Software engineeringPython, C++, TypeScript, Flutter, WebRTC, CI/CD
Software Engineer at InfineonTUM M.Sc. candidate
Aziret Ashyrbai uulu / Software & AI Systems Engineer
I engineer software.I make it intelligent.
Software engineer and AI engineer building typed backends, cross-platform products, reliable automation, and inspectable machine-learning systems from prototype through production.
01 Architecture → delivery02 Models → production03 Evidence → trust
At a glanceSignal / 2026
AI engineeringLLMs, RAG, vision, on-device ML, inference systems
Production rangeIndustrial automation, mobile/web, simulation, cloud
Selected recognitionHACK:XR+AI 2025 — Best Idea
/01
Selected systems
Software and intelligence, engineered together.
Seven evidence portals across software architecture, AI, products, infrastructure, and impact. Use the role lens to inspect the work most relevant to you.
Architecture labPublic code · 2 repositories
Language models, from first principles
Two inspectable PyTorch implementations turn attention, positional structure, latent projections, and expert routing into code you can trace.
Open system note
- Problem
- Modern language models are easy to call and difficult to truly understand. The work makes their internal mechanics concrete enough to inspect, test, and explain.
- Ownership
- Implemented a compact GPT-style transformer with multi-head self-attention, sinusoidal positions, and weight tying, plus a DeepSeek-inspired educational architecture with latent attention, decoupled RoPE, and sparse top-k experts.
- Decision
- Small datasets and explicit modules keep the architecture legible. The projects favour learning and reproducibility over claims of frontier-scale training.
- Next iteration
- Add component-level tests, ablation notes, and side-by-side memory measurements for the latent-attention path.
Field constraints · GISPublic code · field systems
Sustainable pastureland systems
A cross-platform field system connecting GIS, WebSocket telemetry, WebRTC communication, and resource decisions for pasture management.
Open system note
- Problem
- Useful field software must continue to support decisions when connectivity, device capability, and data freshness are uneven.
- Ownership
- Connected the economic and sustainability context with GIS algorithms, live telemetry, peer-to-peer communication, and a field-oriented product experience.
- Decision
- Use direct peer-to-peer communication and lightweight telemetry paths to keep collaboration useful under limited connectivity.
- Next iteration
- Validate the workflow with field users and publish only media and implementation detail cleared for public disclosure.
Spatial AI · rapid product workHACK:XR+AI 2025 · Best Idea
ThXRd Space
Responsive XR avatars joining a Unity experience, language-model inference, and speech into a memorable human–AI interaction concept.
Open system note
- Problem
- Conversational intelligence often feels detached from space, expression, and the shared context that makes interaction feel natural.
- Ownership
- Helped shape and prototype the end-to-end interaction loop under hackathon constraints, from spoken input to model response and avatar delivery.
- Decision
- Optimise the concept around one clear experiential loop instead of spreading effort across a broad virtual world.
- Next iteration
- Package the approved demo footage, team context, and award evidence into a short cinematic case study.
No public source link is shown. Publication stays limited to approved narrative and evidence.
Production systemsEmployer-safe summary
Industrial AI & engineering automation
A confidentiality-safe view of typed Python automation, RAG-assisted specification work, and CAD infrastructure built for engineering environments.
Open system note
- Problem
- Engineering teams need reliable paths through dense specifications, compliance material, data-processing work, and release infrastructure.
- Ownership
- Built and supported Python automation and AI data-processing workflows across vehicle-development and semiconductor contexts.
- Decision
- Keep public evidence at the level of methods and responsibilities until architectures, metrics, and internal names receive explicit permission.
- Next iteration
- Replace the summary with a measured, approved case study once evidence and disclosure reviews are complete.
No public source link is shown. Publication stays limited to approved narrative and evidence.
Applied ML studiesPublic code · reproducible notebooks
Vision & diffusion notebook lab
Compact DETR and DDPM studies make object queries, bipartite-style detection, noise schedules, and reverse diffusion tangible.
Open system note
- Problem
- Research papers become more useful when the central mechanism can be run, visualised, and modified in a small environment.
- Ownership
- Built a PyTorch DETR implementation with custom inference and bounding-box visualisation, plus a DDPM with a custom U-Net, ELBO objective, and linear/cosine noise schedules.
- Decision
- Use focused datasets and visible intermediate outputs so visitors can reason about behaviour without specialised infrastructure.
- Next iteration
- Turn the notebooks into interactive explainers with fixed examples, accessible annotations, and documented experiment seeds.
Mobile · web · deliveryProduction experience · employer-safe
Cross-platform, real-time product engineering
A disclosure-safe engineering case spanning Flutter delivery, WebRTC video, payments, state architecture, on-device ML, and release optimisation.
Open system note
- Problem
- A feature is only useful when it behaves consistently across platforms, survives imperfect networks, and remains understandable as the product grows.
- Ownership
- Delivered product logic and user-facing features across iOS, Android, and web, including peer-to-peer video, payments, state management, performance, and release work.
- Decision
- Separate domain behaviour from framework details, keep state transitions explicit, and treat runtime, memory, and release health as product requirements.
- Next iteration
- Publish an architecture walkthrough when product-specific details and measurable outcomes are approved for disclosure.
No public source link is shown. Publication stays limited to approved narrative and evidence.
C++ · Python · AzureProduction experience · employer-safe
Simulation-to-cloud engineering systems
Middleware and telemetry work connecting vehicle simulation, computer vision, IoT data, cloud services, and repeatable delivery pipelines.
Open system note
- Problem
- Mixed simulation and telemetry systems fail at their boundaries: data contracts, timing, observability, and deployment ownership.
- Ownership
- Worked across C++ and Python middleware, anomaly detection, IoT-to-Azure telemetry, and CI/CD support in an applied engineering environment.
- Decision
- Make interfaces typed and observable, isolate failure domains, and automate the delivery path so experiments can become repeatable systems.
- Next iteration
- Add approved diagrams, test strategy, and operational evidence without exposing client or employer-confidential implementation details.
No public source link is shown. Publication stays limited to approved narrative and evidence.
/02
Experience trajectory
Production software → applied AI → intelligent infrastructure.
Official software-engineering roles and disclosure-safe responsibilities, arranged as evidence rather than a stack of logos.
Published from the owner-controlled evidence studio
01
Jun 2026 — present
Software Engineer
Infineon Technologies AGMunich, GermanyPython applications and automation for CAD and chip-design infrastructure, AI data-processing pipelines, releases, documentation, and Linux support.
Inspect responsibilities04
- Build Python tools around engineering and CAD workflows.
- Develop AI-driven data-processing pipelines for repetitive engineering tasks.
- Automate software releases and maintain technical documentation in Linux environments.
- Resolve user issues and functional requests for cross-functional hardware-design teams.
02
Jan 2026 — May 2026
Software Engineer
HOLONMunich, GermanyLLM and RAG pipelines integrated with CodeBeamer for technical-specification and compliance retrieval, supported by typed and tested Python automation.
Inspect responsibilities04
- Integrated LLM workflows with CodeBeamer for requirements analysis.
- Applied retrieval-augmented generation to technical specifications and compliance documents.
- Built Python automation for specification tracing across the vehicle-development lifecycle.
- Structured mission-critical components with type hints and unit tests.
03
Jun 2025 — Oct 2025
Software Engineer
Froxy LabsRemote · EnglandFlutter delivery across iOS, Android, and web, including on-device ML, runtime and memory optimisation, payments, and release work.
Inspect responsibilities04
- Designed Flutter architecture across iOS, Android, and web targets.
- Integrated on-device inference for product personalisation.
- Profiled memory, algorithmic complexity, and rendering latency.
- Separated business logic with Clean Architecture and delivered through mobile-store and container CI/CD pipelines.
04
Mar 2024 — Nov 2024
Software Engineer
Entwicklerburg LLCBishkek, KyrgyzstanC++ and Python simulation middleware, NeRF work, drone anomaly detection, IoT-to-Azure telemetry, and CI/CD support.
Inspect responsibilities04
- Developed C++ and Python middleware for simulation workflows.
- Worked on fault-tolerant IPC, NeRF-based reconstruction, and volumetric rendering.
- Built CNN-based anomaly detection for high-resolution aerial imagery.
- Connected asynchronous IoT telemetry to Azure and supported repeatable CI/CD delivery.
05
Jun 2023 — Mar 2024
Flutter Developer
Move MatesBishkek, KyrgyzstanFitness-product business logic and state management, WebRTC peer-to-peer video, SOLID design, and dependency injection.
Inspect responsibilities04
- Built a reactive cross-platform fitness experience with explicit state and business logic.
- Integrated real-time peer-to-peer video with WebRTC.
- Optimised bandwidth use and connection stability.
- Applied SOLID boundaries and dependency injection for maintainability and testing.
Parallel education lanes
Mathematics, economics, and responsible AI.
Concurrent programs are labelled explicitly so the chronology stays accurate.
Technical University of Munich
M.Sc. candidate · Artificial Intelligence in Society
Munich, GermanyCurrent interdisciplinary study connecting technical AI systems with bias, governance, and responsible societal deployment; expected completion August 2027.
Current · expected Aug 2027Focus on AI-bias mitigation and effective GenAI/ML implementation.MITx
MicroMasters Program · Statistics and Data Science
OnlineGraduate-level probability, statistics, machine learning, and applied data analysis.
Completed program period · credential record availableStudied concurrently with the TUM master’s program.Verify MITx recordAmerican University of Central Asia
M.Sc. Economics · GPA 3.77
Bishkek, KyrgyzstanEconomics and quantitative analysis with AI/software applications in sustainable economic modelling.
Completed · GPA 3.77Connected engineering methods with economics and sustainability.Humboldt-Universität zu Berlin
Integrated Natural Resource Management · exchange period
Berlin, GermanyExchange period within the M.Sc. program, including a GIS software project for responsible pasture utilisation.
Erasmus Mundus SM+ exchange · GPA 1.7 (German scale)Exchange period—not represented as a separately completed degree.Kyrgyz National University
B.Sc. Applied Mathematics & Informatics · GPA 3.87
Bishkek, KyrgyzstanApplied mathematics, informatics, and mathematical modelling for real-world computational and AI problems.
Graduated with honors · GPA 3.87Built the mathematical foundation for software and AI engineering./03
Evidence map
Capabilities linked to where they were used.
No self-assigned percentages. Each cluster points back to a project, role, or field of study.
Software architecture
Python · C++ · Dart · C# · SQL · TypeScript · SOLID
Infineon, HOLON, Entwicklerburg, public repositoriesProduct delivery
Flutter · WebRTC · REST · payments · release systems
Froxy, Move Mates, ThXRd SpaceModel architecture
PyTorch · TensorFlow · Transformers · vLLM · RAG · LoRA / QLoRA
Transformer, DeepSeek-inspired, HOLONPerception
Computer vision · DETR · DDPM · NeRF · OpenCV · on-device ML
DETR, Entwicklerburg, FroxyInfrastructure
Linux · CUDA · Docker · Kubernetes · Azure · Bash · CI/CD
Infineon, HOLON, EntwicklerburgApplied context
NumPy · Pandas · scikit-learn · SciPy · statistics · GIS
TUM, MITx, AUCA, pastureland systems/04
Interactive engineering lab
Don’t just read the work. Run it.
Real-time model internals, an original systems game, and a responsibly integrated open-source arcade—all isolated from the fast recruiter path.
01 / LIVE MODEL
Inside a convolutional network
Draw a digit. Watch its real feature maps, pooling, dense activations, and class scores transform live.
Enter neural path 02 / OPEN SOURCEFour games. Near-zero homepage cost.
A legal Doom-engine campaign, a tiny FPS, a canvas shooter, and a touch puzzle—loaded only when you choose one.
Open arcadeHuman layer
Intelligent systems matter when they survive contact withpeople, constraints, and consequences.
I am a software engineer and an AI engineer. My path runs through applied mathematics, production product engineering, economics, sustainability, and AI in society. That range gives me a practical question for every technical choice: what must be true for this to work in the world?
Make it legible
Inspectable architecture beats unexplained magic.
Measure honestly
Evidence and caveats belong next to the claim.
Design for reality
Latency, access, trust, and maintenance are product features.
Open channel / public profile
Let's build something
worth trusting.
For engineering conversations around production software, applied AI, reliable automation, research systems, or ambitious product work, email me directly or connect through LinkedIn and GitHub.