Ege Hakan Karaağaç
Verified Expert in Engineering
Full-stack AI Developer
Istanbul, Turkey
Toptal member since June 29, 2026
Ege is a full-stack AI engineer who builds the whole product, from retrieval and agents to the React dashboard and the AWS infrastructure it runs on. As Founding Engineer of BrandVox AI, he shipped a multitenant AI SaaS solo to paying customers across 5 services: a FastAPI back end, a forked agent engine, an embeddable widget, and real-time voice. Before that, Ege built back-end systems at Amazon and Dream Games. He treats evals and cost control as first-class, not afterthoughts.
Portfolio
Experience
- Python 3 - 7 years
- Amazon Web Services (AWS) - 5 years
- FastAPI - 4 years
- Artificial Intelligence (AI) - 4 years
- Large Language Models (LLMs) - 3 years
- LangChain - 3 years
- Retrieval-augmented Generation (RAG) - 3 years
- AI Agents - 3 years
Preferred Environment
Claude Code, Python 3, FastAPI, TypeScript, Java, PostgreSQL, Next.js, LangChain, Amazon Web Services (AWS), Retrieval-augmented Generation (RAG)
The most amazing...
...thing I've built solo is BrandVox AI, a multitenant AI SaaS with hybrid retrieval, multi-provider agents, and a real-time voice pipeline serving 30+ customers.
Work Experience
Founding Engineer | CEO
BrandVox AI
- Architected a multitenant AI SaaS solo, from zero to 1,000+ users and 30+ paying customers across 5 services (around 1,557 commits, dominant author), led by an around 90,000-LOC FastAPI back end and a forked Flowise agent engine.
- Built a hybrid retrieval in the forked Flowise engine: Qdrant dense vectors fused with BM25 sparse through a custom Reciprocal Rank Fusion ranker, with a weighted 0.7/0.3 ensemble live in production and RRF switchable.
- Built a multi-provider generation layer (OpenAI, Anthropic, Gemini, Grok) with cascading fallback so generation does not fall over when a provider does, plus provider routing and credit-based per-model cost accounting.
- Shipped a real-time voice pipeline over WebSockets (OpenAI Realtime API, PCM16/24kHz, server VAD, barge-in) with exactly-once turn persistence, plus image and video generation through Fal.ai.
- Built a Qdrant-versus-DynamoDB orphan-reconciliation garbage collector (dry-run gate) to keep vectors consistent after distributed deletes, with agentic tool-calling, structured outputs, and prompt-injection handling.
- Ran the platform solo on AWS ECS Fargate behind an ALB (39 DynamoDB tables, Redis/MemoryDB, Elasticsearch for chat/lead search, S3, Cloudflare), with Terraform, GitHub Actions, and Sentry, Grafana, PostHog, and CloudWatch.
Back-end Engineer
Dream Games
- Built live in-game event services for a top-grossing mobile game with 6+ million active players, using Spring Boot and gRPC over MySQL, Cassandra, Redis, and Elasticsearch.
- Owned features end to end on a 5-engineer back-end team, from the design through production rollout under high-throughput, low-latency constraints.
- Operated event services against live player traffic, tuning data access across the polyglot store for production load rather than test load.
Software Development Engineer
Amazon.com
- Ran a zero-downtime migration of 5+ million Turkish-market user records from a legacy relational store to DynamoDB—(checkout and tax): dual-write, backfill, validation—then a progressive read cutover completed in around 1 hour.
- Delivered cross-team launches as an Away team, including Vendor Powered Coupons on Amazon.com (Netherlands and Sweden).
- Built and operated back-end services in Java/Spring on AWS (DynamoDB, EC2, ECS, Lambda, S3, SQS, SNS, DLQ, IAM) under production checkout traffic.
- Designed for correctness under load, applying idempotency, dead-letter handling, and progressive cutover so a checkout-path change could ship without downtime.
Digital Tech Developer Analyst
Accenture
- Built the back end for Roche's customer-experience dashboard in Spring/Java on AWS, delivering the data and API layer behind the client-facing reporting views.
- Worked toward the client's reporting requirements, shipping the API contracts and data access that the dashboard read from.
- Wrote and integrated the SQL data access that fed the dashboard, keeping the API responses aligned with the client's reporting schema.
Junior Software Engineer
JotForm
- Joined as a full-stack intern and was kept on part-time as a junior engineer, moving from triaging production issues to owning a feature end to end across the team's product surfaces.
- Led the intern team's back end for a commenting and feedback system on an online PDF editor, designing and shipping the PHP APIs over MySQL from specifications to production.
- Fixed live production bugs across the stack, in both the PHP back end and the React/TypeScript front end, on services used by paying customers.
- Worked inside an established engineering org with code review and a release process, before later moving to build-from-scratch founding work.
Webmaster
IEEE Bilkent Örenci Kolu
- Built the new IEEE Bilkent Student Branch website on a React front end with a Django back end over SQLite, replacing the previous static site.
- Built the GRC2020 conference site, an HTML5, CSS, and JavaScript front end with a PHP and MySQL signup and login flow that let graduates register and upload their projects.
- Contributed to a Bilkent course-schedule planning tool written in Vue, used by students to lay out their semesters.
- Administered the Debian server hosting ieee.bilkent.edu.tr, handling deployment and uptime for the branch's web presence.
Software Engineer Intern
SEN Sistem Entegrasyon
- Built a route-validation system for UAV flight planning that checked planned routes against impact-threat data, flagging unsafe paths before flight.
- Worked on the geospatial layer in Python with GDAL, reading and manipulating geoTIFF terrain and threat files to drive the validation logic.
- Owned the system's API, building it in Flask so the validation could be called as a service rather than run as a one-off script.
Web Developer Intern
Plum Media Agency
- Built an internal CRM in PHP and jQuery covering the agency's development process, receipts, contacts, and client messaging.
- Built client websites in WordPress across eCommerce, business, and media use cases, handling theme and feature work.
- Built a website cost calculator in jQuery, HTML5, and CSS that gave prospects an estimate before a sales call.
Experience
BrandVox AI – Multitenant AI SaaS (RAG, Agents, Real-time Voice)
https://brandvoxai.comI built it solo as the dominant author (around 1,557 commits) across 5 production services: an around 90,000-LOC FastAPI back end, a forked Flowise (Node.js/TypeScript) agent engine extended with around 8,700 lines of custom nodes, a React 19 dashboard with a Shadow-DOM embeddable widget, a scraper, and a Next.js SEO site.
The hard part was the retrieval that survives real customer data, not demo corpora. I built hybrid retrieval in the Flowise engine: Qdrant dense vectors fused with BM25 sparse through a custom Reciprocal Rank Fusion ranker, with a weighted 0.7/0.3 ensemble live in production and RRF switchable.
A multi-provider LLM layer (OpenAI, Anthropic, Gemini, Grok) with cascading fallback keeps generation up when a provider fails, and a browser-to-OpenAI Real-time voice agent (PCM16, server VAD) handles barge-in with exactly-once turn persistence. Cost is controlled with provider routing and credit-based per-model accounting per request, on AWS ECS Fargate (Terraform + GitHub Actions). It runs at 1,000+ users and 30+ paying customers, operated solo.
Kodwai – AI-fluency Assessment & AI-native Interview Platform
https://kodwai.comI built the full stack solo: a published TypeScript CLI (@kodwai/cli), a FastAPI scoring back end on Turso/libSQL, and Next.js 16 clients for a B2C developer leaderboard and a B2B interview product that shares infrastructure.
Scoring runs in 2 phases. Phase 1 is deterministic: test pass rate, lint and nesting-depth heuristics, and iteration signals, with skipped-signal renormalization so a submission without an API key still gets a meaningful objective score. Phase 2 is an anchored 0-10 Claude Sonnet judge over spec precision, verification rigor, and decomposition, median-of-N sampled for stability.
The piece clients ask about most is the proxy: an async Anthropic proxy that swaps a per-session token for a server-held key, parses streaming SSE deltas to compute true per-token USD cost in real time, and enforces per-session time and budget limits (HTTP 402 at the cap), so a real key never touches an untrusted candidate machine and interviewer spend stays bounded. Stored 3rd-party keys use AES-256-GCM with a fresh nonce per write.
Ksenda – Multitenant BYOK Cold-email SaaS
https://ksenda.comI built Ksenda on Next.js 16, React 19, and Prisma 7 over Turso (libSQL) with 14 data models and row-level isolation across a 73-route API. The AI work is a batched, concurrent, per-domain-cached Gemini classifier (urlContext plus Google Search grounding) that gates lead imports on a company's AI presence at roughly $0.0018 per company, with self-healing retries for verdicts the model drops on large batches.
Follow-ups are RFC822-correct: Ksenda captures the initial Message-Id and replays it as In-Reply-To and References with a forced Re: subject, so Day 3/7/14 sequences collapse into the original Gmail thread. The automation runs as an Inngest pipeline that decomposes each per-row import, send, and follow-up into its own step (retries:0 to prevent double-sends, per-tenant concurrency 1) so multi-minute runs never hit serverless step timeouts, with daily caps and a timezone-aware send window. It runs at around 15% reply rate across 5,000+ emails (observed operating metric; reply detection is manual today).
Pitch IQ - Agentic live-sports companion
https://github.com/egehakan/pitch-iqA single supervisor graph implements all seven canonical agent runtime patterns end to end: a structured-output intent router; a ReAct agent bound to eight live-data tools whose prompt forbids stating any unverified fact; parallel provider fan-out; an orchestrator-worker briefing writer over the Send map-reduce API; a generator-critic prediction loop that checks the model's win probabilities against a de-vigged market line and revises up to two rounds; two-tier memory over a Postgres checkpointer and store; and a human-in-the-loop interrupt that pauses to confirm before it locks a bracket.
The chat streams the agent's working to the UI, the routed specialist, each tool call with live status, then tokens, over a custom NDJSON protocol.
Behind it sits an async FastAPI back end, a two-schema Postgres, Alembic, APScheduler, and a provider layer with caching, rate-limiting, and failover pulling real live data.
Education
Bachelor of Science Degree in Computer Science
Bilkent University - Ankara, Turkey
Certifications
Data Engineering Bootcamp
Trendyol
Skills
Libraries/APIs
OpenAI API, Claude API, REST APIs, API Development, Pydantic, LinkedIn API, Node.js, React, Turso libSQL, Stripe, Stripe Connect, SQLAlchemy, APScheduler, Playwright, Facebook SDK, Facebook API, Vue 3, GDAL, jQuery
Tools
Claude Code, Okapi BM25, Git, ChatGPT, Claude, GitHub, Amazon Elastic Container Service (ECS), Terraform, Amazon CloudWatch, Amazon Simple Queue Service (SQS), Amazon Simple Notification Service (SNS), AWS IAM, Claude Agent SDK, Shell, Shadcn UI, BigQuery, Sentry, Grafana, Google Analytics, Prisma
Languages
Python 3, TypeScript, Python, Java, SQL, JavaScript, HTML, CSS, Bash, Scala, PHP, Go, XML
Frameworks
Next.js, Spring Boot, Spring 6, LangGraph, Django, Flask, JSON Web Tokens (JWT), Express.js, Tailwind CSS, Spring, Alembic, Apache Spark, gRPC, LlamaIndex
Paradigms
Object-oriented Programming (OOP), Model Context Protocol (MCP), Microservices, Automation, Back-end Architecture, Event-driven Design (EDD), Event-driven Architecture, B2C
Platforms
Amazon EC2, Amazon Web Services (AWS), Docker, PostHog, AWS Lambda, Linux, Vercel, LangSmith, Apache Kafka, Apache Flink, Debian Linux, WordPress, Kubernetes
Storage
PostgreSQL, Databases, Amazon DynamoDB, Redis, Amazon S3 (AWS S3), NoSQL, JSON, MySQL, SQLite, Turso, SQL Injection Protection, Elasticsearch, Cassandra, Database Migration, Database Architecture, MongoDB
Other
FastAPI, LangChain, RAG Systems, Algorithms, Data Structures, Programming Languages, Artificial Intelligence (AI), Software Engineering, Retrieval-augmented Generation (RAG), Hybrid Search, Qdrant, Vector Databases, Vector Search, Reciprocal Rank Fusion, Gemini API, AI Agent Orchestration, AI Agents, Agentic Workflow Design, Tool Calling, Prompt Engineering, Server Sent Events (SSE), Flowise, Multi-tenant SaaS, AWS ECS Fargate, Multimodel AI, Model Agnostic Development, Agent Orchestration, LLM Orchestration, Pgvector, Keyword Searches, LLM-as-Judge, Evaluation, LLM Evals, Per-token Cost Metering, AI Engineering, Multimodal GenAI, Multimodal Models, Embedding Models, Semantic Search, Chunking, Generative Artificial Intelligence (GenAI), Human-in-the-loop (HITL), Large Language Models (LLMs), Agentic AI, Agentic AI Systems, Conversational AI, AI Chatbots, AI Integration, LLM Integration, OpenAI, Anthropic, Gemini, Back-end, Full-stack, Full-stack Development, Web Development, APIs, API Design, API Integration, API Documentation, RESTFul APIs, Migration, Authentication, Multitenancy, Software Architecture, System Architecture, Agentic RAG Systems, Agentic Coding, Communication, Structured Outputs, Generator-Evaluator, Streaming, Web Crawlers, Web Scraping, Data Scraping, Leads, LinkedIn, Lead Generation, Multi-agent Orchestration, Workflow, Workflows, Multi-agent Systems, Integrations, Reviews, Debugging, Workflow Automation & System Integration, LLM Agents, API Management, Computer Organization, Machine Learning, OpenAI Realtime API, WebSockets, Server VAD, Fal.ai, Cloudflare, Distributed Systems, Low-latency Software, Idempotency, Code Review, Realtime Voice AI, RFC82, Apollo.io, Re-ranking, Amazon Bedrock AgentCore, Image Generation, Video Generation, Natural Language Processing (NLP), Function Calling, Generative Pre-trained Transformers (GPT), Large Language Model Operations (LLMOps), AI Automation, Supabase, Real-time Data, Scalability, Data Modeling, Data Architecture, Enterprise Software, eCommerce, Amazon RDS, Command-line Interface (CLI), Architecture, Cloud Architecture, Cloud Infrastructure, System Integration, Integration, AI Tools, AI Product Strategy, Workflow Automation, Minimum Viable Product (MVP), User Experience (UX), User Interface (UI), Single Sign-on (SSO), Async.js, Facebook, CRM, Security, Front-end, Operating Systems, GitHub Actions, Server Administration, Geospatial Data, GeoTIFF, Security (AES-CCM), Google Search Console, Inngest, Nodemailer, SMTP, Pinecone, Weaviate, Hugging Face, Machine Learning Operations (MLOps), Design
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring