Bálint Balázs
Software Engineer01 / Own product · mobile app and backend
Sightway
Strength training, logged fast and explained clearly.
- Role
- Solo, end to end
- App
- React Native · Expo · TypeScript
- Backend
- ASP.NET Core · PostgreSQL
- Quality
- Jest · xUnit · Testcontainers
01 / How it is built
The stack
01
Interface
React Native · Expo
02
Feature logic
TypeScript · pure modules
features/workout/ confirm-start.ts next-set.ts rest-timer.ts set-suggestion.ts session-stats.ts
03
Offline storage
SQLite · Drizzle
CREATE TABLE set_logs ( weight_kg real NOT NULL, reps integer NOT NULL, rpe real, e1rm_kg real, completed_at integer
04
API
ASP.NET Core · .NET 10
POST /api/sync GET /api/today GET /api/analytics/records POST /api/coach GET /api/nutrition/meals
05
Database
PostgreSQL · EF Core
Npgsql.EntityFrameworkCore Api → Core → Data layers xUnit + Testcontainers
06
Server
Docker · Coolify · own VPS
FROM dotnet/aspnet:10.0 coolify → traefik → api api.sightway.app · TLS
In the app
Solo project
01Today
02Live workout
03Progress
04Exercises
05AI coach
06Nutrition
02 / Study project · team of two
Team of two
Food discovery as a vertical video feed, with ordering built in.
- My part
- Web, content studio, payments, CI
- Web
- Next.js 16 · React 19 · TypeScript
- API
- .NET 10 · EF Core · PostGIS
- AI
- Gemini · Qdrant · FFmpeg
02 / Content studio
The lineup
Six agents turn one dish into a short video.
- 01
Description
What the dish is
- 02
Storyboard
Shots for a 9:16 clip
- 03
Voiceover
Script, then speech
- 04
Caption
Burned-in subtitles
- 05
Hook
Opening line from trending hooks
- 06
QA
Score gate before rendering
- Guardrails
- JSON schema check on every input and output
- Trace per agent: input hash, latency, model
- Hook retrieval from a Qdrant vector index
- Prompt-injection guard on restaurant input
02 / What I built
My part
- 01
Kitchen board
Drag and drop, live over SignalR
- 02
Analytics
Revenue, feed and hook insights
- 03
Subscriptions
Stripe billing and Connect payouts
- 04
Split checkout
Group orders with a refund saga
- 05
Onboarding
Restaurant setup wizard
- 06
Delivery
Multi-arch Docker images, SSH deploy
CreditMobile app and personalisation by my project partner.
03 / Bachelor's thesis · with Dynatrace
Security for AI agents
Agent skills are plain instructions plus helper files. Can a scanner tell when one would make an agent do something it should not?
01
Rules
Static analysis
02
Data flow
Sources and sinks
03
Semantics
LLM-assisted review
04
Runtime traces
Recorded, not a live sandbox
scan · SKILL.md
$ scan ./release-notes/SKILL.md
✓ruleschecked instructions and helper files
✓dataflowtraced inputs to sensitive sinks
✓semanticreviewed intent of the instructions
✓runtimecompared traces with declared permissions
>reportfindings with evidence
- 01
Findings with evidence, not an opaque score
- 02
Results integrated into Dynatrace
- 03
Benchmarks on detection quality, runtime and cost
- 04
Evaluated on own and independent data
04 / Experience
So far
- 2025
FH Hagenberg
BSc Software Engineering
Education · Since 10/2025
Fifth semester, direct entry into the third.
- 2026
Dynatrace
Software Engineering Intern
Work · 08/2026 – 10/2026
Bachelor's thesis on LLM agent security: a four-layer TypeScript scanner.
- 2020
HTL Leonding
Diploma in Computer Science
Education · 09/2020 – 06/2025
Technical high school, diploma thesis with Avocodo.
- 2024
Avocodo
Software Developer Intern
Work · 07/2024 – 08/2024
Java connector with SAP JCo and a TypeScript generator for Mendix domain models.
- 2023
Avocodo
Software Developer Intern
Work · 07/2023 – 09/2023
C# connector for Azure DevOps tickets, code reviews and team presentations.
Skills
- TypeScript
- React
- Next.js
- React Native
- C#
- ASP.NET Core
- EF Core
- PostgreSQL
- SQLite
- SignalR
- Java
- Python
- Docker
- Kubernetes
- GitHub Actions
- Linux