Data Sovereignty & Responsible AI

The Projects

Memory Receipt
Memory With a Receipt is an AI assistant for finding reliable project decisions across emails, reports, and meeting transcripts. Every answer links to a literal quote, original source file, and exact line range. It distinguishes proposals, commitments, reversals, and current status instead of treating every statement as a decision. The system retrieves locally, limits external model use, masks direct contact details and credentials, and tracks token usage. It also supports derived-data deletion and previews the impact on remaining decision evidence.
Submitted:
20 Sept 2026
Learn More about Memory Receipt (opens in a new tab)
Scand AI
Getting real and reliable data from project files is difficult. What was really said? Was was taken back later? What was only proposed? This is where Scand AI comes in. With a strong focus on sovereign infrastructure, Scand AI organizes information and give you exactly what you need. Ask any questions about the data set, and Scand AI accesses a database of extracted statements and their current status. Thus, you get real, current and reliable answers from all of your sources. Our unique selling points: - Sovereignty - We host everything on one machine, including an open-weights LLM, in Finland - Portable - By shipping our own self-contained application, including the LLM inference, you can easily spin up a full instance in your own office, no data leaving your premise - Accuracy - We are using the power of the LLM instead of relying on flaky RAG performance and possibly losing context - Scalability - By extracting only relevant statements from the documents, we reduce the context size enormously to enable a larger statement context and thus more accurate answers
Submitted:
20 Sept 2026
Learn More about Scand AI (opens in a new tab)
PS2T (personalized speech-to-text
An end-to-end working application of speech-to-text transcription with a backend to allow for personalised, fine-tuned speech-to-text model generation.
Submitted:
20 Sept 2026
Learn More about PS2T (personalized speech-to-text (opens in a new tab)
TPM - Trustworthy Process Monitor
Norrin Trustworthy Process Monitor takes an undocumented table of sensor readings - any format, no labels, no documentation, gigabytes if needed - and explains in plain language what each column is, whether the data can be trusted, when the process behaved unusually, and why. It runs the entire analysis on your own computer: the default profile makes no network call at all, and the optional Claude and EU-hosted routes send only aggregated, rounded and anonymised results through a guard that inspects every payload and records it in a ledger. Raw readings never leave the machine. Seven stages build on each other: understanding the columns and how sensors move together, data quality, monitoring for drift and sudden changes, a ranked diagnosis challenged by a built-in critique, an assessor that says whether more data would help, reports in HTML, PDF, PowerPoint and e-mail, and a live monitor for a running feed. Every page opens with a short summary and next steps, with charts, tables and evidence behind "Show technical analyses", in Basic, Operator and Engineer modes and in English, Finnish and Swedish. A person can accept, question or override any conclusion, rename a sensor, write operating rules in plain English and ask "why?" in a chat that counts from the run's own artifacts, and every decision lands in a tamper-evident log. The app is built to be checked rather than believed: where a list stops at a limit it says so and gives the true total, a finding its own critique weakened can never be called "very sure", and every score states what it measures. When the data happens to carry a truth column, it scores its own detection - precision, recall, detection delay - and says plainly what those numbers mean. It installs on Windows in minutes with no Python, no administrator rights, no account and no API key; with a local model (Ollama) the explanations are written on your machine, and without one it says so and keeps working.
Team:
aBARofchocolate
Submitted:
20 Sept 2026
Learn More about TPM - Trustworthy Process Monitor (opens in a new tab)
TraceGuard
TraceGuard — Privacy-Preserving Analytics Pipeline for Mobile Network Telemetry TraceGuard is a data-sovereignty engineering project that turns raw mobile-network event-window telemetry (per-subscriber/device throughput, latency, data-volume, and location metadata) into analytics-ready datasets that minimise re-identification risk while preserving statistical utility. The dataset combines direct identifiers (MSISDN, IMSI, IMEI) with fine-grained spatiotemporal fields (enb, time_start, province) and behavioural signals (app-category usage, video/audio volume, tethering). Individually these look like standard network KPIs; combined, they form a movement and behaviour trace — de Montjoye et al. (2013) showed just four spatiotemporal points uniquely identify ~95% of individuals in comparable mobile datasets. TraceGuard treats that as the central design constraint, not an afterthought. What it does: Strips and re-keys direct identifiers — MSISDN/IMSI/IMEI are dropped or replaced with keyed HMAC-SHA-256 pseudonyms (key held outside the repo/dataset), never plain hashes. Generalises quasi-identifiers — enb is clustered to coarser cell-areas, time_start rounded to hour/day, rare cell/time combinations suppressed. Enforces measurable k-anonymity — rather than assuming anonymisation worked, the pipeline computes and reports uniqueness/k-anonymity on its own output (docs/anonymisation-report.md), following ISO/IEC 20889 and NIST SP 800-188 methodology. Aggregates behavioural fields (data volume, app category, media usage) into bands to reduce profiling/inference risk. Documents every privacy design decision against its legal basis — GDPR Art. 5/25/32, ePrivacy Arts. 6 & 9, WP216's three anonymisation tests (singling out, linkability, inference) — so outputs can be defended to a DPO, not just an engineer. Why it matters for a Data Sovereignty hackathon: the project demonstrates that "anonymisation" is a verifiable engineering property, not a label — pseudonymisation ≠ anonymisation (GDPR Art. 4(5), EDPB Guidelines 01/2025), and claims of anonymity must be tested, not assumed. It also models safe agent/AI-assisted workflows around sensitive telecom data (strict raw-data handling rules, synthetic-data-only testing, no data egress to third-party tools), which is directly relevant to any hackathon environment where external participants or AI tooling touch real personal data. Tech scope: Python-based pipeline (src/) run only by humans in a controlled environment; unit tests use synthetic/Faker-style fixtures only; raw CSVs are gitignored and never committed or fed to external services.
Team:
Team Fih
Submitted:
20 Sept 2026
Learn More about TraceGuard (opens in a new tab)
trace.
A lightweight neural network that keeps an eye out for user activities in a dashboard and tries to detect fraud, scams, stolen credentials devices, and so forth.
Submitted:
20 Sept 2026
Learn More about trace. (opens in a new tab)
confidentiality-broker
AI harness which accounts for data and provider confidentiality.
Team:
a-studio enjoyers
Submitted:
20 Sept 2026
Learn More about confidentiality-broker (opens in a new tab)
RecallX
A cited-answer customer memory archive built over a SQLite corpus, featuring a Python standard-library backend and a React single-page application frontend. The system answers customer and project questions by binding every generated claim directly to exact, stored source line offsets, ensuring provable provenance and non-hallucinated citations.
Team:
RecallX2
Submitted:
20 Sept 2026
Learn More about RecallX (opens in a new tab)
Datalight
Datalight helps teams understand their data, detect meaningful changes, and investigate them with traceable evidence—while keeping raw data local. Upload a numeric CSV to explore data quality, statistics, correlations, and short-term forecasts. Replay observations in batches to monitor abrupt changes, sustained drift, and deviations from an initial reference. Interactive charts connect each flagged interval to the measurements and rules behind it, with an explicit assessment of evidence strength. An optional AI assistant explains findings, answers follow-up questions with evidence citations, and translates natural-language requests into simple monitoring rules for users to review and apply. Only computed summaries and sanitized requests with anonymous channel identifiers reach the model; raw observations and original column names stay local. Monitoring continues even when AI is unavailable. Human judgment stays central: users can accept, question, or override decisions with a reason, while preserving the original findings and review history. Statistical changes are treated as signals for investigation, not proven diagnoses. Our demo uses industrial process data from the Tennessee Eastman benchmark and a synthetic web-service dataset to demonstrate the same monitoring workflow across domains.
Team:
Datalight
Submitted:
20 Sept 2026
Learn More about Datalight (opens in a new tab)
Relio
An AI teammate for RELEX customer teams that turns scattered conversations into shared, traceable project memory. The idea is simple: include the agent in a project's email conversation, and relevant knowledge becomes available to the team. It connects decisions, commitments, open questions and stakeholder information across conversations, helping people understand what happened, what changed and what needs attention. Key features: - Personal morning briefs tailored to each employee's role, priorities and upcoming meetings. - Natural-language questions with answers linked to exact source passages. - Proactive email participation when the agent can contribute useful, evidence-backed information. - Stakeholder guidance grounded in recorded interactions and project history. - An explorable knowledge graph connecting people, sources and project facts. - Access controls and traceable deletion, separating customer-shareable, internal and restricted information while keeping personal memory private. The demo uses RELEX's Acme archive and an Outlook-style email simulator to demonstrate how new conversations update project memory, calendar entries and briefings. It also demonstrates conflicting evidence, changing facts and deletion propagating through stored and derived information. Inference runs on GLM 5.3 Flash hosted by Verda, formerly DataCrunch, in Helsinki.
Submitted:
20 Sept 2026
Learn More about Relio (opens in a new tab)
Recourse
HR Automated Advisor is an open-source platform that helps HR managers handle the flood of applications that AI tools now make possible. Instead of reading hundreds of CVs by hand, the manager defines what a role needs, and the system does the first pass. For each vacancy, the manager sets hard requirements (must-haves that decide whether a candidate is considered at all) and soft requirements (nice-to-haves with adjustable weights). Uploaded CVs are parsed into structured data, scored against these criteria, and shown as a ranked list. Transparency is at the core. For every candidate, the platform explains its reasoning: which requirements were matched, where the gaps are, and how the score was calculated, with the relevant parts of the CV highlighted. Protected attributes are never used in screening. The platform also drafts personalised feedback emails so every applicant gets a clear, respectful answer. Nothing is sent or decided automatically: the manager reviews and approves every outcome. Built with EU regulations in mind (GDPR, the EU AI Act, NIS2), HR Automated Advisor keeps humans in control while saving them hours of repetitive work. Login credentials to the projects website: demo/demo
Team:
Yabloko
Submitted:
20 Sept 2026
Learn More about Recourse (opens in a new tab)
Trim & Jitter V3
Trim & Jitter V3 makes telecom data useful for product development and service-quality analysis through anonymization. It helps analysts explore gaming and streaming opportunities and prioritize application-quality improvements. Our approach removes identifying fields, suppresses rare combinations and perturbs numeric measurements while preserving useful usage and quality patterns. We compare analytical results and measure singling out, linkage and inference before and after transformation. Presented by the Pivot Pilots: Mikael and Patrik.
Team:
Pivot Pilots
Submitted:
20 Sept 2026
Learn More about Trim & Jitter V3 (opens in a new tab)
Anh Phan
A. A general-purpose statistical analysis pipeline for numeric, multi-run tabular data. TPM takes raw datasets, builds a statistical baseline, detects changes in new data, and uses an LLM to help explain what the analysis may mean. B. Key Features 1. Automatic dataset discovery — detects delimiters, timestamps, labels, run IDs, sample IDs, and numeric columns. 2. Statistical profiling — uses ydata-profiling for per-run statistics, correlations, time-series diagnostics, and quality alerts. 3. Causal discovery — uses Tigramite PCMCI with partial correlation tests. 4. Automatic clustering — groups related variables using hierarchical clustering and silhouette-based cluster selection. 5. Multi-perspective anomaly detection — checks mean shifts, variance shifts, correlation changes, and fault onset. 6. Autocorrelation-aware statistics — adjusts effective sample size for sequential data. 7. CUSUM fault detection — identifies when changes begin and provides a possible propagation order. 8. LLM-assisted interpretation — explains variables, cluster roles, and possible root causes from statistical evidence. 9. Data-quality rules — supports built-in checks and human-confirmed custom rules. 10. Audit logging — records LLM calls and important system decisions. 11. Reusable reports and plots — exports structured JSON/CSV data plus a readable report.md. 12. Web dashboard — inspect results and upload new runs through a FastAPI/Jinja2 interface. 13. Dataset-agnostic design — not tied to the original Tennessee Eastman dataset.
Submitted:
20 Sept 2026
Learn More about Anh Phan (opens in a new tab)