Hello, it’s Kolja

Highlights

• Inference loop: designed a frame → landmarks → model → output flow
• Structure: created clear data objects for gestures and predicted tokens
• ML: experimented with RandomForest and TensorFlow-based approaches
• Realtime: used MediaPipe for hand tracking on webcam frames

Tech

What I’d do next

• Improve temporal smoothing to reduce irregular variations in predictions
• Turn it into a simple web demo and publish a short case study

Highlights

• Real-time Sync: utilized Flask-SocketIO to broadcast game state changes instantly to all clients
• Game Logic: implemented complex rules for attacking, defending, and turn rotation in Python
• Idvividual Play: designed server-authoritative logic that only sends a player their own hand data
• Repeatability: created a reconnection system using localStorage so players can rejoin active sessions

Tech

What I’d do next

• Make user experience more fun
• A tutorial for new players
• Add bots with a variety of difficulty

Highlights

• Data cleaning: merged multiple sources across 90 neighborhoods
• Geo: used GeoPandas for neighborhood-level joins and spatial context
• Visualization: created 10+ Matplotlib charts to show patterns + tradeoffs
• Composite indicator: built a simple scoring system to compare neighborhood well-being

Tech

What I’d do next

• Document data sources and assumptions more formally
• Turn the score into an interactive map
• Add sensitivity analysis for weighting choices

Highlights

• No Frameworks: Built with pure HTML/CSS to ensure lightweight performance
• Responsive: Uses CSS Grid and clamp functions for fluid scaling on mobile/desktop
• Accessibility: Semantic tags, high contrast, and keyboard navigation support
• Design: Custom glass-morphism aesthetic with CSS variables for easy theming

Tech

What I’d do next

• Implement a light/dark mode toggle
• Add a blog section for technical writing
• Make the cards more engaging

Highlights

• OOP Design: Structured game objects (player, obstacles, terrain) using inheritance and polymorphism
• Game Loop: Implemented a real-time update and render cycle for smooth gameplay
• Collision Detection: Created bounding-box logic to handle interactions between the player and traffic
• State Management: Managed game states (menu, playing, game over) and score tracking

Tech

What I’d do next

• Add sound affects and animations to the game
• Improve the looks of objects and the background
• Add more fun and exciting features to the game

Operations

• Create: Inserts a new stat row
• Read: Shows stats (all, sorted, or filtered)
• Update: Changes one stat for a row
• Delete: Removes a row by ID

Tech

What I’d do next

• Create more tables for the stats, like for games, seasons, player stats, team stats
• Use it in an actual project involving basketball
• Make the UI more pleasant and easier to interact with

Highlights

• Agentic Build: Directed an AI agent to implement the app based strictly on a comprehensive SPEC.md file.
• Data & Auth: Engineered a local login system, JSON file I/O for question banks, and secure, non-human-readable score tracking.
• Adaptive Logic: Implemented a feedback loop that adjusts future question selection based on user ratings.
• AI Code Review: Prompted a fresh AI agent session to audit the raw codebase for bugs, unhandled errors, and security flaws.

Tech

What I’d do next

• Experiment with parallel AI reviewers (e.g., separate agents for security and code style).
• Automate the generation of new study subjects into the JSON format.
• Refine the terminal UI with richer text formatting and colors.

Highlights

• Spec-driven build: Two-tab flow—My Accomplishments (alphabetical checklist with per-country filters) and My Stats (completion %, summed elevation, and “regions fully covered” for the active filter).
• Persistent state: Visited locations saved locally with a versioned JSON shape, ID sanitization, and user-visible handling when storage read/write fails.
• Derived metrics: Stats recompute from filtered subsets (e.g. elevation totals and region completion when every sight in a region is checked).
• Agent workflow: Implement from spec → structured code review (REVIEW.md) → targeted fixes and visual improvements (navigation shell, clearer hierarchy, haptics-friendly interaction).

Tech

What I’d do next

• Add automated tests for filter/stats edge cases and persistence migrations.
• Optional cloud backup or export so progress survives device changes.
• Accessibility pass (labels, contrast, larger touch targets) and optional localization.

Highlights

• Six games & tools: Guess Who, Wordle, Student of the Day, Impostor, Connections, and a Study Abroad Matcher
• Daily rotation: a scheduled pick per game so every day surfaces a new mystery Panther
• Auth & profiles: Supabase handles sign-up, login, opt-in, and per-user stats (wins, streaks, daily status)
• Study Abroad Matcher: swipeable deck that re-ranks 35 Pitt programs by Euclidean distance on a normalized feature vector

Tech

What I’d do next

• Add more community-driven games and tools
• Improve mobile UX and add push/email daily reminders
• Expand beyond Pitt to other universities

What I did

• Paper: first-authored a submission to an IROS 2026 workshop — all three models read tools and objects at 81–84%, but dropped to 57–68% on motion and grip
• Benchmark: built the 311-claim Ego4D benchmark behind that result, hand-verifying every claim across 85 clips from six countries and giving each model the same 16 frames
• Scale: labeled 7,688 Epic Kitchens clips on binary technique axes for a robot-learning collaboration, running Hugging Face VLMs across four GPUs
• Clinical study: owned the vision-language side of an independent-living study, converting a paper self-care assessment into 58 automated checks on first-person kitchen video

Skills used

Key takeaway

A careful benchmark can expose a gap that headline accuracy numbers completely hide.

What I did

• Pipeline: four-stage system following Loughran and McDonald that downloads, cleans, and scores SEC 10-K filings — 3,179 filings from 468 companies (2019–2025)
• LLM Scoring: designed the scoring pipeline the project now runs on, rating five kinds of AI capability, and extended it to Factiva news so both sources are measured identically
• Cost Engineering: cut a full 3,179-filing run to under $60 by scoring only text around keyword hits, skipping the model on empty categories, and checkpointing every filing
• Benchmarking: engineered logic to score brands against "Digital Pillars," the project's key performance indicators

Skills used

Key takeaway

Reliable scoring at scale comes from careful prompt and cost design, not from throwing more tokens at the problem.

What I did

• Led: weekly lab sessions for ~20 students
• Debugging: helped students systematically find and fix issues
• Teaching: explained fundamentals clearly (loops, arrays, methods, OOP basics)
• Support: held office hours and guided students through assignments

Skills used

Key takeaway

Teaching made me faster at diagnosing code and explaining technical ideas in simple steps.

What I did

• SQL: performed joins on internal company datasets to support reporting needs
• SAP workflows: observed and assisted with structured business processes
• APIs: reviewed and documented integrations using M-Files
• Collaboration: worked with IT + business teams on how systems connect

Skills used

Key takeaway

I got my first real view of how “messy real life” data becomes structured business decisions.

What I did

• Discrete Structures: logic, proofs, sets, relations, induction
• Intro Programming: Java fundamentals, debugging strategies, code reading
• Coaching: helped students develop step-by-step approaches to problems

Skills used

Key takeaway

The best explanations are structured like algorithms: clear inputs, steps, and checks.