EHK
§ 00 · COVER

AI-native MVPs.End to end.

I help founders and product teams ship AI products in production: agents, RAG, multi-modal, realtime voice, plus the mobile, web, and AWS infrastructure around them.

Ex-AmazonAccentureFounder / CEO, BrandVox AI
Book a discovery callor email
FIG. 00.A · INFERENCE STREAMlive on desktop
↓ 01SCROLL
§ 01 · SELECTED PROJECTS

Three live builds, 2024 to now.

Each shipped to production. Real users, real revenue, real metrics. Specifics, not screenshots.

  1. § 01.012024 → present
    BrandVox AI
    brandvoxai.com

    Brand teams need on-brand content at scale across channels. Off-the-shelf LLMs hallucinate brand voice; generic AI writers don't know the brand.

    ROLEFounder, CEO. Architected and shipped the platform solo.

    STACK
    • FastAPI
    • React 19
    • Flowise (forked)
    • Qdrant + BM25 + RRF
    • DynamoDB
    • ECS Fargate
    • OpenAI Realtime
    • Fal.ai
    OUTCOME

    1000+ users. 30+ paid customers across $59-$449/mo tiers. Multi-tenant SaaS spanning 5 services, 4 LLM providers, 6 image models, 4 video models, realtime voice. Solo from zero to revenue.

    FIG. 01.ABrandVox AI · System topology
    CLIENTSSERVICESDATA STORESPROVIDERSVOICE WSAPP · MEDIATRAIN PIPELINECHATAUTH · QUOTAS · BILLINGBVConditionBVKnowledgeBVAnswerGenBVHumanSupportAppKnowledgeBVAnswerBVLeadsBVButtonsSUPPORT BVchat + voiceSOCIAL MEDIA BVcontent + schedulingIMAGE GEN BVfal.ai · 5 modelsVIDEO GEN BVfal.ai · wan 2.6TRAIN UI7 sourcesDYNAMODBprimaryQDRANTvector / RAGREDIScache · queuesS3assetsELASTICanalyticsOPENAIANTHROPICGEMINIGROKFAL.AISCRAPERcrawl4ai · playwrightBRANDVOX-APIfastapi · 205+ servicesFLOWISE (forked)8 BV agentflow nodesHYBRID RAG · BVKNOWLEDGE_AGENTFLOWQdrant (k=12) + TFIDF · RRF q=0.7 / t=0.3 · LRU 24h
    ·

    Hover a node for detail. Realtime voice (PCM16, server-side VAD), hybrid RAG (Qdrant + TFIDF + RRF), 4 LLM providers + Fal.ai, 5 image and 1 video model. Solo-built, in production.

    ↓ 02
  2. § 01.022025 → present

    Kodwai

    kodwai.com

    Modern engineers work with AI agents (Claude Code, Cursor) all day, but interview platforms still test “no-AI” coding under artificial constraints. Companies cannot measure how well a candidate collaborates with an AI agent; developers have no public credential for AI-agent fluency.

    ROLESolo build: API, CLI on npm, client app, landing site, two-phase scoring engine.

    STACK
    • FastAPI
    • Turso / libSQL
    • Next.js 16
    • React 19
    • TypeScript CLI
    • Anthropic proxy
    • Railway
    • Vercel
    OUTCOME

    400-person waitlist. 100 active beta users. 22 challenges live. @kodwai/cli published on npm. B2C developer leaderboard + B2B interview SaaS sharing one two-phase scoring engine: objective (test pass rate, lint, complexity, iteration speed) and analytical (Claude Sonnet 4.6 quality assessment). AES-256-GCM encryption for stored Anthropic keys, per-session budget enforcement.

    FIG. 02.AKodwai · System topology
    CLIENTSSERVICESDATAPROVIDERSPROXYAUTHCHALLENGESSUBMISSIONSLEADERBOARD · BADGES · ENCRYPTIONTEST PASSCOMPLEXITYITERATIONANALYTICAL · CLAUDE SONNET 4.6LATE PENALTYBADGE EVALSESSION CLEANUPEMAIL · RESENDBADGE ENGINECLI@kodwai/cli · npmAGENTclaude code · cursorWEB CLIENTnext.js · ~13K LOClibSQL · TURSOusers · challenges · submissions · developer_profiles · badges · api_keys · sessions · leaderboard · admin_audit_log · 17+ tables · 10 migrationsANTHROPICRESENDnpmRAILWAYVERCELBACKGROUNDasync workersKODWAI-APIfastapi · 24+ routersSCORING ENGINEchallenge_scoring.py · 633 LOC
    ·

    Hover a node for detail. Anthropic-proxy with per-session budget enforcement, AES-256-GCM key encryption, two-phase scoring engine (objective + Claude Sonnet 4.6 analytical), CLI captures git/code/agent traces, badge engine on every score.

    ↓ 03
  3. § 01.032025 → present
    Ksenda
    ksenda.com

    Founders and SDR leads running outbound need three things existing tools (Apollo + lemlist / Instantly / Smartlead) don't deliver: per-prospect AI writing that doesn't read as AI, threaded follow-ups that look like replies, not fresh outbound, and review-first defaults that scale to hands-off once trust is built. Existing tools also lock users into shared sending infrastructure and per-seat fees most outbound operators don't want to pay.

    ROLESolo build: multi-tenant SaaS. BYOK per-user Apollo / Gemini / SMTP keys, 5-state pipeline with threaded Day 3/7/14 follow-ups, AI-presence targeting, 30-day campaign planner, 14 Prisma models.

    STACK
    • Next.js 16 · App Router
    • React 19
    • Prisma 7 · libSQL adapter
    • Turso / SQLite
    • Tailwind 4 · Radix UI
    • TanStack Query · Zustand
    • JOSE · bcryptjs auth
    • Apollo API v1
    • Gemini 3 Pro (BYOK)
    • Nodemailer · user SMTP
    • Resend · platform transactional
    OUTCOME

    Battle-tested on BrandVox AI's outbound: 5,000+ emails, ~15% reply rate vs 1-5% industry benchmark. Multi-tenant SaaS, every user brings their own Apollo + Gemini + SMTP keys. 14 Prisma models. 5-state pipeline + Day 3 / 7 / 14 threaded follow-ups (Gmail-threaded via In-Reply-To). Gemini AI-presence detection for 'no_ai' targeting; five independent automation toggles for hands-off operation.

    FIG. 03.AKsenda · multi-tenant pipeline
    CLIENTSSERVICESDATAPROVIDERSSEARCH ORGPEOPLE · 115 TITLESAI DETECTPENDING_GENERATIONEMAIL_NOT_GENERATEDPENDING_REVIEWAPPROVED_TO_SENDSENTGEMINI GENFOLLOW-UP ENGINENODEMAILER · BYOK SMTPDISCOVERY UIapollo searchPIPELINE BOARD5-state company listREVIEW MODALapprove · edit · regenPRISMA · libSQL · TURSOuser · company · email · follow_up · prompts · titles · ai_cache · audit · saved_search · campaign_day · jobs · 14 modelsAPOLLO APIGEMINI 3 PROSMTP · GMAIL · OUTLOOKRESENDAPOLLOdiscovery · 115 ICP · AI detectPIPELINE STATE MACHINE5 states · 3 follow-ups · audit-loggedDELIVERYgen · follow-up · send
    ·

    Hover a node for detail. Multi-tenant SaaS — each user owns their Apollo / Gemini / SMTP keys (BYOK). 5-state initial pipeline (pending_gen → not_gen / pending_review → approved → sent) plus Day 3 / 7 / 14 threaded follow-ups. 115-title ICP enforcement, Gemini-powered AI-presence detection for 'no_ai' targeting, mandatory human-review default with five independent automation opt-out toggles, Resend for platform verify, Nodemailer for user campaigns.

§ 02 · HOW I WORK

Three engagements. Fixed price. Written down.

No discovery-call sales funnel, no scope-creep negotiation. Pick the one that maps to where you are. The audit converts to an MVP at a locked price within 14 days.

§ 02.01

AI Opportunity Audit

Duration
2 weeks
Best for

Founders and CTOs who know they should be doing AI but cannot articulate which use case to bet on. Used as a low-risk wedge into a 6-week MVP.

Deliverables
  1. 01Three stakeholder calls, codebase walkthrough, current AI / data inventory.
  2. 02Architecture diagram of the existing system + AI integration map.
  3. 03Prioritized list of 3 to 5 AI use cases ranked by ROI vs effort, with build / buy per case.
  4. 0490-minute live readout call + written report (10 to 15 pages).
§ 02.02

6-Week AI MVP

Duration
6 weeks
Best for

Teams with one specific AI feature to ship who want it production-grade, observable, and cost-controlled. Not a notebook demo.

Deliverables
  1. 01One AI feature shipped end-to-end: RAG chatbot, agentic workflow, vision pipeline, voice agent, or multi-modal generator.
  2. 02Production deployment to your stack, or AWS / GCP greenfield (ECS Fargate, DynamoDB / Postgres, Redis, S3).
  3. 03Sentry + PostHog + per-request token-cost tracking dashboard.
  4. 04Eval harness with 20+ test cases, regression-proof for future model migrations.
§ 02.03

Fractional AI Engineer

Duration
3-month minimum
Best for

Post-MVP clients, or teams with engineering muscle who need a senior AI specialist on speed-dial. Maximum two to three concurrent retainers.

Deliverables
  1. 01One day per week dedicated, flex within the month.
  2. 02Architecture review + PR review on AI-relevant code (RAG, prompts, evals, agent flows).
  3. 03Model migration & cost optimization (Opus → Sonnet → Haiku routing, prompt caching, RRF tuning).
  4. 04Direct Slack channel + weekly written status. No standing meetings required.

Not sure which one fits? Book a discovery call. Thirty minutes, no slide deck, you tell me the problem, I tell you which engagement to start with. No charge.

Book a discovery call
§ 03 · EXPERIENCE

Where I've shipped before.

Nine years, three cities. Backend at Amazon and Accenture before going solo. The full record below.

  1. 2024 → now
    BrandVox AI
    Delaware / Istanbul (remote)
    Co-Founder & CEO

    Architected and built the multi-tenant AI SaaS platform from scratch. 5 services, ~23K source files. 1000+ users, 30+ payed customers.

  2. 2024
    Dream Games
    Istanbul, TR
    Backend Engineer

    High-throughput services on Spring (Java). MySQL, Cassandra, Redis. Kafka, gRPC, AWS DynamoDB and Lambda.

  3. 2022-2024
    Amazon
    Luxembourg City, LU
    Software Development Engineer

    Checkout & Tax systems, Away team. Designed a data-migration system on DynamoDB, EC2, ECS, Lambda, S3, SQS, SNS, DLQ. Backend in Java and Clojure.

  4. 2022
    Accenture
    Istanbul, TR
    Digital Tech Developer Analyst

    Backend engineer for the Roche customer-experience dashboard. Spring on AWS.

  5. 2017-2022

    Bilkent University

    Ankara, TR
    BS, Computer Science
§ 04 · CONTACT

Let's talk.

The right way to start is a thirty-minute discovery call. Bring the problem; I'll tell you which engagement to start with, or recommend someone else if I'm not the right fit.

Ege Hakan Karaagac, photographed in Istanbul, 2026.
FIG. 04.AEge Hakan Karaagac · Istanbul · 2026

I work with founders and product teams shipping AI in production. Most engagements happen end-to-end: architecture, build, deploy, and observability. Engineering is the entire deliverable, not the half that gets handed back to the client.

Email
hakan@egehakankaraagac.com
Calendar
cal.com / ege-hakan-karaagac-s7l6lv / 30min
GitHub
github.com / egehakan
LinkedIn
linkedin.com / in / egehakankar
Phone
+90 507 738 5277
Location
Istanbul, TR · UTC+03 · Remote-first
Book a discovery call30 minutes · no slides · no charge
REV. 2026.05 · EHK-PORTFOLIO-A© 2026 EGE HAKAN KARAAGAC ·↑ TOP