Maksym S.

Project overview:

The project involved developing an AI-driven demand forecasting system for a large network of distributed retail locations (scalable to thousands). The solution enabled fully automated inventory replenishment by predicting product demand for fast-moving SKUs across all categories (food, beverages, non-food). The system addressed inefficiencies in manual ordering processes, preventing both stockouts and overstock. Technical details: Forecast horizon: aligned with operational constraints (e.g., delivery cycles) Error metric: MAPE (baseline: simple average-based forecasting) Clustering: HDBSCAN on behavioral patterns Inference mode: batch (triggered by operational events)

Responsibilities:

• Designed and implemented a Temporal Fusion Transformer (TFT) for multi-horizon time series forecasting
• Built feature pipelines incorporating historical sales, seasonality patterns, external signals, and location grouping
• Applied clustering techniques to group locations by behavioral patterns, enabling model generalization
• Implemented cold-start handling via data-driven grouping for onboarding new locations without historical data
• Configured forecasting aligned with operational constraints (e.g., delivery cycles)
• Set up automated pipelines for recurring forecast generation

Achievements:

Forecast accuracy improved ~3x compared to baseline, significantly improving supply chain efficiency and enabling automated ordering without human intervention.

Technology stack:

Project overview:

AI support agent automating Tier-1 customer support across chat-based channels (web, mobile, etc.), handling hundreds of thousands of tickets annually. Technical details: Automation metric: AI-handled vs total tickets Agent scope: full conversation handling (classification, resolution, escalation) Embeddings: locally hosted models Summarization: locally hosted models Clustering: HDBSCAN Vector DB: Qdrant

Responsibilities:

• Architected a multi-step reasoning agent using LangGraph with specialized chains for complex cases
• Built a RAG system on top of support knowledge sources (transcripts, documentation, communication logs)
• Implemented a lightweight validation model to verify agent responses before delivery
• Created a clustering-based self-learning loop to identify emerging issues and continuously improve the knowledge base
• Developed knowledge ingestion pipelines (chunking, embedding, deduplication, augmentation)
• Defined QA processes with human-in-the-loop validation

Achievements:

AI now handles more tickets than human agents, significantly reducing support workload and enabling human agents to focus on complex issues.

Technology stack: