ps

Data Scientist & AI Engineer

Patryk Seweryn

Turning complex data into clear decisions, manual processes into automated workflows, and ambitious ideas into production-grade AI systems.

Patryk Seweryn

Experience

Production systems I've designed and shipped.

AI Platform for Regulatory Screening

Automating regulatory screening so compliance teams never miss a flag.

Python LLMs pydantic-ai Streamlit Docker FastAPI
Read more

LLM-Powered E-commerce Search Engine

Making products findable through natural language search.

Python LLMs RAG LangChain FAISS
Read more

Smart Delivery Detection for Home Cameras

Spotting package deliveries from home camera feeds.

Python PyTorch ONNX OpenCV Azure Cloud
Read more

DPV Detector (Drones / People / Vehicles)

Real-time detection of drones, people, and vehicles on edge hardware.

Python NVIDIA Jetson Detectron2
Read more

Proactive Failure Detection for Smart Thermostats

Detecting thermostat failures before customers notice.

PySpark Databricks Delta Tables Azure
Read more

ML-Driven Efficiency Optimization for Gas Water Heaters

Squeezing more efficiency out of home water heaters through data.

Pandas NumPy Scikit-Learn PySpark
Read more

Automated Document Router for Business Workflows

Sorting business documents automatically so nobody has to read them first.

Python DVC Poetry Tesseract OCR SciPy
Read more

Revenue Prediction by Location

Predicting store revenue before opening the doors.

Python XGBoost Scikit-Learn SHAP PostgreSQL Dash Plotly Leaflet OpenStreetMap
Read more

Training Data Pipeline for Detection Models

Building better training data through systematic curation.

Python PyTorch Pandas NumPy Matplotlib
Read more

Customs Tariff Classifier

Assigning customs codes automatically so humans don't have to look them up.

Python Scikit-Learn Flask
Read more

Automated Product Recommendations for E-commerce

Recommending products by what they actually are, not what others bought.

Python Decision Trees Dash
Read more

Deep Learning Denoiser for Thermal Cameras

Making thermal camera footage actually usable.

TensorFlow
Read more

Fine-tuning Counter-UAV Detection Model

Cutting false alarms in drone detection without losing real threats.

Python Computer Vision
Read more

3D Body Scan Denoiser for Mobile LiDAR

Cleaning up noisy 3D body scans from phone cameras.

Python PyTorch Open3D Blender
Read more

Semantic Label Clustering for Large-Scale Detection

Making sense of model predictions when there are hundreds of labels.

NLP Computer Vision
Read more

Detection Model Benchmark for Production Deployment

Finding the best detection model through systematic evaluation.

Python PyTorch NumPy Matplotlib
Read more

Flying Objects Classification on Radar Data

Teaching radar systems to tell drones from birds.

Python MLflow
Read more
LLMs & AI Agents

AI Platform for Regulatory Screening

Automating regulatory screening so compliance teams never miss a flag.

Banks must screen every transaction against regulatory watchlists — fragmented across authorities, constantly updated, arriving in every document format. I built a platform that extracts entities from incoming documents, matches them against consolidated watchlists, and produces full audit trails.

Python LLMs pydantic-ai Streamlit Docker FastAPI

Problem

Financial institutions must screen transactions against regulatory watchlists, but those lists come from multiple authorities, in different formats, and change frequently. Compliance teams checked documents manually — slow turnaround, real risk of missed matches.

Approach

Built a platform that consolidates watchlist data into a single searchable index, then uses AI to extract entities from incoming documents regardless of format. Each match produces a fully traceable result.

Result

Analysts upload documents and get instant, reliable matches with full auditability — replacing hours of manual cross-referencing.
LLMs & AI Agents

LLM-Powered E-commerce Search Engine

Making products findable through natural language search.

Customers of a large e-commerce platform struggled to find what they needed — especially when they didn't know the exact product name. I built a search engine that understands everyday language and returns the right results, even in grammatically complex Polish. Then took it further by iterating on accuracy in a second phase.

Python LLMs RAG LangChain FAISS

Problem

A large e-commerce platform had a classic search problem: customers who knew exactly what they wanted could find it, but everyone else was lost. Polish language made it harder — a single noun can have over a dozen grammatical forms, and users naturally mix technical terms with colloquial descriptions.

Approach

Built a search layer that understands the intent behind queries — extracting product categories, desired features, and price constraints from conversational input. Special attention went to handling Polish morphology so different word forms map to the same products. After the initial build, continued with a second phase focused on improving accuracy further.

Result

Delivered a search engine where non-technical users could describe what they needed in their own words and actually find it — no need to know exact product names or navigate complex category trees.
Computer Vision

Smart Delivery Detection for Home Cameras

Spotting package deliveries from home camera feeds.

Home cameras generate hours of footage, but owners only care about specific moments — like when a package shows up. I built a detection system that identifies deliveries from camera feeds, backed by a full cloud pipeline for training and deploying the model.

Python PyTorch ONNX OpenCV Azure Cloud

Problem

Home security cameras record continuously, but what owners actually want to know is simple: did a package arrive? Without automated detection, they’d have to scrub through footage or just hope they noticed in time.

Approach

Built an object detector for packages and set up a complete cloud pipeline — from data collection and model training to deployment — so the model could be iterated on and improved continuously.

Result

Delivered a working detection system with a production-ready cloud infrastructure for ongoing model updates.
Computer Vision

DPV Detector (Drones / People / Vehicles)

Real-time detection of drones, people, and vehicles on edge hardware.

Security systems need to detect threats in real time, on-site, without relying on cloud infrastructure. I built an object detector that identifies drones, people, and vehicles from both thermal and RGB camera feeds, running directly on edge hardware at ~15 FPS.

Python NVIDIA Jetson Detectron2

Problem

Security operators need to know what’s happening on-site right now — not after a round trip to the cloud. The system had to detect drones, people, and vehicles from both thermal and standard cameras, all running on compact edge hardware fast enough for real-time use.

Approach

Built an object detector optimized for edge deployment — running on NVIDIA Jetson hardware with both thermal and RGB sensor inputs. Focused on keeping inference fast enough for live monitoring while maintaining reliable detection across object types.

Result

Delivered a detector running at ~15 FPS on edge hardware, supporting live demonstrations that helped the company acquire new clients.
Data & Analytics

Proactive Failure Detection for Smart Thermostats

Detecting thermostat failures before customers notice.

Instead of waiting for customers to report a broken thermostat, I built an analytics framework that detects failures proactively — analyzing temperature and operational data to catch problems early and classify what went wrong.

PySpark Databricks Delta Tables Azure

Problem

When a thermostat fails, the company typically found out the hard way — a customer calling to complain. By then, the problem had already caused discomfort and frustration. There was no systematic way to detect failures early from the data the devices were already generating.

Approach

Built a time-series analytics framework that monitors temperature readings and operational data to spot anomalies that indicate different types of failures. Used a decision-tree approach to classify what went wrong, so support teams know what they’re dealing with before reaching out.

Result

Delivered a proactive detection system that identifies failing thermostats from operational data — turning reactive customer support into preemptive outreach.
Data & Analytics

ML-Driven Efficiency Optimization for Gas Water Heaters

Squeezing more efficiency out of home water heaters through data.

Home water heaters waste energy when they run suboptimally — but without data, it's hard to know what 'optimal' even looks like. I analyzed operational data and built ML models to understand how to get the most efficient performance out of gas water heaters.

Pandas NumPy Scikit-Learn PySpark

Problem

Gas water heaters consume energy regardless of whether they’re running efficiently. Without analyzing actual operational data, there was no clear picture of what optimal usage looks like — or how much energy was being wasted.

Approach

Analyzed operational data and built ML models to identify what drives efficiency — understanding how usage patterns, settings, and conditions affect energy consumption.

Result

Delivered actionable insights into how gas water heaters can be operated more efficiently, backed by data and predictive models.
LLMs & AI Agents

Automated Document Router for Business Workflows

Sorting business documents automatically so nobody has to read them first.

Contracts, invoices, forms — every document needed to land in the right category, and someone had to read it to decide where. I built a system that classifies documents automatically based on their content, removing the need for manual review.

Python DVC Poetry Tesseract OCR SciPy

Problem

Every incoming document — contracts, invoices, forms — needed to be categorized before it could be processed further. Someone had to open it, skim it, and decide where it belonged. At volume, this was a bottleneck.

Approach

Built a classification pipeline that extracts text from documents using OCR and analyzes their content with keyword-based matching to assign each document to the right category automatically.

Result

Delivered a system that classifies incoming documents on the spot — no human review needed to route them correctly.
Data & Analytics

Revenue Prediction by Location

Predicting store revenue before opening the doors.

Where should the next store go? I built predictive models that estimate revenue for potential locations based on geospatial data, plus an interactive dashboard for exploring predictions and comparing sites — turning expansion planning from gut feel into data-driven decisions.

Python XGBoost Scikit-Learn SHAP PostgreSQL Dash Plotly Leaflet OpenStreetMap

Problem

Opening a new store is a big bet — and the decision of where was largely based on experience and intuition. There was no systematic way to estimate how much revenue a location would generate before committing to it.

Approach

Built predictive models using geospatial features — what’s nearby, foot traffic patterns, demographics — to estimate revenue for any given location. Wrapped it in an interactive dashboard with map visualizations so expansion teams could explore and compare sites visually.

Result

Delivered a decision-support tool where planners can evaluate potential locations, see predicted revenue, and understand what drives performance — backed by an automated data pipeline that keeps everything up to date.
Computer Vision

Training Data Pipeline for Detection Models

Building better training data through systematic curation.

A model is only as good as its training data. I built a pipeline that systematically evaluates and filters video datasets to select the most effective training samples — not by hand, but through statistical analysis.

Python PyTorch Pandas NumPy Matplotlib

Problem

Object detection models need high-quality training data, but figuring out which data is high-quality isn’t obvious. Manual curation doesn’t scale, and using everything available often means feeding the model noise alongside signal.

Approach

Built a pipeline that statistically evaluates which video samples contribute most to model performance, then automatically filters datasets to keep only what matters. Replaced subjective curation with measurable, repeatable selection criteria.

Result

Delivered a systematic dataset curation process that consistently produced higher-quality training sets, directly boosting model accuracy.
LLMs & AI Agents

Customs Tariff Classifier

Assigning customs codes automatically so humans don't have to look them up.

Every product that crosses a border needs a customs tariff code — and someone has to assign it. I built a classifier that learns from previously coded products to automatically assign tariff codes to new ones.

Python Scikit-Learn Flask

Problem

Every imported or exported product needs a customs tariff code, but assigning them requires domain expertise and manual lookup. As product catalogs grow, this becomes a bottleneck that slows down operations.

Approach

Built a classification model that learns the mapping between product descriptions and tariff codes from historical assignments, then predicts codes for new products automatically. Wrapped it in a web service for easy integration.

Result

Delivered an automated tariff classifier that assigns codes to new products based on patterns learned from the existing catalog.
LLMs & AI Agents

Automated Product Recommendations for E-commerce

Recommending products by what they actually are, not what others bought.

Most recommendation engines rely on purchase history — but that doesn't help when you want products with similar specs. I built a system that matches products by their technical features, replacing a manual process where people paired 'similar' items by hand.

Python Decision Trees Dash

Problem

When a customer looks at a product, they want to see alternatives with similar specs — not just what other people happened to buy. But pairing products by technical features was done manually, which couldn’t keep up with a large and constantly changing catalog.

Approach

Built a recommendation engine that compares products by their actual technical attributes rather than purchase history. Used decision trees to learn which features matter most for each product category.

Result

Delivered automated, feature-based recommendations across the full catalog — replacing a manual curation process that couldn’t scale.
Computer Vision

Deep Learning Denoiser for Thermal Cameras

Making thermal camera footage actually usable.

Thermal cameras produce inherently noisy images — artifacts and distortions that make the footage hard to use for anything downstream. I applied deep learning denoising techniques to clean up infrared imagery, significantly reducing noise across multiple frequency bands.

TensorFlow

Problem

Thermal cameras capture what visible light cameras can’t, but their output is inherently noisy — low-frequency distortions, high-frequency grain, and visual artifacts that obscure the actual scene.

Approach

Applied deep learning denoising techniques based on recent research to tackle different types of noise separately — each frequency band required a different treatment to clean up effectively.

Result

Delivered significantly cleaner infrared imagery, removing the noise that made raw thermal footage impractical for downstream use.
Computer Vision

Fine-tuning Counter-UAV Detection Model

Cutting false alarms in drone detection without losing real threats.

A counter-UAV detection system was triggering too many false alarms — especially confusing birds with drones. I fine-tuned the model to reduce false positives while keeping detection reliable across all target categories.

Python Computer Vision

Problem

A counter-UAV detection system was generating too many false alarms — birds in particular were frequently misclassified as drones. Operators couldn’t trust the alerts, which defeated the purpose of automated detection.

Approach

Fine-tuned the existing model with a focus on separating visually similar classes — specifically birds and drones. Delivered performance benchmarks and validation videos showing before/after behavior.

Result

Brought false alarm rates down to a practical level while maintaining or improving detection across all other categories.
Computer Vision

3D Body Scan Denoiser for Mobile LiDAR

Cleaning up noisy 3D body scans from phone cameras.

Phone LiDAR sensors can scan a person's body in 3D, but the raw data is noisy — full of artifacts that make measurements unreliable. I built a deep learning model that cleans up these scans, trained on synthetic 3D data to achieve reliable accuracy on real scans.

Python PyTorch Open3D Blender

Problem

iPhone LiDAR cameras can capture 3D body scans, but the raw point clouds are noisy — scattered points, surface artifacts, and distortions that make the data unreliable for any downstream use.

Approach

Built an encoder-decoder neural network that learns to reconstruct clean geometry from noisy input. The model was trained on synthetic 3D data generated alongside the project in Blender, then validated on real iPhone scans.

Result

Delivered a denoising model that cleans up raw 3D body scans, making phone-based scanning practical.
Computer Vision

Semantic Label Clustering for Large-Scale Detection

Making sense of model predictions when there are hundreds of labels.

When an object detector has hundreds of classes, analyzing its predictions becomes overwhelming. I built a tool that semantically groups similar labels — so instead of reviewing hundreds of classes one by one, analysts can focus on meaningful clusters.

NLP Computer Vision

Problem

Object detectors with hundreds of classes produce a wall of predictions that’s hard to make sense of. When similar classes get confused with each other — say, different types of vehicles or bird species — the patterns are buried in noise.

Approach

Built a semantic grouping tool that clusters similar labels based on their meaning, so analysts can see confusion patterns at the group level instead of sifting through individual classes one by one.

Result

Turned an unmanageable prediction matrix into a structured, human-readable analysis — making it easy to spot where the model struggles and why.
Computer Vision

Detection Model Benchmark for Production Deployment

Finding the best detection model through systematic evaluation.

When you have multiple detection model candidates, picking the right one for production isn't just about accuracy — it's about reliability and deployment feasibility. I led a team that trained and evaluated multiple model variants to find the one ready for real-world use.

Python PyTorch NumPy Matplotlib

Problem

The client needed an object detection model in production, but had multiple architectural options to choose from. Without a rigorous comparison, picking one would be a gamble — the wrong choice could mean poor real-world performance.

Approach

Led a team through a systematic evaluation — training multiple model variants and assessing each on performance, reliability, and deployment feasibility. Built clear benchmarks so the comparison was objective, not subjective.

Result

Delivered a validated recommendation backed by thorough benchmarks, giving the client confidence to move forward with the right model.
Radar

Flying Objects Classification on Radar Data

Teaching radar systems to tell drones from birds.

Early warning systems need to know what's flying toward them — and radar alone doesn't make that obvious. I built a classifier that distinguishes drones, birds, helicopters, and airplanes from raw radar signals by engineering features that capture how each type of object moves and reflects.

Python MLflow

Problem

Early warning systems detect flying objects, but raw radar data doesn’t reveal what is flying. A bird and a small drone can produce similar signatures, making it hard for operators to assess threats quickly and accurately.

Approach

Engineered features from radar signal characteristics — how objects move, reflect, and behave over time — to build a classifier that distinguishes between drones, birds, helicopters, and airplanes. Delivered with a detailed accuracy analysis across different range and signal quality conditions.

Result

Significantly improved classification accuracy, giving operators a reliable way to identify airborne objects from radar data alone.

Research

Academic work and technical explorations.

Search Intent Classifier for E-commerce

Understanding what users actually want when they search.

Python XGBoost UMAP Sentence Transformers FastAPI Docker
Read more

LLM-Powered Customer Feedback Explorer

Turning customer feedback into answers you can just ask for.

Python LLMs Prompt Engineering
Read more

Radar Object Classification R&D

Proving that flying objects can be classified from impulse radar data.

Python DVC
Read more
LLMs & AI Agents

Search Intent Classifier for E-commerce

Understanding what users actually want when they search.

Search queries rarely say exactly what people mean. I built a classification system that identifies the intent behind search queries — what the user is actually trying to accomplish — using an ensemble of text embeddings and machine learning.

Python XGBoost UMAP Sentence Transformers FastAPI Docker

Problem

Users type short, ambiguous queries into search boxes. Without understanding the intent behind a query, a search system can only match keywords — which often means irrelevant results.

Approach

Built a classification pipeline that converts search queries into semantic embeddings, reduces their dimensionality, and feeds them into an ensemble classifier. Deployed as a standalone service for easy integration.

Result

Demonstrated reliable intent classification from search queries in a proof-of-concept, ready to be integrated into a production search pipeline.
LLMs & AI Agents

LLM-Powered Customer Feedback Explorer

Turning customer feedback into answers you can just ask for.

Customer feedback comes in piles of documents — surveys, spreadsheets, reports. I built a system that extracts questions and answers from these files and lets users query the results conversationally, like asking a colleague who's read everything.

Python LLMs Prompt Engineering

Problem

Customer feedback was scattered across documents in different formats. Getting insights meant someone had to read through everything manually — and even then, spotting patterns across dozens of files was nearly impossible.

Approach

Built a pipeline that extracts Q&A pairs from uploaded documents regardless of format, feeds them into an LLM, and exposes a conversational interface where users can ask questions about the feedback naturally.

Result

Delivered a research prototype where users upload feedback documents and immediately start asking questions about what customers said — no manual reading required.
Radar

Radar Object Classification R&D

Proving that flying objects can be classified from impulse radar data.

Can you tell a drone from a bird using only impulse radar data? That was the question. I built a proof-of-concept that tracks detected objects and analyzes their signal characteristics to classify what's flying — proving the approach is viable.

Python DVC

Problem

The client wanted to know if their radar data could be used for more than just detection — specifically, whether it was possible to classify what was flying. No one had tried it with this type of radar signal before.

Approach

Built a proof-of-concept that processes radar detections into object tracks, then analyzes signal characteristics along those tracks to classify flying objects.

Result

Demonstrated that classification from impulse radar data is feasible, providing the evidence needed to greenlight further development.

Interested in working together?

Let's talk

Start typing to search...