Cases
Cases
Straight to the point: context → risk → solution → impact. NDA-safe where needed.
AI‑Powered Analytics Cockpit for Vehicle Auction Platform
Embedded AI assistant that turns raw behavioral data into real‑time user portraits, LLM‑generated vehicle summaries, explainable recommendations, and automated engagement tools for platform staff.
Context
- —Analysts had to manually cross‑reference user activity, vehicle specs, and market context to make any engagement decision.
- —No quick way to obtain a buyer portrait, an objective car summary, or a ranked interest score – outreach was gut‑driven.
Risk
- —Manual processes limited personalisation and made scaling impossible without adding headcount.
- —The lack of transparent, explainable recommendations caused missed cross‑sell and upsell opportunities.
- —Ad‑hoc LLM experiments had no cost control, making budget forecasting and ROI measurement impossible.
Solution
- —Built an embedded Next.js dashboard, statically exported into the Django admin, sharing auth and session state.
- —Designed async Celery pipelines that ingest platform events, compute embeddings, run clustering, and invoke LLM enrichment.
- —Surfaced real‑time user profiles, vehicle AI‑analysis, business arguments, and a one‑click notification form for staff.
- —Added per‑request AI cost logging, token tracking, and consistency audits to keep spending transparent and under control.
Impact
- —Staff see AI‑curated user portraits, vehicle summaries, and interest scores the moment they log in – zero manual prep.
- —Targeted notifications reach the right buyers, increasing bid density and conversion rates.
- —The platform gained a scalable, auditable intelligence layer that actually gets cheaper per unit as it grows.
DjangoNext.jspgvectorCeleryRedisLLMPrompt Engineering
Custom BI Platform: Code‑First Analytics That Outperforms Vendor Tools
Fully custom business intelligence platform built with FastAPI, Next.js, and TypeScript. Replaces rigid vendor tools with domain‑specific metrics, real‑time what‑if modeling, automated alerts, and AI‑accelerated development.
Context
- —Management needed precise capacity planning based on individual norms, production calendars, and complex project allocations.
- —Off‑the‑shelf BI tools could not model the company’s unique workflows without expensive customisation.
Risk
- —Relying on generic BI would force the business to work around the tool rather than the tool adapting to the business.
- —Manual reporting and spreadsheet‑based forecasts were error‑prone, slow, and impossible to trust for critical resourcing decisions.
Solution
- —Designed a code‑first analytics platform where every metric is implemented as typed, testable code—not a visual configuration.
- —Integrated production calendar, vacation data, and individual work norms into a central calculation engine.
- —Delivered role‑based dashboards (executive, department, individual) with interactive filtering and instant recalculation.
- —Built what‑if scenario modeling, an intelligent alert system, and one‑click XLSX exports.
Impact
- —Managers save 5–10 hours per week previously spent compiling and validating manual reports.
- —Resource decisions are now based on accurate, calendar‑aware data, eliminating capacity overestimation.
- —The platform’s flexibility allows new metrics to be added in days, not vendor release cycles.
FastAPINext.jsTypeScriptPostgreSQLRedisDocker Compose
Cross‑market arbitrage analytics: real‑time signals & decision engine
Production‑grade infrastructure that synchronises MOEX futures and Forex spot quotes, computes arbitrage coefficients with sub‑second latency, models hidden costs, enforces risk rules, and delivers deterministic deal plans through a real‑time dashboard.
Context
- —Traders needed a single source of truth for multi‑leg arbitrage opportunities across fragmented brokers and data formats.
- —Manual calculation of swap costs, margin, and commissions made profitability estimates unreliable and slow.
Risk
- —Stale or misaligned quotes created phantom arbitrage signals and direct financial exposure.
- —Without alert cooldowns and explicit risk checks, operators risked over‑leveraging and notification fatigue.
Solution
- —Designed a microservice‑oriented pipeline: dedicated adapters for MOEX (Alor/Finam) and cTrader OpenAPI, Redis‑backed quote cache, and a decoupled calculation engine.
- —Built deterministic financial calculator with triple‑swap logic, multi‑leg margin accounting, and commission/slippage modeling.
- —Implemented independent entry/exit alert cooldowns, Telegram integration, and a deal‑plan simulator returning VIABLE / MARGINAL / NOT VIABLE verdicts.
- —Delivered a Next.js dashboard with live WebSocket updates, financial reports, and manual override controls.
Impact
- —Operators receive synchronised, validated arbitrage signals that respect all real‑world costs, eliminating phantom trades.
- —Risk‑aware decision support with explicit margin usage, position concentration checks, and cooldown‑protected notifications.
- —Platform architecture remained responsive and auditable under continuous market activity, with full deployment and operational documentation.
FastAPIRedisWebSocketFIX/OpenAPINext.jsPostgreSQLDocker Compose
Chat bot & mini-app: data collection and marketing measurability
Automation of scenarios, contact collection, transparency of sources and conversions.
Context
- —Need to collect data from packaging (QR) and understand channel effectiveness.
Risk
- —Marketing without measurability and loss of source data.
Solution
- —Bot on n8n + notification and reminder scenarios.
- —Mini-app for extended profile.
- —UTM tracking and click proxying.
Impact
- —Automated data collection.
- —Channel transparency and controlled effectiveness.
n8nTelegramMini AppTracking
Voice-to-spec: transcription and structuring
Concept: voice → spec draft to accelerate start and lower client barrier.
Context
- —Client finds it easier to dictate ideas than write a document.
Risk
- —Loss of requirements and chaos in task setting.
Solution
- —Transcription + structuring and initial decomposition.
Impact
- —Fast and clear start of requirement discussion.
WhisperLLMTelegram
R&D / internal product.
Data platform: ingestion → management → metadata
Platform layer for collecting, managing, and preparing data for BI/AI.
Context
- —Disparate sources and need for standardized data delivery.
Risk
- —‘Manual’ integrations, data inconsistency, rising change costs.
Solution
- —Connectors and primary data delivery layer.
- —Management interfaces and change control.
- —Metadata and governance layer.
Impact
- —Unified predictable data layer.
- —Higher quality, lower change cost.
ConnectorsMetadataBI Integration
BI product: visualizations from task descriptions
Product prototype: business describes a task — system suggests visual answers.
Context
- —Need to bring analytics closer to business language.
Risk
- —High BI entry barrier and long ‘question → answer’ cycle.
Solution
- —App + metadata layer + visualization generation.
Impact
- —Faster management answers.
Next.jsMetadata LayerLLM
Concept can be shown as a direction.
Elderly care monitoring: AI event detection and alerts
System where reliability, privacy, and timely response matter.
Context
- —Need to detect incidents and notify caregivers.
Risk
- —False positives or missed events.
Solution
- —Detection layer + alerts and scenario quality control.
Impact
- —Stable event response while preserving privacy.
AIAlertingObservability
BI rollouts across different industries
Experience building management analytics in companies of various profiles.
Context
- —Different sources, roles, and access requirements.
Risk
- —‘Raw’ dashboards without data trust.
Solution
- —Semantics, marts, quality control, governance.
Impact
- —Unified metrics and daily management value.
WarehousingETL/ELTBI
Phrased without revealing sensitive details.
Global macro on currencies: factors and regular updates
Research layer: data collection, factor structuring, reporting.
Context
- —Need to regularly update factor picture for FX.
Risk
- —Decisions without up-to-date factor structure and evidence.
Solution
- —Data collection, factor framework, regular updates.
Impact
- —Transparent analytical base for strategies.
Data PipelinesResearch Framework
R&D / internal layer.