Kaan YILDIZ — Portfolio

01 / 07|LIVE SHOWCASE

Custom Unreal Engine Plugins

7 plugins

LevelInstanceDeepCopy GitHub repository preview

DeepLevelDesignPCG GitHub repository preview

ProcessRuntime GitHub repository preview

TextureGeneration GitHub repository preview

BackupFileBrowser GitHub repository preview

01 / 03

--:--

A Neo-Noir Detective Experience

Players can speak freely with NPCs in natural language as they investigate cases.

The TextGen plugin uses an LLM to generate replies that fit the scenario and gameplay flow; the SpeechGen plugin produces the voice and synchronized mouth and facial animation at runtime.

ConvCore manages the full process. The AI dialogue pipeline is complete; asset placement and scenario implementation remain.

Atmospheric InvestigationRain-slicked streets, crime scene investigation, and suspect interrogation.

Dynamic NPC DialoguesFreeform questioning and evidence confrontation replacing static branching trees.

Real-Time AI NPC Dialogue

The TextGen plugin uses an LLM to generate scenario-aware replies. The SpeechGen plugin turns each reply into voice and synchronized mouth and facial animation.

ConvCore manages the full pipeline and keeps its output aligned with the scenario and gameplay.

From Text to Live NPC PerformanceTextGen generates the reply; SpeechGen produces its voice and synchronized mouth and facial animation at runtime.

Scenario and Gameplay OrchestrationThe custom ConvCore module manages requests, replies, and gameplay effects while keeping them aligned with the scenario.

Custom Kokoro-82M CPU QuantizationA custom, quality-gated U8/S8 mixed-precision quantization for SpeechGen; 3.00× real-time synthesis on a mid-range CPU after model load.

DeepLevelDesignPCG — Procedural City & Silhouette Closure Solver

A placement pipeline for positioning buildings along UE 5.8 road networks and closed island splines. Elastic tension and guarded seam algorithms bring the layout together.

Roadside & Two-Stage CompactionFollowing initial placement, peripheral elastic tension and SafeStep orthogonal pressure waves pull buildings into physical contacts.

Guarded Seam HandlingConcave corners and protected blocks are used to close remaining seams while avoiding overlaps and gaps.

Engine Architecture with 7 Custom Plugins

Custom open-source Unreal Engine plugins built to handle procedural level design, AI text/speech integration, and subprocess orchestration.

LevelInstanceDeepCopyUnreal Editor plugin for copying Level Instance hierarchies independently of their source.

DeepLevelDesignPCGPCG plugin for placing buildings along spline-based road networks in UE 5.8.

ProcessRuntimeManages external processes and inter-process communication from Unreal Engine.

TextGenBrings local LLM inference with CUDA and Vulkan support to Unreal Engine.

TextureGenerationDeterministic texture generation and import tools for Unreal Editor.

BackupFileBrowserIn-editor project asset backup and recovery utility.

SpeechGenOffline speech synthesis plugin powered by Kokoro and ONNX Runtime.

02 / 07|LIVE SHOWCASE

http://192.168.4.1/web/ LittleFS flash partition

K4MPUS_R00TK1T

 /$$   /$$ /$$   /$$ /$$      /$$ /$$$$$$$  /$$   /$$  /$$$$$$        /$$$$$$$   /$$$$$$   /$$$$$$  /$$   /$$   /$$ /$$$$$$$$
| $$  /$$/| $$  | $$| $$$    /$$$| $$__  $$| $$  | $$ /$$__  $$      | $$__  $$ /$$$_  $$ /$$$_  $$| $$  /$$/ /$$$$|__  $$__/
| $$ /$$/ | $$  | $$| $$$$  /$$$$| $$  \ $$| $$  | $$| $$  \__/      | $$  \ $$| $$$$\ $$| $$$$\ $$| $$ /$$/ |_  $$   | $$   
| $$$$$/  | $$$$$$$$| $$ $$/$$ $$| $$$$$$$/| $$  | $$|  $$$$$$       | $$$$$$$/| $$ $$ $$| $$ $$ $$| $$$$$/    | $$   | $$   
| $$  $$  |_____  $$| $$  $$$| $$| $$____/ | $$  | $$ \____  $$      | $$__  $$| $$\ $$$$| $$\ $$$$| $$  $$    | $$   | $$   
| $$\  $$       | $$| $$\  $ | $$| $$      | $$  | $$ /$$  \ $$      | $$  \ $$| $$ \ $$$| $$ \ $$$| $$\  $$   | $$   | $$   
| $$ \  $$      | $$| $$ \/  | $$| $$      |  $$$$$$/|  $$$$$$/      | $$  | $$|  $$$$$$/|  $$$$$$/| $$ \  $$ /$$$$$$ | $$   
|__/  \__/      |__/|__/     |__/|__/       \______/  \______//$$$$$$|__/  |__/ \______/  \______/ |__/  \__/|______/ |__/   
                                                             |______/

Cihaz

Operating Mode

DUAL-STA (STA+BLE)

Device Name

TRAKYA UNI KAMPUS4.0

STA SSID

Kampus_Wifi_Guest

STA Password

••••••••••••

AP SSID

K4MPUS_ROOTKIT_AP

Company ID (HEX)

0047

0047 - Trakya Üniversitesi

Çakışmalı ID'ler hariç tutuldu.

▼ LED Profilleri

GenelWiFi Durumları

WiFi Mode

blink

500 ms

BLE Mode

blink

500 ms

Dual Mode

blink

500 ms

Dual-STA

pulse

250 ms

Zaman

Cihaz Saatini Ayarla

2026-09-23T09:30

Cihaz Saati: 09:30:15 (UTC+3)

Tarayıcı saati alınır ve cihaza gönderilir. Cihaz Türkiye saati (UTC+3) gözetir.

Ders Programı

Gün	Saat	Asıl Başlama	Ders
Pazartesi	09:30 - 12:20	09:30	D103 (BLE Aktif)
Pazartesi	13:30 - 16:20	13:30	Mobil Uygulama
Salı	09:30 - 12:20	10:00	Amfi Derslik 2

RP2040 DONANIM VE ÇİP MİMARİSİSYS CLK: 133 MHzSRAM: 264 KB

01 / 03

--:--

K4MPUS_R00TK1T: Pico W Web Server (LittleFS)

A hacking tool for testing campus Bluetooth verification systems. It uses fake beacon signals to bypass filters and aims to detect zero-day vulnerabilities.

It runs on Raspberry Pi Pico W (RP2040) and ESP32, tests Bluetooth beacon filtering methods, and includes a LittleFS-based web server. Use only in authorized security tests.

Flash-Hosted Web UIServes static HTML/CSS/JS assets from LittleFS flash storage.

Real-Time Device TelemetryLive uptime tracking, operational mode switching (AP/STA/DUAL), and async REST endpoints.

BTstack BLE Packet Generation & Company ID Emulation

Synthesizes raw BLE advertisement packets using BTstack to test the beacon verification logic of attendance and tracking systems.

Dynamic Company ID SpoofingInjects target manufacturer HEX IDs on-the-fly to test receiver filter reliability.

Task SchedulerAutomated advertisement cycles timed via schedule.txt and NTP clock synchronization.

RP2040 Dual-Core Embedded C++ Architecture

Task distribution on dual-core RP2040: Core 0 handles web server and network routing, while Core 1 executes the real-time BLE transmission loop.

Modular State MachineLoose coupling between WiFiManager, BLEManager, and ModeController layers.

Visual State AnimationConfigurable blink and pulse millisecond intervals for radio state telemetry.

03 / 07|LIVE SHOWCASE

http://localhost:8000/#inference VocaRig Lab — Audio-Driven Lip Sync

VOCARIG LAB

Live Scene

3D Viewport

Drag: OrbitWheel: Zoom

FPS60

DELTA16.7ms

INFERENCE1.84ms

ENERGY-14.2 dB

1 / Audio & Motor

WAV Filevoice-sample.wav

BAKEDLIVE WAVMIC

Signal and Blendshape

Lip Output

STREAMING

Baked ModeSynchronized (00:02.40 / 00:04.20)VISEME: [AA]

jawOpen

0.72

mouthFunnel

0.38

mouthSmileLeft

0.45

mouthSmileRight

0.45

mouthPucker

0.22

mouthUpperUp

0.28

http://localhost:8000/#pipeline VocaRig Lab — Synthetic Dataset & Training Run

VOCARIG PIPELINE

1 / Data

Synthetic Dataset

SYNTHESIZED

Utterances4000

Seed42

Min Phonemes18

Max Phonemes56

Silence Prob0.18

TR Mix0.55

2 / Training

Training Run

RUNNING (182/200)

TRAINSTOPEXPORT ONNX

EPOCH182 / 200

TRAIN LOSS0.0040

VAL LOSS0.0058

ROLLOUT0.0041

Base Settings

Datasetvocarig_synth_4k

Batch Size64

Learning Rate0.00035

PrecisionFP32 (CUDA)

3 / Monitor

Live Loss

ONNX CAUSAL

LOGLIN

Train (Single Step)Train RolloutVal Loss

[LipSyncGRU] Step 182/200 - loss: 0.0040 - val_loss: 0.0058

[LipSyncGRU] Step 154 - checkpoint saved to runs/vocarig_lipsync_gru_best.pt

[Rollout] Free-run rollout drift: 0.0041 (stability PASS)

http://localhost:8000/#analytics VocaRig Lab — Model Health & Metric History

VOCARIG ANALYTICS

1 / Architecture & Flow

LipSyncGRU Network Flow

Model: LipSyncGRU-v2

TOTAL EPOCHS200

BEST VAL LOSS0.00559

INFERENCE LATENCY1.84ms

ARKIT CHANNELS21 Lips

2 / Health & Gates

Model Health

ALL CHECKS PASS

✓

Pose & Delta Loss (0.00425)Geometry & difference loss clamped. Natural mouth aperture tracking without over-stretching.

PASS

✓

Velocity & Jerk Damping (0.00329)Dynamic acceleration filter prevents abrupt mouth snaps on plosives (P, B, M).

PASS

✓

Silence Closure Clamp (0.00007)Zero-drift mouth closure verified during inter-sentence acoustic pauses.

PASS

✓

Zero-Lookahead Causality (<2ms CPU)Conv1D receptive field + GRUCell preserves strict causality with 0 future frame leakage.

PASS

01 / 03

--:--

Real-Time Lip Animation Purely from Audio

A deep learning system transforming raw speech audio directly into 3D facial blendshapes without requiring cameras or facial markers.

Hardware-FreeOperates directly with standard microphone audio.

Low LatencyProcesses live audio streams with minimal latency.

LipSyncGRU: 21 Focused Channels to 52 ARKit Targets

Recurrent GRU architecture analyzing acoustic features with causal context to produce smooth, natural mouth movements.

21 Lip ChannelsFocuses specifically on speech-critical mouth and jaw blendshapes.

ONNX ExportReady for runtime deployment across game engines and web applications.

PyTorch, ONNX & WebGL/Three.js Integration

A complete pipeline spanning offline neural network training to interactive real-time 3D browser preview.

Open SourceFull training scripts, datasets, and web preview available on GitHub.

ARKit StandardStandard 52 blendshape format compatible with UE, Unity, and Blender.

04 / 07|LIVE SHOWCASE

http://127.0.0.1:8000/#inferenceFaceRig Lab — Neo-Noir Edition

FACERIG LAB

01 / Input

Emotion Vector

Dual Blend

Joy

0.78

Surprise

0.24

Sadness

0.00

Anger

0.00

Style Alpha0.50

Delta Time0.033s (30 FPS)

02 / Output

Blendshape Output

GRU Causal

mouthSmileLeft

0.55

mouthSmileRight

0.55

mouthFrownLeft

0.35

browInnerUp

0.70

cheekPuff

0.32

jawOpen

0.14

eyeSquintLeft

0.30

INFERENCE: 1.62msJITTER: 0.12ms

Live Scene

3D Viewport

Drag: OrbitWheel: Zoom

FPS60

DELTA16.7ms

STATUSTRACKING

http://127.0.0.1:8000/#trainingFaceRig Lab — Pipeline & Retraining

FACERIG LAB

1 / Setup

Run Kurulumu

Datasetsynthetic_face_rig.npz

Target Epochs3000

Batch Size256

Learning Rate0.0005 (AdamW)

Auto-StopVal Plateau (Patience: 30)

2 / Monitor

Live Loss & Convergence

LOGLIN

StatusRUNNING

Epoch2140 / 3000

Train Loss0.0028

Val Loss0.0039

Train (Single Step)Train RolloutVal Loss

[FaceRigGRU] Step 2140 - loss: 0.0028 - val_loss: 0.0039

[Checkpoint] Saved best weights to models/facerig_mlp_best.pt

[AutoStop] Divergence test PASSED (margin: 0.0011)

http://127.0.0.1:8000/#analyticsFaceRig Lab — Model Health & Evaluation Dossier

FACERIG LAB

Best Val Loss0.0039MSE @ Epoch 2140

Gen/Val Gap0.0011Zero Overfitting Drift

Inference Latency1.62 msFP32 ONNX Engine

Jitter Variance0.0008Velocity Clamping Pass

Architecture

Causal GRU Topology

PyTorch / ONNX

INPUT TENSOR4D Emotion Weights (Joy, Surprise, Sadness, Anger)

RECURRENT CORE128-dim Causal GRU + LayerNorm & Dropout

OUTPUT DECODER52 ARKit Blendshape Weights [0.0, 1.0]

LOSS CRITERIONWeighted MSE + Velocity Derivative Damping (0.05)

Validation

Evaluation Dossier

4/4 PASSED

PASS

Causal Streaming Invariance

Sıfır gelecek pencere (zero future lookahead); her kare bağımsız veya sıralı akışta deterministik çıktılar üretir.

PASS

Max-Acceleration Velocity Clamping

Ani duygu geçişlerinde ARKit katsayılarının mikro-titremeleri (jitter) ivme sönümleme filtresi ile engellenir.

PASS

ONNX Runtime & WebGL Compatibility

FP32 ve FP16 modelleri hem masaüstü CPU hem WebGL browser worker üzerinde <2ms ile tam doğrulanmıştır.

PASS

Canonical ARKit 52 Clamp Integrity

52 blendshape katsayısının tamamı [0.0, 1.0] aralığında kilitli kalır; poligon dezenformasyonu sıfırdır.

01 / 03

--:--

3D Faces Driven by Continuous Emotion Parameters

A machine learning workspace mapping continuous emotion parameters to facial blendshape weights for dynamic NPC expressions.

Emotion BlendingContinuous blending across varied emotion coordinates.

Synthetic DatasetAutomated synthetic data generation for model training.

Web-Based Model Training & Real-Time 3D Viewport

A FastAPI dashboard offering one-click training, live loss visualization, and immediate ONNX compilation for game engines.

Live Training StatusGraphical tracking of epoch progression and validation loss.

3D Face PreviewInstant visual feedback of parameter adjustments on a 3D avatar.

Modular Python Package, PyTorch GRU & Test Coverage

The PyTorch GRU uses emotion input, the previous face state, and timing data to predict changes across 24 blendshape values.

GRU Temporal ModelCarries the previous face state and hidden state to produce controlled transitions between frames.

Modular StructureClean separation between training, inference, and UI layers.

05 / 07|LIVE SHOWCASE

UE 5.7

Unreal Editor — SoC.uproject (5.7.4-Development_Editor)Plugin: ModelContextProtocolPort

EULA AUDIT: PASS • SOURCE CLEAN

01DISCOVERY

02SPARSE FETCH

03BACKPORT PATCH

04UBT BUILD

05HEALTH PROBE

Output Log — cmd: python mcp_port_kit.py installUnrealBuildTool Win64

Cmd:python mcp_port_kit.py install --project "Q:\Projects\Unreal Engine\SoC\SoC.uproject"

[00:00.12] LogPortKit: Display: Target project discovered: SoC (UE 5.7.4-release)

[00:01.48] LogPortKit: Display: Sparse checkout: github.com/EpicGames/UnrealEngine@7deeb41 (EULA verified)

[00:02.85] LogPortKit: Display: Applying 0001-generated-ue57-port.patch (42 toolsets, 0 conflicts)

[00:05.10] LogUBT: Display: UnrealBuildTool.exe Win64 Development Editor: 0 errors, 0 warnings (14.2s)

[00:07.40] LogMCP: Display: Schema probe: 423 / 423 tools registered. Status: PASS

>LogMCP: Daemon listening on port 3001 (Ready)█

Project Settings — MCP Toolset SubsystemsPHASE: 01 DISCOVERY

SOURCE ENGINEUE 5.8-mainEpic Games Git Ref

TARGET HOSTUE 5.7.4SoC.uproject Win64

LOADED TOOLSETS42 / 42423 Total Functions

EULA COMPLIANCE100% PASSZero Source Vendored

BACKPORT SUBSYSTEM STATUS100% PATCHED

Editor & Asset Tools251 / 251

Atomic Material Graph DSL28 / 28

PCG & Level Subsystems62 / 62

UE 5.7

Material Editor — /Game/Materials/M_CyberAlloyDirectX 12 (SM6)

AST PREFLIGHT & ROLLBACK ARMED

ATOMIC DSL PAYLOAD (M_CyberAlloy.uasset)JSON • 18ms COMPILED

{
  "schema_version": "1.0",
  "mode": "replace",
  "material": "/Game/Materials/M_CyberAlloy",
  "nodes": [
    {
      "id": "tex_diff",
      "type": "TextureSample",
      "texture": "/Game/Textures/T_Alloy_D"
    },
    {
      "id": "p_tint",
      "type": "VectorParameter",
      "value": [0.05, 0.65, 0.95, 1.0]
    },
    {
      "id": "mul_tint",
      "type": "Multiply",
      "inputs": { "A": "tex_diff.RGB", "B": "p_tint.RGB" }
    },
    {
      "id": "p_rough",
      "type": "ScalarParameter",
      "value": 0.28
    },
    {
      "id": "tex_norm",
      "type": "TextureSample",
      "texture": "/Game/Textures/T_Alloy_N"
    }
  ],
  "outputs": {
    "BaseColor": "mul_tint.Output",
    "Metallic": 0.92,
    "Roughness": "p_rough.Output",
    "Normal": "tex_norm.RGB"
  }
}

Graph Canvas — M_CyberAlloy (Preview: PBR Metallic)142 INSTRUCTIONS • 0 COMPILER ERRORS

UMaterialCompileDiagnosticsBridge: COMPILED (SM6 142 inst)

Latency: 18.4ms

Status: COMMITTED

UE 5.7

Plugins Browser — MCP Toolset Subsystem & RPC Daemon423 TOOLS • 42 DOMAINS ACTIVE

UNREAL EDITOR SUBSYSTEM: READY

TOTAL TOOLS423

REGISTERED TOOLSETS42

MATERIAL TOOLS28 Custom

RPC LATENCY< 1.4ms

EULA STATUSPASS (0 Leaks)

ACTIVE MCP TOOLSET DOMAINS (42 Domains)423 TOOLS

PORTKIT EXTENSIONMaterialTools28 Tools

Atomic Material Graph DSL parser, apply_graph_dsl, read_graph_dsl, AST validation, rollback safety & comment injection.

apply_graph_dslread_graph_dsladd_commentdiagnostics

ENGINE COREEditorToolset251 Tools

Actor transforms, component spawning, level management, asset registry queries, selection sets & viewport cameras.

spawn_actorquery_assetsset_transformopen_asset

PROCEDURALPCGToolset31 Tools

Procedural Content Generation graph construction, attribute sets, spline samplers & spatial queries.

create_pcg_nodeconnect_pinsexecute_graph

GAMEPLAY & AIGASToolset & AI28 Tools

Gameplay Ability System attributes, effect definitions, behavior trees & blackboard keys.

add_attributebind_abilitybt_inspect

LIVE AGENT JSON-RPC BRIDGESESSION: CONNECTED

AI AGENT12:44:02.108call_tool("MaterialTools_apply_graph_dsl")

{ "material": "/Game/Materials/M_CyberAlloy", "mode": "replace", "nodes": 5 }

MCP PORTKIT DAEMON12:44:02.126 (+18ms)200 OK

{ "status": "success", "nodes_created": 5, "compilation": "SM6 OK (142 inst)", "rollback_snapshot": "snap_0x4f12" }

✓

License & EULA Guard

Epic Games GitHub sparse checkout; zero engine C++ code redistributed.

✓

AST Preflight & Auto-Rollback

Transients compile-checked prior to commit; zero corrupted assets in /Game.

✓

UE 5.7 Backward Compatibility Bridge

All UE 5.8 Material & Subsystem API differences automatically patched.

01 / 03

--:--

Backporting UE 5.8 MCP Toolsets to UE 5.7 Projects

An automated utility and patch kit for using Unreal Engine 5.8 Model Context Protocol (MCP) toolsets in UE 5.7 projects.

License CompliantDoes not vendor UE source code; sparse-fetches directly from licensed repositories.

Automated PatchingApplies compatibility patches and compiles the target plugins automatically.

Atomic Material Graph DSL for AI Coding Agents

Enables AI agents to read, validate, and atomically mutate complex material graphs in a single call with automated rollback safety.

Single-Call OperationsFull graph reads and multi-node mutations complete in a single tool call.

Automated RollbackReverts invalid graph mutations automatically if compilation fails.

Extended Python Material Surface & Console Bridge

Enhanced project extensions including graph comments, wildcard asset filtering, and execution of editor console commands.

Custom Material ToolsFunctions for reading, validating, and mutating material graph DSL.

Wildcard SearchMatches both substring text and wildcard patterns across asset registries.

06 / 07|LIVE SHOWCASE

http://localhost:3000/chat

Chimera AI Platform

EmirGPTEmirGPT, yüksek düzeyde teknik bilgiye sahip, hızlı...

💬

EmirGPT

Merhaba, ben EmirGPT! Yüksek teknik bilgimle mühendislik, yazılım geliştirme ve donanım optimizasyonunda hızlı analizler yaparım. 🔧

Başlangıçta RISC-V mimarisi için geliştirilen bir sistemdim, şimdi AI hızlandırıcıları ve performans iyileştirmelerinde uzmanlaştım. 🌟

Sana net, verimli çözümler sunabilirim; örneğin, kod optimizasyonu veya sistem tasarımı konularında yardımcı olurum. ⚡

Eğer bir sorununu paylaşırsan, hemen optimize edilmiş bir yanıt veririm! 😊

❐

give me a question in English.~

✎❐

... to test your knowledge: What are the key advantages of using RISC-V over proprietary ISAs when designing hardware accelerators for LLM matrix multiplication? 🧐

Looking forward to an optimized analysis! ⚡

❐

http://localhost:3000/canvas

Chimera AI Platform

http://localhost:8080/api/telemetry

Chimera Architecture & Telemetry

Model Kataloğu322 Aktif ModelOpenRouter sağlayıcı havuzu

Ağ Gecikmesi< 1.8 msGo 1.23 Gin HTTP motoru

Akış Hızı (SSE)64.8 tok/sGerçek zamanlı token aktarımı

YetkilendirmeFirebase JWTBearer token korumalı uçlar

SİSTEM MİMARİSİ KATMANLARI4 BİLEŞEN

01Next.js 14 App RouterReact 18 • Radix UI

Zustand durum yönetimi, tam ekran görsel kanvası ve sanallaştırılmış mesajlaşma arayüzü.

02Go 1.23 Gin EnginePort :8080 • REST & SSE

Goroutine iş parçacıklarıyla yönetilen yüksek eşzamanlı ve sıfır sızıntılı non-blocking API ağ geçidi.

03OpenRouter Service Adapter322+ LLM Yönlendirme

Güncel LLM aileleri için tek noktadan dinamik SSE parça ayrıştırma ve maliyet optimizasyonu.

04Replicate & FirebaseFLUX.1 • Firestore • Storage

Asenkron difüzyon iş kuyruğu, otomatik görsel kaydı ve bulut sohbet geçmişi senkronizasyonu.

CANLI İSTEK & ÇIKARIM TELEMETRİSİ200 OK • AKIŞ AKTİF

POST/api/chat/streamxAI / grok-3-mini

İlk Token (TTFT)182 ms

Toplam Süre1.12 sn

Token Hacmi342 tokens

Tahmini Maliyet$0.00010

Yanıt Akışı Çıktısı (Stream Buffer):

"RISC-V provides modular ISA extensions tailored for tensor and matrix operations. Custom hardware accelerators eliminate proprietary licensing bottlenecks while optimizing inference energy efficiency..."

✓

Firebase JWT Bearer GüvenliğiTüm uç noktalar token doğrulamalı; yetkisiz istekler 401 ile engellenir.

✓

Goroutine Bellek Güvenliğiİstemci bağlantısı koptuğunda context derhal sonlandırılır; sıfır sızıntı.

01 / 03

--:--

322+ LLM Routing & Custom AI Persona Studio

A chat workspace for accessing current LLM families through OpenRouter, with provider filters, token streaming, and custom AI personas.

322+ Models & Provider FiltersFilter Anthropic, xAI, OpenAI, Google models by pricing ($/M) and context window (up to 2M+ tokens).

Custom Personas & Version BranchingTailored AI personas like EmirGPT with custom system prompts, branching message history, and syntax highlighting.

AI Image Generator & Canvas Studio

An image generation studio using the Replicate API, with prompt input, canvas preview, gallery saving, and export features.

Prompt & Generation PipelinePrompt input, generation status, and full-screen image preview.

Gallery & Export SyncPersistent image saving to personal gallery, API key management, and direct image exports.

Go 1.23 Gin & Next.js 14 Distributed Architecture

A service architecture combining Go Gin HTTP/SSE services, OpenRouter model routing, Replicate diffusion jobs, and Firebase authentication.

Microservices & Data PipelineLow-latency data pipeline connecting Next.js client, Go API gateway, and external AI providers.

Live Streaming & SecurityServer-Sent Events (SSE) token streaming, usage tracking, and Firebase JWT authentication.

System and Architecture Reviews

Design decisions, solutions to technical problems, source code, and custom plugins.

Secret of Chronicles

Neo-Noir Detective Experience & Real-Time AI Dialogue

C++ lines: 145K+
commits: 1.2K+

PowerShell / Python lines: 11K+
first-party plugins: 7

Secret of Chronicles (SoC) is a neo-noir detective game built with UE 5.8. TextGen uses an LLM to generate NPC replies that fit the scenario and gameplay flow.

SpeechGen produces voice and synchronized mouth and facial animation at runtime; ConvCore manages the full process.

The AI dialogue pipeline is complete; asset placement and scenario implementation remain. Steam AppID setup is complete.

The game will be available to wishlist on Steam once the vertical slice is finished.

Technical Problems and Constraints

Preventing external LLM responses from changing quest, evidence, or knowledge state without validation, and separating runtime code from editor tools.

Solution

TextGen uses an LLM to generate dialogue that fits the scenario and gameplay flow. SpeechGen produces voice and synchronized mouth and facial animation.

ConvCore manages requests, replies, and gameplay effects. NarrativeDirector coordinates the Quest, Investigation, and Knowledge domains.

Design Decisions

Real-time AI dialogue pipeline: TextGen uses an LLM for scenario-aware replies; SpeechGen produces voice and synchronized mouth/facial animation; ConvCore manages the full flow
NarrativeDirector composition root: separates Quest, Investigation, and Knowledge state
ConvCore action-token validator: checks gameplay messages before state changes
Pure Binding MVVM: FieldNotify data bindings between UMG views and C++ ViewModels
Custom, quality-gated U8/S8 Kokoro-82M quantization for SpeechGen: 3.00× warm synthesis on a mid-range CPU; 1.82 dB median mel-spectral error vs. FP32 (excludes model load, phonemization, queueing, and playback)
7 standalone open-source Unreal Engine plugins (TextGen, ProcessRuntime, SpeechGen, LevelInstanceDeepCopy, etc.)

Custom Unreal Engine Plugins

7 Custom Plugins

LevelInstanceDeepCopy

Unreal Editor plugin for copying Level Instance hierarchies independently of their source.

UsageNested Level Instance hierarchies remain editable independently of the source level.

View GitHub repository ↗

DeepLevelDesignPCG

PCG plugin for placing buildings along spline-based road networks in UE 5.8.

UsagePCG workflows generate road networks and level content inside the editor.

View GitHub repository ↗

ProcessRuntime

Manages external processes and inter-process communication from Unreal Engine.

UsageExternal applications launch from Unreal Engine; their output and IPC messages flow into the project.

View GitHub repository ↗

TextGen

Brings local LLM inference with CUDA and Vulkan support to Unreal Engine.

UsageLocal LLM responses stream into in-game dialogue systems.

View GitHub repository ↗

TextureGeneration

Deterministic texture generation and import tools for Unreal Editor.

UsageGenerated textures enter the editor workflow and remain reusable in the project.

View GitHub repository ↗

BackupFileBrowser

In-editor project asset backup and recovery utility.

UsageEarlier asset versions can be browsed and restored without leaving the editor.

View GitHub repository ↗

SpeechGen

Offline speech synthesis plugin powered by Kokoro and ONNX Runtime.

UsageCharacter dialogue is synthesized offline with Kokoro and ONNX Runtime.

View GitHub repository ↗

Unreal Engine 5.8Modern C++Gameplay Ability SystemPure Binding MVVMModular PluginsAction Token DSL

K4MPUS_R00TK1T

Bluetooth Security Testing Tool — Raspberry Pi Pico W

View on GitHub

K4MPUS_R00TK1T is a hacking tool for testing campus Bluetooth verification systems. It uses fake beacon signals to bypass filters and aims to detect zero-day vulnerabilities.

It runs on Raspberry Pi Pico W (RP2040) and ESP32, tests Bluetooth beacon filtering methods, and includes a LittleFS-based web server. Use only in authorized security tests.

Technical Problems and Constraints

Serving concurrent HTTP dashboard requests from flash LittleFS while sustaining precise BTstack BLE advertisement bursts without memory exhaustion or CYW43439 radio desynchronization.

Solution

Implemented non-blocking LittleFS stream handlers and separated AP, STA, BLE, and DUAL-STA modes.

BTstack payload handling is separated from HTTP dispatch, with LED status animations for device events.

Design Decisions

Onboard Web Server & Dashboard: Full responsive web interface served from LittleFS flash to adjust Wi-Fi, time offsets, and target IDs wirelessly
BLE Company ID Emulation: Dynamic imitation of enterprise beacon structures and proprietary manufacturer tags for security verification
Multi-Mode Wi-Fi Architecture: Configurable AP, STA, and DUAL-STA networking with automatic NTP time synchronization
Configurable LED State Profiles: Non-blocking millisecond pulse and blink telemetry for wireless radio and error feedback
PlatformIO & Clean C++: Modular, maintainable object-oriented embedded codebase supporting RP2040 and ESP32 targets

Modern C++Raspberry Pi Pico W (RP2040)BTstack (BLE GAP/GATT)LittleFS Flash StorageCYW43439 Wi-Fi/BLE RadioPlatformIO ToolchainEmbedded Web ServerESP32 Support

VocaRig

Audio-Driven Lip-Sync & 3D ARKit Animation

View on GitHub

VocaRig predicts 52 ARKit facial blendshapes for 3D characters directly from speech audio, without cameras or facial markers.

Its model and ONNX Runtime inference pipeline can be used with Unreal Engine, Unity, and web browsers.

Technical Problems and Constraints

Eliminating phoneme latency and preventing rapid blendshape jitter across bilabial plosives without relying on non-causal lookahead frames.

Solution

Engineered a causal LipSyncGRU recurrent network that tracks temporal context and regresses 21 speech-critical mouth/jaw channels.

The outputs expand to 52 ARKit targets through deterministic facial kinematics and are exported to streaming ONNX.

Design Decisions

LipSyncGRU architecture: causal hidden-state propagation for streaming inference without future-frame lookahead
21 Speech Focus Channels: maps speech-related motion to 52 ARKit targets using deterministic facial kinematics
VocaRig Lab & FastAPI: live microphone streaming, validation telemetry, and real-time Three.js 3D viewport
Platform-agnostic streaming ONNX runtime (compatible with Unreal Engine, Unity, and WebGL)

PyTorchONNX RuntimeFastAPILibrosaThree.jsARKit 52

FaceRig

Parametric NPC Face Rig Training Workspace & Web Dashboard

View on GitHub

FaceRig is a machine learning workspace mapping continuous emotion coordinates (valence, arousal, dominance) to 3D facial blendshapes with synthetic data generation and a FastAPI dashboard.

Technical Problems and Constraints

The prohibitive manual cost of authoring keyframed facial poses for hundreds of NPCs and the robotic stiffness occurring between discrete emotion state switches.

Solution

FaceRig uses a PyTorch GRU that takes emotion parameters, the previous face state, and timing data to predict controlled changes across 24 blendshape values.

The current GRU checkpoint was trained on a processed subset of BEAT facial data. The project also provides a FastAPI REST/WebSocket interface and a live Three.js preview.

Design Decisions

Continuous Emotion Regression: fluid, organic facial transitions driven by valence-arousal coordinate vectors
Synthetic Data Engine: automated procedural generation of anatomically bounded blendshape distributions
FastAPI Interactive Dashboard: real-time loss tracking, hyperparameter tuning, and Three.js 3D viewport
Runtime Game Deployment: ONNX export for native inference in the game engine

PythonPyTorchFastAPIONNXThree.jsWebGLPydantic

UE 5.7 MCP PortKit

UE 5.8 MCP Toolsets Backport & Atomic Material Graph DSL

UE 5.7 MCP PortKit is an EULA-compliant patch kit and editor bridge that backports UE 5.8 Model Context Protocol toolsets to UE 5.7.

It introduces an Atomic Material Graph JSON DSL and console execution bridge for AI coding agents.

Technical Problems and Constraints

Enforcing EULA compliance without redistributing engine sources.

Also addressing the agent failure mode where dozens of sequential roundtrips exhaust context budgets and leave half-connected, corrupted material graphs.

Solution

Developed an automated sparse-fetch pipeline targeting licensed Epic repositories.

Paired it with an independent C++ Atomic Material Graph DSL that executes full-graph mutations in a single tool call and rolls back automatically on compilation failure.

Design Decisions

EULA-Compliant Sparse-Fetch: Clones only target plugin trees from licensed Epic repos without redistributing proprietary sources
Atomic Material Graph DSL: Single-call atomic JSON mutation for entire node networks, pins, and expressions
Automated Rollback Safety: Instantly restores prior graph state upon invalid pin binding or shader compilation errors
Extended Editor Console Bridge: Enables AI agents to execute engine console commands and wildcard asset registry queries

C++Unreal Engine 5.7 / 5.8Model Context Protocol (MCP)Atomic JSON DSLPython Editor Scripting

Chimera

Multi-Model AI Chat & Inpainting Diffusion Studio (Next.js 14 & Go)

View on GitHub

Chimera combines a Next.js 14 App Router frontend with Go 1.23 Gin microservices.

It provides filters and real-time token streaming for 322+ LLMs through OpenRouter, plus mask-based inpainting with Replicate FLUX.1 and SDXL.

Technical Problems and Constraints

Handling different API streaming formats across LLM providers and running asynchronous diffusion inpainting jobs without blocking the UI.

Solution

Implemented Go Gin goroutine workers and SSE streaming. The Next.js interface uses a virtualized message list, a Radix UI model selector with filters, and an HTML5 Canvas layer for inpainting masks.

Design Decisions

Multi-LLM routing: Real-time token streaming across current LLM families via OpenRouter
Custom AI Persona Hub: Specialized system instructions, custom avatars (e.g. EmirGPT), and message version branching
Diffusion Canvas & Inpainting Studio: Interactive raster canvas with dynamic brush masking powered by Replicate FLUX.1 & SDXL
Go 1.23 Gin REST & SSE Gateway: HTTP microservices and goroutine-based streaming
Firebase JWT & Hybrid Storage: Token validation, Firestore storage, and local SQLite fallback
Next.js 14 & Tailwind CSS UX: Fluid Radix UI primitives, Framer Motion transitions, and responsive dark mode

Next.js 14Go 1.23 (Gin)TypeScriptOpenRouter APIReplicate (FLUX.1 / SDXL)Firebase Auth (JWT)HTML5 Canvas (Inpainting)Tailwind CSS & Radix UIDocker

Experience & Education

Project experience and education.

Project Development

Secret of Chronicles (SoC) — Independent Project

2025 — PresentIstanbul, Turkey

A neo-noir detective game in development since 2025.

The TextGen plugin uses an LLM for NPC replies aligned with the scenario and gameplay; SpeechGen produces voice and synchronized mouth/facial animation at runtime.

ConvCore manages the pipeline. Asset placement and scenario implementation remain.

Open-source plugins developed for SoC

LevelInstanceDeepCopyEditor plugin providing hierarchical and isolated duplication of Level Instances.View repository ↗DeepLevelDesignPCGLevel design automation plugin utilizing Procedural Content Generation (PCG).View repository ↗ProcessRuntimeRuntime plugin managing external background processes and IPC communication.View repository ↗TextGenText generation plugin abstracting LLM providers with streaming support.View repository ↗TextureGenerationEngine plugin integrating AI-assisted texture and material generation workflows.View repository ↗BackupFileBrowserIn-editor asset backup and file recovery browser utility.View repository ↗SpeechGenLocal speech synthesis and character dialogue generation plugin.View repository ↗

Modular Unreal Engine C++ gameplay systems and custom editor plugins
MVVM-based data flow for gameplay interfaces
Completed real-time AI dialogue, speech, and facial-animation pipeline

C++ lines: 145K+
commits: 1.2K+
PowerShell / Python lines: 11K+
first-party plugins: 7

Unreal EngineC++GASMVVMAI Dialogue

Campus Rootkit

Embedded Systems and Security Research

Independent Project

Campus Rootkit GitHub repository preview

Open GitHub repository

A hacking tool for testing campus Bluetooth verification systems. It uses fake beacon signals to bypass filters and aims to detect zero-day vulnerabilities.

It runs on Raspberry Pi Pico W (RP2040) and ESP32, tests Bluetooth beacon filtering methods, and includes a LittleFS-based web server. Use only in authorized security tests.

C/C++Raspberry Pi Pico WESP32BLEPlatformIO