Cases from our data projects. We build them on open data, the client's own data and machine-learning models. Each case names the industry, the problem, the data that was used and the result.
Problem: The bank had to assess a borrower before issuing a loan with more signals than the credit bureau alone provides.
Data: The bank's own data plus open sources: public registries and databases (EGRUL, FSSP, NBKI), Rosstat statistics, social networks, job boards (HH, Superjob) and classifieds (CIAN, Avito, Avto.ru).
What was done: We trained a behaviour model on two sources at once — the bank's data and open data — and used it as an additional pre-loan signal. The model was never meant to decide on its own: the scoring itself stayed with the bank.
Outcome: The model reports the borrower's behavioural profile from aggregated data and helps the bank judge credit risk. Of 146 million residents of Russia, 137 million had open information about them, 98 million were in social networks and 42 million in public databases.
Problem: Raise loyalty to the car brand and sell cars. The brand's audience in the project was 37 million people.
Data: Published content, likes in thematic groups and posts, audience activity and the speed at which content spread across the social network.
What was done: We filtered the audience by interests, activity and content spread, measured the share of bots and cynical followers, processed the top 150 accounts by reach and selected those whose followers are on target and whose content travels fastest.
Outcome: Seven niche auto bloggers were selected to work with the brand and a new content plan was built. Result: +41% customers from the blogger channel and +6% from friend recommendations.
Problem: Cut advertising costs, attract new customers and make targeted advertising more effective: know who buys and what people buy.
Data: Loyalty cards, customer data collected from social networks, and the matching of purchases against the buyer's social portrait.
What was done: A loyalty card individual for every customer and activated through a social network; we built the audience portrait — who buys, what content they consume, what they are interested in and what they buy — then found similar users and targeted the advertising at them.
Outcome: The project covers the full loop of scoring, profiling and precise targeting with consumer-behaviour forecasting; client results are not published.
Problem: Predict a customer's needs from their digital footprint and use it for accurate segmentation, micro-targeting and reactivation.
Data: Social networks, sites, blogs and forums, dating apps and games, public registries (EGRN, EGRUL, FSSP), job boards (HH, Superjob) and classifieds (CIAN, Avito, Avto.ru).
What was done: We built remote identification, precise segmentation, micro-targeting and forecasting; the reactivation case relied on personal interaction and the analysis of the user's own experience rather than on mass mailing.
Outcome: Reactivation cut communication costs by 20% and returned 12% of "old" customers; in a neighbouring case the promo-code channel added 41% of customers and friend recommendations 6%.
Problem: Customise creatives to what a particular audience prefers, and then decide which user sees which creative.
Data: Colours and images extracted from social networks, voice recordings processed with homomorphic MFCC analysis, comments and reviews under product reviews, web-analytics of visited pages, environment parameters (geo, time, weather, holidays, events) and A/B test results.
What was done: Colours were extracted from the largest silhouette in a photo; voice was categorised by tone; the average face was computed from the audience's opinion leaders rather than from the brand's own audience; cropping was driven by face position and the useful background area. For customisation we settled on 10 key parameters, and for behaviour — on more than 60 parameters derived from the content people consume.
Outcome: Engagement uplift: up to +25% from colour customisation, up to +40% from voice, up to +30% from images and up to +10% from cropping. Of the 60+ behavioural parameters, demography and online activity weigh the most (41%), then finance and the consumer model (23%), then personality traits and interaction with advertising (17%).
Problem: Enrich a brand's marketing data with open sources and estimate in advance how the recommendation algorithms of social networks will promote a creative.
Data: Open profiles and published content, social connections, geolocation reconstructed from content and photos, plus additional metrics from mobile operators, Wi-Fi providers in shopping centres and fiscal data operators.
What was done: Content profiling, behavioural modelling and social-graph analysis: we select niche opinion leaders by audience intersection and clustering, forecast how a trend will develop, and compare the brand's audience with competitors'.
Outcome: Up to 90% of content in social networks is distributed by recommendation algorithms, so the analysis shows in advance how a creative will be promoted and reduces the campaign budget through organic spread of content.
Problem: Find future stars in time, evaluate how much the social network's recommendation algorithm likes the content, and compare the label's position with competitors.
Data: Data from any social networks and services in real time, tracks (similar-track search by mp3), brand and competitor audiences, and content consumption patterns.
What was done: We set up real-time trend monitoring, predictive models for future releases, competitive analysis of the audience with potential audiences, a search for opinion leaders who interacted with competitors' content, and similar-music analytics.
Outcome: A monitoring and forecasting loop for releases with cost optimisation of marketing communications.
Problem: Understand which properties of a track and its metadata drive promotion in recommendation systems, and separate real audience reactions from algorithmic noise.
Data: Audio features of tracks (tempo, rhythm, danceability, energy, instrument prominence, key, instrumentation, a cappella, cover/remake, mood), lyrics, metadata (authors, composers, historical data, playlists) and similar-track vectors.
What was done: We built the weighting of the main factors and a similar-track search on spectral spread, flux, centroid, rolloff, MFCC, chroma vector with deviation and zero-crossing rate; we also identified what distorts the picture — one account used by several people, music played at parties and nostalgia.
Outcome: The weight of the factors is 45% rhythm, 33% musical harmony, 17% metadata and 5% lyrics. About 20% of queries to recommendation systems return irrelevant results that are in fact audience A/B tests. For artists we computed an affinity index (demography, finance, consumer model, politics, advertising) — for example 0.87 / 0.89 / 0.97 / 0.72 / 0.90 for Taylor Swift.
Problem: Shorten the recruiter's work, verify the accuracy of resumes and widen the reach of the candidate search.
Data: Open sources: social networks, sites, blogs and forums, dating apps and games, public registries (EGRN, EGRUL, FSSP), job boards (HH, Superjob), classifieds and real-estate registers.
What was done: We searched for and verified candidates in open sources and ranked employees by efficiency for the HR service. The rules of safe use are set explicitly — do not disclose sources, do not disclose algorithms, justify legally as an expert opinion — and names the risks: racist algorithms and unjustified dismissals.
Outcome: Recruiter time is reduced, resumes are verified and the reach of the search grows. The client's figures are not published.
Problem: Collect the behaviour of tens of millions of visitors a week and give clients analytics in real time, not in a nightly batch.
Data: Data from the JS code and IMG pixel on client sites, logs from switching equipment, and 65 behaviour metrics per visitor read from a separate DBMS.
What was done: We built a storage on Ceph + MongoDB + PHP 7.4 FPM + Nginx: 12 backend servers of 64 cores and 500 GB behind a cloud router that spreads traffic by geography, 240 SSDs of 800 GB in five baskets behind a single controller (RAID 10 — about 85 TB and 1.5M IOPS, RAID 6 — 157 TB, RAID 5 — 152 TB), three levels of data with monthly-to-weekly and yearly-to-monthly roll-up, and Apache Solr kept alongside for full-text search. We chose Ceph over GlusterFS, MooseFS, HDFS and DRBD for speed and self-healing.
Outcome: About 50 million unique visitors a week are processed; the delay between an event and the client report is no more than 5 minutes, and the ML model refreshes behaviour metrics every 5 minutes.
Problem: For a music festival brand — find opinion leaders, count the real visitors and attract new people. The potential audience was 25 million.
Data: Offline: receipts from the fiscal data operator, Wi-Fi traffic, radiometry and loyalty cards. Online: published content and likes in thematic groups and posts.
What was done: We measured the active audience (the share of people who interacted with the content, the share of bots and of cynical followers) and the speed at which content spread — how fast posts travelled from one point of the network to another, how many people distributed them and what audience each of them engaged. Then we processed the top 150 accounts by reach.
Outcome: For the selected accounts 70% of followers were on target and content spread in 2–3 days. Face-recognition systems and over-targeted content are among the risks of the approach.
Problem: Cut the cost of servicing and minor repairs of locomotives by relocating repair crews instead of moving the machines.
Data: Real-time telemetry from rolling stock (more than 7,000 vibration sensors on engines, compressor units, wheel pairs and rolling bearings), technical certificates of admissible deviations, train and crew schedules, a map of free repair bays, the railway station graph and an internal CRM with failure marks.
What was done: The model receives the locomotive's repair index, the crew schedule, the train schedule, free repair bays, the station graph and CRM marks about breakdowns, and looks for anomalies in engines, compressors, wheel pairs and bearings.
Outcome: −12% rolling-stock maintenance costs, −20% payroll for staff waiting for tasks, +3% schedule adherence. The main mistake of the project: the increased cost of buying and delivering spare parts to repair sites was not taken into account.