rather than year-round inference boundary.

**Likely supervision fit.** Data Science, Machine Learning, Data Systems; possible meteorological/environmental-data co-supervision.

---

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

**One-sentence pitch.** Define and validate observable labels for requested obligations, artifact creation, verification and delivery in tool-using agent sessions, then compare simple predictors of verified completion.

**Primary research 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?

**Scope cuts.** Recovery-policy design, checkpoint implementation, broad fault injection, cost optimization and cross-runtime generalization are optional after the benchmark exists. The thesis is the evaluation framework and label-validity study.

**Methods and baselines**

- Freeze one runtime/event-schema version and deduplicate retries, health checks and nested agent runs.
- Operationally define requested obligation, tool execution, artifact creation, verification, delivery and later correction.
- Double-annotate a stratified gold sample; report agreement, adjudication and prevalence.
- Baselines: current rule checks and logistic regression on aggregate event features; add one tree or time-to-event model only after label audit.
- Session-level temporal holdout; primary metric AUPRC for incomplete obligations, plus calibration and error analysis.
- Privacy/publication review for session text, tool output, user identifiers and third-party data.

**Existing assets**

- **12,278 retained sessions** and **522,761 message/event rows** from 28 April through 19 October 2026, with platform/thread metadata and operational artifacts.
- Implemented reliability mechanisms and incident cases suitable for defining candidate labels and controlled examples.

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

**Closest defended precedents.** *Designing an Evaluation Framework for a General AI Tutor in K-12 Education* (2026), *Detecting and Mitigating Data Drift in Call Center Data* (2025), and *Exploring SQL-based Near-realtime Conformance Checking Using Stream Processing Engines* (2024).[4]

**Main boundary.** Raw log volume does not equal labelled outcome readiness. If verified failures are too rare, stop at the framework, taxonomy and reliability analysis rather than forcing a predictive model.

**Likely supervision fit.** Software Engineering Analytics, Software Engineering & Information Systems, Machine Learning, Natural & Artificial Intelligence.

---

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

**One-sentence pitch.** Compare amount/date rules, the current heuristic and a compact learned matcher, then calibrate abstention so only high-confidence receipt–transaction links are automated.

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

**Scope cuts.** General multimodal document understanding, explanation-quality experiments and an LLM reranker are optional. The core is selective matching, not a universal finance assistant.

**Methods and baselines**

- Freeze transaction normalization, document extraction and candidate generation before labelling.
- Audit existing positive links and construct realistic hard negatives, no-match cases and ambiguous cases.
- Baselines: amount/date tolerance rules, current weighted heuristic and text-similarity retrieval or logistic pair classification.
- Measure candidate-generation recall before ranking quality.
- Temporal holdout plus a merchant-disjoint sensitivity analysis.
- Primary outcomes: precision at automatic-link coverage and selective risk; secondary metrics recall@k and calibration.

**Existing assets, observed 19 October 2026**

- **11,031 transactions** from **325 source files**, covering 1 January 2017 through 17 October 2026.
- **732 source documents**, **2,141 line items**, and **184 document–transaction links**, including **172 user-reviewed links**.
- Existing provenance tables, parsers, link explanations and privacy-aware reports.

**Minimum defensible result.** A labelled benchmark and calibrated rules-versus-compact-ranker comparison showing when the system should abstain, even if a multimodal or language model is unjustified.

**Closest defended precedents.** *Repair Decisions for Used Smartphones Using Machine Learning Methods* (2025), *Machine Learning Methods in Anti-Money Laundering* (2025), and *From Clinical Trial Free Text to Structured Disease Facts: A Grounded LLM Validation Framework* (2026).[4]

**Main boundary.** The positive set is small and private. Expand/audit labels first, retain raw data locally and publish only schemas, aggregate results, synthetic examples or permitted derived features.

**Likely supervision fit.** Data Systems, Information Systems, Machine Learning, Natural Language Processing.

---

## 4. Provenance-Preserving Reconstruction of Public-Office Tenures from Estonian Official Records

**One-sentence pitch.** Evaluate whether typed legal/administrative events plus explicit provenance reconstruct person–office tenures more accurately and honestly than flat person–role rows.

**Primary research question.** How accurately can a bounded set of Estonian official source families be transformed into person–office tenure intervals while preserving event semantics, provenance, conflicts and unknown dates?

**Scope cuts.** Baltic-wide completeness, unrestricted source discovery, entity resolution, event extraction, temporal reasoning and contradiction detection cannot all be co-equal theses. Freeze one Estonian source universe and make final tenure reconstruction the primary outcome.

**Methods and baselines**

- Freeze the computational definition of in-scope offices, source families and collection date.
- Create a manually adjudicated set of entities, event types, dates, intervals, conflicts and unknowns.
- Baselines: current flat rule extraction, untyped person–role rows and a deterministic source-priority/date rule.
- Evaluate extraction, entity resolution and final interval reconstruction separately; include source-stratified errors and abstention/provenance completeness.
- Treat the graph representation as the mechanism, not a second independent contribution.

**Existing assets**

- A normalized official-source dataset with **1,267 tenure rows**, **798 distinct people**, **777 current**, **485 former** and **5 announced-future** records.
- Validators reported zero structural/date/status errors and preserve explicit gaps and source conflicts rather than inventing dates.

**Minimum defensible result.** A bounded temporal/provenance schema, adjudicated gold subset and comparative tenure-reconstruction benchmark.

**Closest defended precedents.** *Evaluation Framework for Schema-Constrained Knowledge Graph Extraction Using Large Language Models* (2026), *Framework for Neural Network Based Fact Extraction Workflows* (2022), and *Determining Estonian Usual Residents Using Machine Learning Methods* (2023).[4]

**Main boundary.** Keep legal interpretation and source assertions separate from computational ground truth; avoid claiming complete national truth from the selected source set.

**Likely supervision fit.** Data Systems, Natural Language Processing, Information Systems, applied data science.

---

## 5. Reconstruction Fidelity and Storage Cost of Semantic Preservation for Raspberry Pi OS Images

**One-sentence pitch.** Compare full-image, used-block, file-level and semantic-preservation methods on a fixed Raspberry Pi OS image suite using reconstruction fidelity, storage, read volume, runtime and failure analysis.

**Primary research question.** Under what controlled modifications does semantic preservation reduce storage cost while retaining byte-level or declared semantic reconstruction fidelity compared with established image/backup baselines?

**Scope cuts.** Do not combine semantic backup, deleted-data forensics, universal Linux support, secret management, automatic recipe synthesis and byte-identical reconstruction into one claim. Freeze one OS family, image suite and threat model.

**Methods and baselines**

- Baselines: compressed full image, filesystem-aware used-block image, file-level archive/rsync-style backup and the current semantic package.
- Use several versioned Raspberry Pi OS images with controlled modifications plus at least one held-out real card image.
- Primary outcomes: reconstruction hash/semantic manifest success and storage size; secondary outcomes: bytes read, wall time and failure modes.
- Ablate public-package references versus private deltas.
- Use boot or provisioning checks when the claim is semantic rather than byte-identical.

**Existing assets**

- Working semantic-dehydration and reconstruction tooling, exact manifests and hash checks.
- Controlled image experiments, package-aware assets and a clear offline/rehydration evaluation path.

**Minimum defensible result.** A reproducible benchmark that identifies where semantic preservation saves space and where it fails—even if a conventional used-block image remains preferable overall.

**Closest defended precedents.** *A Comparative Study of Database Schema on Query Performance in Data Warehousing* (2024), *Design and Implementation of an Incremental ELT Pipeline for a Jira Data Warehouse* (2023), and *Comparative Evaluation and Implementation of an MLOps Platform* (2026).[4]

**Main boundary.** Programme fit depends on a comparative, data-driven research question and careful operational metrics rather than the breadth of the tool.

**Likely supervision fit.** Data Systems, distributed systems, software engineering, systems-oriented applied data science.

---

## 6. Source-Aware Time-to-Removal Modelling for Estonian Online Vacancies

**One-sentence pitch.** Define source-specific vacancy lifecycle events and test whether survival models improve calibrated time-to-removal estimates over source-stratified empirical baselines.

**Primary research question.** How well can vacancy time-to-removal be modelled when disappearance, explicit deadlines, restoration and cross-source syndication are kept distinct?

**Scope cuts.** Personalized relevance ranking is excluded from the core thesis. The outcome is removal/lifecycle time, not “job quality” or causal employer behavior.

**Methods and baselines**

- Freeze first-seen, last-seen, removal, restoration and explicit-deadline semantics by source.
- Deduplicate syndicated postings while preserving source-level observations.
- Baselines: source-stratified empirical survival curves and Cox proportional hazards.
- Candidate: one accelerated-failure-time or tree-based survival model if event counts support it.
- Rolling temporal evaluation; concordance, integrated Brier score, horizon calibration and source-specific error analysis.

**Existing assets, observed 19 October 2026**

- **9,171 postings**, **7,846 canonical jobs**, **8,571 active** and **600 removed** records.
- **13 configured sources** (12 active, 1 blocked), **46,361 raw captures**, **24,857 snapshots**, **25,860 lifecycle events** and **58,932 classification records**.

**Minimum defensible result.** A validated lifecycle semantics/provenance dataset and descriptive source-aware survival benchmark, even if a complex model does not beat source-stratified curves.

**Closest defended precedents.** *Analysis of Psoriasis Comorbidities Using Cox Regression* (2024), *A Machine Learning Approach for Predicting Operation Times from Production Feedback Data* (2026), and *Utilising Machine Learning and RFM Analysis for Customer Retention in an Online Grocery Delivery Startup* (2023).[4]

**Main boundary.** The archive began on 17 October 2026. Keep collection running, freeze a future cut and confirm enough completed lifetimes before committing to a complex model.

**Likely supervision fit.** Data Systems, Software Engineering Analytics, Natural Language Processing, applied Machine Learning.

---

## Specialized alternatives

- **Event-Anchored Backtesting of Early-Warning Rules in Resolved Prediction Markets — 84.5/100.** Strong public-data temporal study; event anchors and selection bias must be frozen before modelling.
- **Personalized Versus Pooled Activity Recognition from Smartphone Inertial Data — 83.0/100.** Strong acquisition/model stack; requires a consented multi-session participant dataset.
- **Material-Change Classification for Monitored Web-Page Diffs — 82.0/100.** Good dataset-construction plus NLP route once independent gold labels exist.
- **Reliability of Time-Shared Low-Cost SDR Spectrum-Occupancy Detection — 80.5/100.** Interesting sensor/time-series topic; controlled ground truth is the blocker.
- **Stability of Recurring Software-Procurement Themes in TED Notices — 77.0/100.** Official multilingual corpus opportunity; freeze deduplication and validation semantics.
- **Software-Test Coverage of a QEMU-Based Raspberry Pi Environment — 74.5/100.** Reproducible systems comparison; less central to DS/AI.

## Which options to show which group

- **Data Science / Machine Learning:** precipitation calibration + vacancy survival or prediction-market backtesting.
- **Software Engineering Analytics / Natural & Artificial Intelligence:** agent completion evaluation + vacancy lifecycle.
- **Data Systems / NLP:** receipt matching + public-office tenure reconstruction.
- **Systems / Data Systems:** semantic preservation + QEMU test coverage.

## Outreach email draft

**Subject:** Master’s thesis supervision request — existing dataset and bounded research directions 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 either of these directions could fit your group:

1. **[Primary title]** — [one-sentence research question]. I already have [one sentence naming the verified dataset/prototype].
2. **[Related backup title]** — [one-sentence research question]. I already have [one sentence naming the verified dataset/prototype].

I want the thesis contribution to be a bounded empirical comparison with explicit baselines, a frozen data cut, leakage-safe evaluation, failure analysis and a minimum defensible result—not only a software implementation. I have prepared a short note with existing assets, comparable recent UT theses, scope cuts and data/evaluation prerequisites.

Because the ICS deadline for registering the supervisor and preliminary title is 2 November, would you be available for a 20–30 minute discussion this week? I am happy to narrow the scope to the direction you consider most academically defensible.

Best regards,  
Kristjan Variksoo

## Ready-to-paste primary version

**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 the 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, prevent valid-time leakage 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 send the short topic note and adapt the scope to your recommendations.

Best regards,  
Kristjan Variksoo

## Questions for the supervisor meeting

1. Is the primary question sufficiently novel and narrow?
2. Which simple baseline and one primary metric would you require?
3. Is the independent-unit count and label quality sufficient for the proposed split?
4. What should the minimum defensible result be if the candidate model fails?
5. Which attractive components should be explicitly excluded?
6. Are ethics, privacy, IP, data-rights or publication approvals needed?
7. What data cut, supervision cadence and interim milestones do you expect?

## Suggested schedule

- **By 21–23 October:** send one targeted email with a primary and one related backup.
- **By 27 October:** hold supervisor discussions; fix one question, source universe, primary metric and minimum result.
- **By 30 October:** agree on supervisor and preliminary title, retaining a two-day safety margin.
- **November–December:** literature review, protocol/data freeze, label audit and simple baseline.
- **January–March:** main experiments and error analysis.
- **April:** writing, robustness checks and supervisor revisions.
- **12 May 2027:** official submission deadline.[1]

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