I build production AI systemsthat real users pay for.
Retrieval, multi-agent orchestration, and realtime voice, shipped as products and not demos. I founded BrandVox and built it to 1,000+ users and 30+ paying customers, mostly alone, across five services on AWS. Before that I spent two years at Amazon on checkout and tax, where the money has to be right every time.
- GitHub
- github.com/egehakan
- linkedin.com/in/egehakankar
What I build
Not a tag cloud. Five things I have actually put in front of paying users, each with the project that proves it.
- § 01.1
Retrieval & RAG
Qdrant dense vectors and BM25 sparse, fused live by a weighted ensemble (0.7/0.3) with a custom Reciprocal Rank Fusion ranker switchable, all in my forked Flowise engine on OpenAI text-embedding-3-small.
BrandVox - § 01.2
Multi-agent orchestration
~8.7K LOC of custom nodes on a forked Flowise engine, with a multi-provider answer path that cascades across OpenAI, Anthropic, Gemini, and Grok, 2 retries per model and a 60s per-invoke timeout.
BrandVox - § 01.3
Realtime voice
a browser-to-OpenAI Realtime pipeline over WebSockets, PCM16 at 24kHz with server VAD, dual-path barge-in and exactly-once turn persistence across a known concurrency race.
BrandVox - § 01.4
Evals & cost control
a two-phase scoring engine plus a token-swapping Anthropic proxy that meters per-model USD cost from streaming SSE deltas and cuts a session off at its budget with an HTTP 402.
Kodwai - § 01.5
Distributed systems at scale
a zero-downtime migration of 5M+ user records to DynamoDB, dual-write then backfill, validation, and a progressive read cutover that finished in about an hour.
Amazon
Selected work
Three systems I designed, built, and still run myself. Real stacks, real users, honest numbers.
Multi-tenant AI SaaS for the customer-facing busywork a small team would grind through by hand.
I architected and shipped BrandVox from zero as the dominant author, around 1,557 commits across five services. It answers on chat and voice from each customer's own content, drafts content, and surfaces leads. The proof is not a demo, it is that people renew every month.
- 1,000+
- users
- 30+
- paying $59–449/mo
- 5 services
- one engineer
Each tenant gets an isolated workspace; retrieval lives in my forked Flowise engine where Qdrant dense and BM25 sparse are fused live by a weighted ensemble (0.7/0.3), with a custom Reciprocal Rank Fusion ranker switchable and a per-chatbot LRU cache; ingestion runs through an async Redis-backed queue with bounded concurrency and a 30-min watchdog, and cost is credit-based per-model accounting with per-request tracking, monitored on Sentry, Grafana, and PostHog.
Scores how well an engineer directs AI coding agents, not how fast they type.
I built Kodwai end to end: a published npm CLI that launches Claude Code or Cursor and captures the trace, a FastAPI backend that scores the submission, and a Next.js client. It hands the repetitive part of technical hiring to AI while a person keeps the call.
- 22
- scored challenges live
- v1.7.0
- @kodwai/cli on npm
- 494
- backend tests
A two-phase scoring engine pairs deterministic signals (test pass rate, lint and nesting-depth heuristics, iteration signals) with a Claude Sonnet 4.6 judge, dropping skipped signals from the denominator so a keyless run still scores, layered with a calibrated operator-ability model: continuous-response IRT that reports a latent-skill estimate with a Fisher-information confidence interval and picks the next challenge adaptively, a counterfactual lift measure against a solo-AI baseline, and mutation-tested outcome grading whose axis weights are learned rather than hand-set. A token-swapping Anthropic proxy meters per-model USD cost from streaming SSE deltas and enforces a per-session budget so a real key never touches an untrusted machine, with stored keys under AES-256-GCM.
Multi-tenant BYOK cold-email system. I built it, then ran my own outbound on it.
Ksenda is how I filled the paying side of BrandVox. It finds companies by a real signal, drafts a first message tied to something specific about each one, and threads follow-ups so nobody falls through the cracks. A human approves every send.
- 5,000+
- emails sent
- ~15%
- reply rate vs 1–5% norm
- 1
- human reviewer
Day 3/7/14 follow-ups thread correctly through Gmail with RFC822 In-Reply-To and References headers, a batched and cached Gemini classifier gates imports by AI-presence signal, and an Inngest per-row pipeline survives serverless timeouts; row-level tenant isolation runs across 73 API routes, and reply tracking is counted by hand, so the ~15% is a number I counted, not one a dashboard flattered me with.
By the numbers
What the work adds up to, with nothing rounded up.
Trajectory
From correctness-at-scale inside Amazon to a product I own end to end.
Founding EngineerDelaware, USA / remoteMy own company. I built a multi-tenant AI SaaS from zero, dominant author at ~1,557 commits, now serving 1,000+ users and 30+ paying customers.
Backend EngineerIstanbul, TRLive in-game event services behind a top-grossing mobile game with 6M+ players, in Spring Boot and gRPC over MySQL, Cassandra, Redis, and Elasticsearch. Five-engineer team, features owned end to end.
Software Development EngineerLuxembourg City, LUCheckout and Tax across Sweden, Poland, the Netherlands, and Turkey, the systems that move customer money and have to be right every time. I ran the 5M-record zero-downtime migration to DynamoDB and shipped Vendor Powered Coupons.
Digital Tech Developer AnalystIstanbul, TRBuilt the backend for Roche's customer-experience dashboard in Spring and Java on AWS.
Fixed production issues across a PHP and React/TypeScript stack, and led an intern team building a commenting system for an online PDF editor.
Education
- 2017 – 2022Bilkent University, AnkaraBS, Computer Science
Algorithms, data structures, computer organization, operating systems, databases, programming languages.
- 2021Trendyol (remote)Data Engineering Bootcamp
BigQuery, Scala, Apache Spark, Apache Kafka, Apache Flink.
About
I'm an AI engineer based in Istanbul. I spent two years on backend systems at Amazon and Dream Games, then two years building an AI product on my own. The backend work taught me production discipline; building solo taught me how much you can ship when you own every layer.
I care most about the parts that are easy to skip: retrieval that holds up on real customer data, agent orchestration that stays stable under load, and cost measured per request instead of estimated. I build the AI, then run it, support it, and keep an eye on the numbers.
The AI is the product, not a feature bolted on. Everything else is plumbing I am happy to own
Get in touch
If you're working on something with production AI, or just want to talk shop, write to me.
- GitHub
- github.com/egehakan
- linkedin.com/in/egehakankar
- Where
- Istanbul, Turkey · UTC+03