Applied AI Systems Engineer · Co-Founder @ Interviewlary
I build AI systems that hold up in the real world.
Computer Engineering Student at Istanbul Medipol University & Co-Founder @ Interviewlary.
Engineering production software from architecture through verification — with deep hands-on focus on applied AI products, voice systems, LLM infrastructure, multi-agent pipelines, and anti-fabrication reliability.
Role-specific voice/text AI mock interview practice with real-time scoring and actionable candidate feedback.
Target Role→Voice Mock Session→Score & Action Plan
01 / Flagship Systems
Selected Case Studies
Four substantial systems representing startup co-foundership, production LLM infrastructure, gamified educational platforms, and winning multi-agent competitive engineering.
01 / Primary Startup · Live Platform
Interviewlary — AI Mock Interview Platform
Full-stack AI simulation platform for job seekers. Conducts role-specific voice and text interviews calibrated by seniority, domain, and skills, returning structured rubrics and actionable feedback.
Role & Context CalibrationTarget position, seniority tier, industry domain, custom skills.
02
Voice or Text SimulationReal-time speech-to-text, low-latency audio, dynamic follow-ups.
03
Diagnostic EvaluationGrounded scoring, answer rubrics, and targeted improvement roadmaps.
02 / LLM Infrastructure · Open Source
TokensCache — Semantic Cache & Optimization Layer
High-performance caching and prompt-optimization engine built for multi-provider AI agents. Eliminates repetitive token expenditure through exact SHA-256 caching, semantic vector similarity matching, and real-dollar budget limits.
Exact & Semantic CachingDual-tier cache with cosine similarity vector matching
Dollar Budget ControlsHard financial circuit breakers on LLM spend
MCP Server IntegrationNative Model Context Protocol tool routing
92/92 Passing TestsVitest suite with comprehensive boundary verification
35.5%Verified Token Reduction in Agent A/B Benchmarks
Universal Provider AdaptersSupports OpenAI, Anthropic, Gemini, and DeepSeek interfaces.
02
MCP Tool RoutingExposes cache telemetry and eviction commands directly to AI agents.
03 / AI Education Product · Digital SAT
Examina AI — Gamified Digital SAT Mastery Platform
Interactive Digital SAT learning ecosystem combining a 3D companion mascot (Navi in Three.js), 5-minute mastery levels, and a Socratic AI tutor (server/tutor.ts) providing step-by-step reasoning without giving away direct answers.
Next.js 16 + React 19Full-stack App Router with server-only data isolation
Three.js 3D EngineProcedural robot companion with dynamic mood states
Dual-Mode PersistenceZero-setup local SQLite & cloud Supabase pgvector
Cross-Browser Test SuitePlaywright tests on Chromium, Firefox & WebKit
Real-time fact-verification pipeline developed and demonstrated live in 7 hours during the Medipol META AI Hackathon. Four specialized agents retrieve citations, debate opposing arguments in parallel, and synthesize a source-grounded verdict.
Adversarial ArchitectureAdvocate vs Skeptic debate loops with citation scoring
FastAPI & StreamingLow-latency SSE stream to responsive React interface
7-Hour Team SprintFrom problem definition to live working stage demo
Award WinnerRanked top project for real-time verification rigor
Organized by engineering domain. Filter across AI products, core LLM infrastructure, voice/agent systems, open-source developer tooling, and architectural research.
Open Source · 92 Tests01
TokensCache
Semantic caching and token optimization middleware for multi-provider AI agents. Eliminates redundant LLM calls using exact SHA-256 hash matching and cosine vector similarity, with Model Context Protocol (MCP) tooling.
Verified with 92/92 automated Vitest tests. Demonstrated 35.5% token reduction across repository mock coding-agent A/B scenarios with identical output fidelity.
Gamified Digital SAT preparation platform. Features 5-minute mastery levels on an interactive campaign map, a Three.js 3D companion mascot (Navi), and a Socratic AI tutor (server/tutor.ts) providing step-by-step guidance without giving direct answers.
Built with Next.js 16 + React 19. Backed by 64 curriculum questions, 16 lesson modules, Vitest unit tests, Playwright E2E cross-browser suites (Chromium/Firefox/WebKit), and dual SQLite/Supabase persistence.
Two-sided admissions operating system for international students applying to Turkish foundation universities. Conversational AI intake in English & Arabic extracts academic profiles into structured cards; consultant triage queues verify documents and handle submissions.
Built with Next.js 15.2, React 19, Prisma 6.5, Jose JWT role-based auth (Student, Consultant, Admin), full RTL bilingual localization, and Playwright E2E test suites.
Next.js 15.2Prisma 6.5RBACi18n (AR/EN RTL)Zod
Active Dev · Monorepo04
Jadwalna AI (جدولنا)
Assistant-first Riyadh outing and itinerary planner. Combines natural Saudi Arabic dialect understanding (Najdi/Riyadh language pack) with a zero-hallucination deterministic constraint solver grounded in 1,000+ verified OpenStreetMap places and real opening hours.
Engineered as a monorepo: Expo 54 client (React Native 0.81 with custom Saudi aesthetic), Fastify 5.2 planning API, Next.js 16 public sharing, and @jadwalna/core solver.
Unified academic operating system and multi-course student workspace. Manages 7 semester courses, 12 class/lab/recitation templates, syllabus progress, embedded MS Office lecture previews, YouTube resources, and attached study session notes.
Vite + React 18 + TypeScript with Supabase private storage and protected Vercel API endpoints. Grounded in deterministic study execution, deliberately distinct from daily routine tracking.
React 18ViteTypeScriptSupabaseAcademic OS
Shipped PWA · 106 Tests06
RoutineAI
Student daily execution OS focused on daily routine execution, deterministic scheduling, focus timers, alarms, streaks, and prayer integrations. Built for fast day-to-day execution.
Verified with 106/106 automated tests. Features offline-first Dexie IndexedDB storage, service worker PWA caching, and Supabase cloud synchronization.
Two-layer reliability framework for autonomous AI coding agents. Implements anti-fabrication boundary rules, premise checking, AST symbol verifiers, and uncertainty calibration to prevent hallucinations on complex codebases.
Validated against 24 documented evaluation test traps covering stale documentation, nonexistent APIs, assumption drift, and source grounding.
AI AgentsReliabilityAST VerifierEvaluationOpen Source
Four-agent adversarial fact-checking pipeline. Gathers external citations, pits advocate and skeptic agents against each other in parallel debates, and synthesizes a source-grounded verdict.
Architected, coded, and demonstrated live on stage in 7 hours with a team during the Medipol META AI Hackathon.
Built and pitched a working 4-agent adversarial fact-checking pipeline with a team: agent orchestration, backend, user interface, and live stage demo within 7 hours.
Live Stage Presentation — Medipol META AI HackathonPitching TruthNet: an adversarial 4-agent fact-checking system built in 7 hours.
04 / Background & Approach
Quality, Structure & Verification
Applied AI systems engineer and computer engineering student at Istanbul Medipol University with extensive hands-on experience shipping real products.
Computer Engineering Student & Founder in Istanbul, Türkiye
Quality-Driven Engineering
I care whether the interface, functionality, edge cases, and backend work harmoniously together — not just whether a feature technically runs.
Verification Mindset
I rely on automated test suites, reliability evaluation scenarios, and rigorous reviews before making confident claims about system readiness.
High-Velocity AI Tooling
I leverage modern AI-assisted engineering workflows (Cursor, Codex, Antigravity, OpenCode, MCP) as force multipliers to architect, debug, test, and iterate systems rapidly.
