# Hakan Baybaş

> AI engineer who thinks out of the box and builds with a product mindset. Retrieval and ranking, agents and evaluation at Rezolve Ai and Writer, with 19 years of building software behind it, much of it as a founder, hands-on CTO and team lead.

Senior AI/ML Engineer, based in London, UK. Website: https://hakanbaybas.dev. How AI agents should use this file: https://hakanbaybas.dev/agents.md

## About

AI engineer who thinks out of the box and builds with a product mindset, based in London.

Today I own the AI/ML behind **Rezolve Ai's shopping assistant**: retrieval and ranking, intent matching and evaluation. Before that I led a team of 4 on **Agent Deployment**, the multi-agent platform inside Writer's Agent Builder.

My product sense comes from building companies. I set up two software companies, **DIGIEGGS** in Turkey and **AlmondSpring** in the UK, and co-founded two startups, **Logg** and **Commented**. Logg was Apple's App of the Week.

My calm comes from leading. As a founding engineer and hands-on CTO I led development teams of up to 12 people, and product and design teams, across time zones. I hire and grow engineers: one junior I hired became a CTO.

I have worked with teams and clients in the UK, the US, Turkey and Saudi Arabia, and I am a British citizen.

## How I work

- **Out-of-the-box thinking**: I question the default before I follow it. Social apps share by default, so at Logg we made privacy the default instead. Apple made it App of the Week and Product Hunt made it App of the Day. (Proof: https://hakanbaybas.dev/projects/logg)
- **Critical thinking with AI**: I use AI agents every day, but I own what ships. User stories and behavioural checks tell them how the code must behave under load and across tangled dependencies, so they work like clones of me. I review their output like a new hire's, and evals and tests, not a confident answer, decide when it is done. That is how broken replies went from about 7% to zero. (Proof: https://hakanbaybas.dev/projects/rezolve-ai)
- **Problem solving under constraints**: When the obvious fix is too slow or too costly, I find another way in, with a bias toward execution. A cloud-only SaaS became an on-premise product in two weeks because I made its data layer point at another database instead of rebuilding it. (Proof: https://hakanbaybas.dev/projects/upsolve-on-prem)
- **Product mindset**: I build what users will pay for, not what's fun to build. At Commented, customers and sales data showed us where the value was, so we moved from a general commenting tool to collaboration between software teams and their clients. Since launch, 800+ teams have left 56,000+ comments on it. (Proof: https://hakanbaybas.dev/projects/commented)
- **Agile**: Product plans become sprint plans, and sprints land. At Decktopus AI I turned product plans into sprints with the team and reviewed the work so the old editor kept serving customers. We delivered 90% of planned sprints on time. (Proof: https://hakanbaybas.dev/projects/decktopus-ai)
- **Root causes, not symptoms**: I start with what success looks like and the people closest to the problem, then fix the cause. A gap I saw in pharma sales became a CRM product, not another spreadsheet. (Proof: https://hakanbaybas.dev/projects/digieggs)
- **Quality by default**: Written plans and reviewed code, not hope. I spot quality gaps early, from broken UX to missing API contracts, and step in to close them myself. (Proof: https://hakanbaybas.dev/projects/karshare-takeover)
- **One design, every team**: Security, DevOps and customer teams rarely want the same thing. I write the trade-offs down, hear each team's constraints, and land one design they all stand behind. (Proof: https://hakanbaybas.dev/projects/writer-agent-deployment)
- **Close to the customer**: I go to the people using the product before I build. At Commented I was in more than 200 sales, demo and customer success meetings, which is where we heard what customers really wanted. (Proof: https://hakanbaybas.dev/projects/commented)

## Systems I have designed

- **Conversational AI, measured**: Retrieval and ranking, intent matching, attribute extraction and vision models for an AI shopping assistant, with a knowledge layer merchants author themselves. Evaluations decide what ships: first responses got about 30% faster. (Proof: https://hakanbaybas.dev/projects/rezolve-ai)
- **Multi-tenant agent execution**: Per-tenant sandboxes on Cloudflare Containers and Durable Objects, with MCP OAuth integrations for tools. (Proof: https://hakanbaybas.dev/projects/writer-agent-deployment)
- **Secrets and security hardening**: Decrypt-on-demand encrypted environment variables, ClamAV malware scanning, Redis rate limiting and a DynamoDB partition key strategy. (Proof: https://hakanbaybas.dev/projects/writer-agent-deployment)
- **Cloud to on-premise**: Containerised a cloud SaaS for on-premise deployment, with a dynamic data layer, migrations and SQL import and export. (Proof: https://hakanbaybas.dev/projects/upsolve-on-prem)
- **Serverless on AWS**: Lambda, GraphQL and CloudFront behind SDKs, a browser widget and an admin panel. (Proof: https://hakanbaybas.dev/projects/commented)

## Certificates and education

- 2026: AWS Certified Solutions Architect, Associate, Amazon Web Services, valid to 2029
- Earlier: SAP and Microsoft certifications, SAP Certified Application Associate, CRM (2013); Microsoft MCTS 70-562 and MCP 70-536 (.NET); Bilge Adam Database and Software Specialist (2010)
- 2007: BSc Computer Engineering, Dokuz Eylül University, Izmir

## Experience

### Senior AI/ML Engineer, Rezolve Ai

Jun 2026 to present · London

Design and build an AI shopping assistant end to end: I write the design doc, then build it across the chat service, the catalog enrichment pipeline, shared libraries and deployment config.

Own the AI/ML behind it: catalog-aware conversation, retrieval and vector search, intent matching, attribute extraction, vision models for product images, and model routing.

Turn real labels into better models: fine-tuning embedding, extraction and vision models, and learning to rank results from shopper signals.

Built a knowledge layer that lets merchants author, validate, publish and serve expert shopping knowledge to the assistant and through an API; in blind comparisons the assistant using it was preferred in about two thirds of conversations.

Ship changes only when an evaluation says they are better: a blind LLM judge plus automated checks over about 100 conversations per run.

Cut time to first response by about 30% and broken replies from about 7% to zero; took a new retail vertical's extraction pass rate from about a third to 100%; made catalogue ingestion reliable past 800,000 products. 260+ merged PRs in five months.

### Staff Software Engineer, Writer

Mar 2025 to May 2026 · London

Tech lead for Agent Deployment, an enterprise multi-agent sandboxed execution platform within Writer's Agent Builder, leading a team of 4.

Worked on Agent Builder v1 and v2, and led the platform and authentication architecture for v2.

Architected Agent Deployment on Cloudflare Containers and Durable Objects with MCP OAuth integrations: isolated, scalable per-tenant agent environments.

Designed the security layer: decrypt-on-demand encrypted environment variables, ClamAV malware scanning, Redis rate limiting and a DynamoDB partition key strategy.

Researched multi-agent orchestration, evaluated agent frameworks (AWS Strands SDK) and MCP as a standard tool interface.

Coordinated across SecOps, DevOps, CE Engineering, customers and solution architects.

### Hands-on Frontend Manager (Consultancy), Decktopus AI

Aug 2024 to Feb 2025 · Remote (Palo Alto, California)

Led development of an AI-driven presentation editor on HTML canvas, built with React, React-Konva and TypeScript.

Led the front-end team and planned feature work with product managers and the backend and AI leads.

Set the technical roadmap with the CTO, and kept the old editor stable for its customers while the team built the new one.

Built reusable, editable canvas components such as tables, stacks and timelines, plus charts, preview and export.

Delivered 90% of planned sprints on time, with regular 1:1s for team growth.

### Principal Engineer (Consultancy), Upsolve AI

Jul 2024 to Aug 2024 · London

Worked with the CTO to turn the cloud SaaS platform into an on-premise product for a major customer.

Built the on-premise infrastructure and Docker containers, and made the data layer point at another RDS dynamically.

Wrote data migration strategies, maintenance commands, and SQL import and export for on-premise clients.

Delivered documented, containerised on-premise applications in two weeks.

### Principal Engineer, AI Engineer (Consultancy), Barkley AI

Jun 2024 to Jul 2024 · London

Reworked the meeting summary flow and its UI, using Azure and OpenAI SDKs to generate summaries and action points from meeting videos.

Joined client meetings to understand pain points and proposed fixes for navigation and general problems.

Designed a customisable deployment strategy for client onboarding.

### Co-Founder, Hands-on CTO and CEO, Commented

Jun 2023 to Sept 2024 · London

Took Commented from idea to product: a tool that lets software teams and their clients comment on any web application.

Since launch: 3,300+ sign-ups, 800+ teams using it, and 56,000+ comments across 39,000+ feedback threads on 850+ customer websites.

Built the JS widget, the Chrome extension, the iOS SDK and the React Native SDK, and the AWS infrastructure with my co-founder.

Used the OpenAI API for summaries and tone suggestions.

Ran weekly 1:1s with a team of 10, one-week sprints and the product roadmap. Made three pivots based on customer and sales data.

Held 40+ investor meetings and 200+ sales, demo and customer success meetings. Secured term sheets for a $1M seed round, which did not close.

### Founding Engineer, Engineering Manager (Interim, Consultancy), Bermudair Airlines

Dec 2022 to May 2023 · London

An interim consultancy, started alongside my Karshare role: one of the founding engineers, setting up the initial work with an outsourcing agency on the airline's mobile apps.

Led a mobile team of 4 and a design team of 2 on React Native apps and a Next.js admin panel.

Built the backend and data structures that wrap the AERO CRS airline API.

### Hands-on Software Technical Manager, Karshare

May 2021 to Dec 2022 · London

Set up product, design and development teams and recruited for them.

Introduced fast, agile sprints that helped the company raise money, hit milestones and scale.

Led native iOS SDK work and the React customer-facing and admin apps; reviewed backend work and PRs.

Worked with the CTO on technical and infrastructure decisions.

### Founding Engineer, Karshare

Nov 2019 to May 2021 · London

Took over the product from an outsourcing agency that was failing to deliver.

Stepped in on front end, UX and parts of the backend where contracts were missing.

### Lead Front-End Developer (Consultancy), Qubit (Coveo)

May 2019 to Jan 2020 · London

Built Qubit's new web app from scratch with React and Hooks, tested with React Testing Library and backed by GraphQL.

Built the integration app between Qubit and the Google Cloud Platform Marketplace.

Fixed bugs in the iOS SDK in Swift and published it as a CocoaPod.

Led the front-end developers through code reviews, pair programming and mentoring.

### Co-Founder, CEO and CTO, Logg

Oct 2017 to Dec 2018 · Izmir

Designed and built a private life journal and tracker for iOS, desktop and web.

Built the native iOS app and the Electron desktop app, and worked on the backend and APIs.

Featured by Apple as App of the Week (August 2018) and on Product Hunt as App of the Day (July 2018).

### Founder, DIGIEGGS

Sept 2014 to Apr 2018 · Izmir

Founded a software company that delivered bespoke web and mobile apps for 15 clients and 20+ projects.

Launched products that earned revenue: some were sold to new owners, others were ended.

Hired junior developers and trained them; the team grew to between 10 and 20 people.

Clients included Yaşar Holding, Decktopus, İnci Holding, Nurol Teknoloji and Eczacıbaşı.

### SAP Certified Consultant, Astron (Yaşar Holding Bilgi Sistemleri)

Sept 2011 to Dec 2015 · Izmir

SAP CRM, Portfolio and Project Management (PPM) and BW consultant for the holding's companies.

Built ABAP solutions and new CRM flows, and solved customisation issues for internal customers.

### Software Engineer, Vestel

Aug 2009 to Aug 2011 · Manisa

Built embedded software for IPTV products.

Worked with hardware partners and the operator's test and QA teams.

### Software Development Engineer, Vestel

Aug 2007 to Aug 2008 · Manisa

Built embedded software for IPTV solutions in hospitals and hotels.

## Projects

### The AI behind Rezolve Ai's shopping assistant

https://hakanbaybas.dev/projects/rezolve-ai

Senior AI/ML Engineer · 2026

Building an AI shopping assistant end to end: retrieval, intent, extraction, vision, and the evaluation that decides what ships. Six case studies with measured results.

Tags: Python, LLM evaluation, Vector search, Fine-tuning, Learning to rank, Retail, LangChain, LangGraph, Loop engineering

#### What I do

I design and build an AI shopping assistant end to end. I write the design doc, then build it across the chat service, the catalog enrichment pipeline, shared libraries and deployment config.

I own the AI/ML behind it: catalog-aware conversation, retrieval and vector search, intent matching, attribute extraction, vision models for product images, and model routing.

Better models from real labels: fine-tuning embedding, extraction and vision models, and learning to rank results from shopper signals.

#### One shopping assistant for any catalogue

The problem: The assistant was built around fashion. Each new kind of merchant, like sports or grocery, meant new code paths, new prompts and new bugs.

What I did: I led the redesign into a neutral core plus thin vertical layers. The intent classifier and search tools work from each catalogue's own schema, not hard-coded product types. Each vertical adds a short prompt overlay and a few interface hints, nothing more. Cart, recall and similar-product tools resolve products generically for any catalogue.

The result: Adding a vertical became configuration instead of engineering. For a new sports vertical, the overlay took the extraction eval pass rate from about a third to 100%, with no regressions on existing cases.

#### Making an LLM pipeline feel fast

The problem: Each shopper turn chained several LLM calls: understand the turn, extract attributes, search, write the reply. Replies took over 10 seconds.

What I did: I ran attribute extraction at the same time as intent classification, so it left the critical path. When the shopper tapped a button and the intent was already known, the classification call was skipped. I reordered prompts so the model server could reuse cached work between turns, and added an instant acknowledgement so the wait felt shorter. I measured with alternating benchmark runs, so drift on the model server couldn't flatter the result.

The result: About 30% faster first response and about 20% faster full reply. A rewrite of the reply prompt also made it about 60% smaller.

#### A knowledge layer merchants write

The problem: A good shop assistant knows what to ask first ("road or trail?") and what to explain before selling. A general model doesn't know that about each merchant.

What I did: I designed a knowledge layer that merchants write: one playbook per shopper situation, each with must-know questions, advice and catalogue filters. The assistant uses it to pick its next move: ask, explain or show products. Every version is checked against the live catalogue and can't go live until a recognition eval passes. AI-assisted authoring drafts playbooks grounded in the merchant's real product values, and the same layer is served through an API that answers in tens of milliseconds, with no LLM call.

The result: In blind comparisons, the assistant with the knowledge layer was preferred in about two thirds of conversations. Recognition reached over 95% top-3 on a second vertical with no vertical-specific code.

#### Measuring an LLM product honestly

The problem: "It feels better" isn't evidence, and one LLM judge on its own is noisy.

What I did: I built a side-by-side harness: scripted shopper conversations, each played several times per variant, about 100 conversations per run. I added a blind LLM judge, then measured its own run-to-run noise. Code checks cover everything verifiable without a model: right category, gender and budget, products actually shown, usable replies, no products named that aren't on screen. Models are chosen by bake-off with decision rules written down in advance: I kept the cheaper embedding model because the alternative gained 3 points at about 4× the query cost. When the numbers showed a trial didn't help, I dropped it.

The result: The code checks matched hand-checked results exactly, so most regressions now get caught without a judge call. One finding, hidden reasoning using up the reply's token budget, took broken replies from about 7% to zero.

#### Vision models for product data

The problem: Merchant feeds often carry wrong or meaningless colours, like SKU codes or guesses, and product photos come with no labels for which angle they show.

What I did: I added a step that checks the listed colour against the product photo and replaces it only when the image clearly disagrees or the value is junk. I built image-angle labelling that sends several photos per vision call, routed footwear, the hardest category, to a larger vision model while everything else stays on a smaller, cheaper one, and added retries so one failed image download no longer blanks a product's visual attributes.

The result: Cleaner colour data for search, with every change counted (replaced, kept or unsure) so merchants can review it.

#### Vector ingestion at scale

The problem: Large catalogue ingests kept dying partway through. About a minute of blocking work made the server miss its health checks, so it was restarted and the job was lost.

What I did: I moved the heavy loading and enrichment off the main event loop, added a clean stop on shutdown, and made restarts resume: they look up which products are already done and embed only the rest.

The result: Reliable ingestion past 800,000 products. A restart now costs seconds instead of re-embedding the whole catalogue.

#### How changes ship

I build with AI coding agents every day, and I own what they write. User stories and behavioural checks tell them what the code must do and how it must behave under load and across tangled dependencies, so they work like clones of me. I review their output like a new hire's work.

A change ships only when an evaluation says it is better: latency comparisons, recognition tests and blind side-by-side conversation judges.

Output: 260+ merged pull requests in my first five months.

### Agent Deployment at Writer

https://hakanbaybas.dev/projects/writer-agent-deployment

Tech Lead, Staff Software Engineer · 2025, 2026

Running AI agents safely for many enterprise customers, in isolated per-tenant environments.

Tags: DynamoDB, Cloudflare Containers, Durable Objects, MCP, Redis, Enterprise AI, LangChain, LangGraph, Langflow

#### The problem

Writer's enterprise customers wanted to run AI agents on the platform. Each customer's agents had to run in their own isolated environment, so that one customer's agent could never see or affect another customer's data or execution.

Agent Builder was Writer's product for building agents. I worked on both versions, and for v2 I led the platform and authentication architecture. Agent Deployment was the part that had to run those agents for real customers, safely, for many customers at once.

#### What made it risky

Four things had to be right at the same time: tenant isolation, customer secrets, untrusted files, and agreement between SecOps, DevOps and the customer-facing teams. Getting any one of them wrong would have been a security incident or a stalled project.

Each of the four needed its own answer. Isolation meant per-tenant execution, so no two customers ever shared a runtime. Secrets meant customer keys never sat in plain text. Files meant anything a customer uploaded was scanned before an agent touched it. And every decision had to work for SecOps, DevOps, customer engineering and the solution architects.

#### What I did

As tech lead and lead of a team of 4 on Writer's Agent Builder, I architected Agent Deployment on Cloudflare Containers and Durable Objects, with MCP OAuth integrations for tools. I designed the security layer: environment variables encrypted at rest and decrypted only on demand, ClamAV malware scanning for files, Redis rate limiting, and a DynamoDB partition key strategy. To get every team behind one design, I wrote the options and trade-offs down, met each team one to one to understand its constraints, and then ran a single large design review.

Before building, I researched the options: multi-agent orchestration patterns, agent frameworks such as the AWS Strands SDK, and MCP as the standard for how agents use tools. That research fed straight into the design. Day to day I led and mentored the team, and former teammates describe me as a good mentor.

#### Result

The platform was live in a limited rollout when I left in May 2026. In the words of Writer's Sr. Director of Engineering, I "consistently turned ambiguous, cross-team problems into a shipped system."

### A new canvas editor at Decktopus AI

https://hakanbaybas.dev/projects/decktopus-ai

Hands-on Frontend Manager (Consultancy) · 2024, 2025

Leading the front-end team building Decktopus AI's new presentation editor on HTML canvas, while the old editor kept serving customers.

Tags: TypeScript, React, Konva, Jest, B2B SaaS

#### The challenge

Decktopus AI makes presentations with AI. The team was building a new editor on HTML canvas, and the old editor still had to keep working for every existing customer while that happened.

I knew the company already: Decktopus had been a client of DIGIEGGS, my first company. This time I joined as a consultant to lead the front end through the move to the new editor.

[Image: Decktopus AI today, from its website.]

#### What I did

I led the front-end team and planned feature work with the product managers and the backend and AI leads, and set the technical roadmap with the CTO. I reviewed the team's work so changes for the new editor did not break the old one for its customers.

Day to day, that meant turning product plans into sprint plans with the team, holding regular 1:1s so everyone kept growing and stayed aligned, and checking the team's work against the old editor so its customers kept a working product.

[Image: The Decktopus editor: the slide list, the canvas and AI editing.]

#### What I built

Hands-on, I built reusable, editable components on the canvas with React, React-Konva and TypeScript: tables, stacks and timelines, then charts with hover and drag, copy and paste, and preview and export.

Canvas components are harder than web page ones: every table cell, stack and timeline item has to be drawn, measured, selected and edited by hand. Between October 2024 and March 2025 I made 168 commits to the web app myself, alongside leading the team.

[Image: Slides made in Decktopus, from its website: the kind of editable blocks the canvas editor renders.]

#### Outcome

The team delivered 90% of planned sprints on time while keeping a healthy backlog, with regular 1:1s to keep everyone growing and aligned.

### Taking over Karshare's product

https://hakanbaybas.dev/projects/karshare-takeover

Founding Engineer, then Technical Manager · 2019, 2020, 2021, 2022

Rescuing a peer-to-peer car-sharing product from a failing outsourcing agency, then building the team that ran it.

Tags: Swift, React, Next.js, GraphQL, DynamoDB, Mobility

#### The situation

Karshare was a UK peer-to-peer car-sharing startup. Its product was being built by an outsourcing agency that kept missing commitments. When I joined as a founding engineer, I saw four problems at once: no clear product owner, a broken delivery process, code that blocked progress, and a vendor relationship that was failing.

Customers needed to find a car nearby, book it and unlock it from their phone. The operations team needed to manage customers and vehicles behind the scenes. Both sides depended on the product the agency was struggling to deliver.

#### What I did first

Before replacing anyone, I worked with the agency for a while, giving them clearer direction to test what they could deliver. The test led to a partial handover. The internal team already owned most of the backend, but the UI and UX were badly broken and some backend contracts were missing, so I stepped in on the front end, UX and parts of the backend.

As the first front-end-leaning founding engineer I built a native iOS app that let customers search and hire vehicles, connect to them and start and stop the engine; a React web app with vehicle search on a map; an admin panel for the operations team in React, TypeScript and GraphQL; and React components for the Webflow landing pages and programmatic SEO pages.

#### Then product, process and code

Next I defined the product, then set up fast sprints with clear ownership, and only then stabilised the code. I told the founders about the risk early and plainly, and backed it with a written handover plan and milestones. Later, as technical manager, I built the product, design and development teams.

As technical manager I worked with the CTO on technical and infrastructure decisions, reviewed the backend team's pull requests, and helped the marketing and sales teams bring new tools into their workflows. The sprint rhythm helped the company hit its milestones while it raised money.

#### Outcome, told plainly

The rebuilt product shipped, and according to public reports the company raised more than £1.4M through crowdfunding in early 2021 and a £3M round in August 2021. Karshare went into administration in June 2023, after I left. My lessons: define the product before building, and plan runway for the market, not just for delivery.

### On-premise in two weeks

https://hakanbaybas.dev/projects/upsolve-on-prem

Principal Engineer (Consultancy) · 2024

Turning Upsolve AI's cloud SaaS into an on-premise product for a major customer.

Tags: B2B SaaS, Next.js, Python, Docker, PostgreSQL, Supabase

#### The problem

A big customer asked Upsolve AI to run its cloud SaaS inside their own infrastructure. It was a good way for the company to move forward, but the platform had only ever run in the cloud.

I joined as a principal engineer on contract and worked closely with the CTO on the plan. The stack was Next.js and React on the front, Node.js and Python behind it, and Supabase with PostgreSQL for data.

#### First steps

I mapped the data layer to see which parts depended on which databases, packaged the system in Docker containers so it could start on a clean machine, and talked to the customer's team about their environment.

The cloud version was wired to its hosted database, so the first real job was to make that connection configurable without breaking the cloud product.

#### Trade-offs

Three tensions shaped the work: keeping the cloud and on-premise versions in sync, two weeks against completeness, and the risk of moving customer data. I made the data layer point at another database dynamically, wrote migration strategies and maintenance commands, and added SQL import and export for on-premise clients.

I also restructured the UI layer to talk to the new data layer, which helped keep the cloud and on-premise versions in step. Everything shipped with documentation, so the customer's team could run it without us.

#### Result

I delivered documented, containerised on-premise applications in two weeks. The customer accepted the build, and it led to follow-on work.

### Commented: from idea to term sheets

https://hakanbaybas.dev/projects/commented

Co-Founder, Hands-on CTO and CEO · 2023, 2024

Co-founded and led a collaboration tool that lets software teams and their clients comment on any web application.

Tags: TypeScript, Chrome Extension, Swift, React Native, AWS, Collaboration

#### The challenge

As CTO and CEO, I faced four problems at once: finding the product customers would pay for, building the core infrastructure, getting investor and customer traction, and coordinating the team. I took them in that order of importance: product clarity first, then traction, then infrastructure, then team rhythm.

Commented is a collaboration tool between software teams and their clients: clients leave comments directly on the web app being built, and the team works from them.

#### What I built

With the web team I built the JS widget at the heart of the product, which lets users comment on any web application in seconds, and a Chrome extension around it. With my co-founder I built the AWS infrastructure on Lambda and CloudFront, the APIs and GraphQL schemas, and the admin panel. I contributed to the iOS SDK and reviewed the Android SDK, and used the OpenAI API for summaries and tone suggestions.

I set up continuous integration and deployment, and designed the subscription features and the plan for turning users into paying customers. With design and product, I took new flows from scratch to stable prototypes.

#### Three pivots

Customers showed us they valued the product for a different use than we planned, and sales data showed where the gap was. We moved from a general commenting tool to collaboration between software teams and their clients, and from consumers to B2B. The mobile SDKs took more time and energy than expected.

Running the team, I held 1:1s with 10 people across development and product, ran one-week sprints and retros, and updated the strategic and product roadmaps as we learned. I was in more than 200 sales, demo and customer success meetings, which is where we heard what customers really wanted.

#### Results

Commented is still running. Since launch in 2023, more than 3,300 people have signed up, and 800+ teams have used it for real feedback: 56,000+ comments across 39,000+ feedback threads.

Feedback came in from more than 850 customer websites, plus iOS and Android apps through the mobile SDKs.

About 72% of feedback threads get resolved, with a median of about two days to close.

The pivot showed up in the data: in one of every five threads with a reply, both the team and their client took part.

Usage peaked in March 2024, with 100 active teams and nearly 3,000 comments that month.

#### The fundraise, told plainly

After 40+ investor meetings we secured term sheets for a $1M seed round. The round did not close: we had applied under the wrong UK tax-relief scheme for investors, HMRC did not allow the amendment, and investors would not invest without the relief. My lesson: check the investor scheme with a specialist before applying, not after the term sheets.

### Logg: a private life journal

https://hakanbaybas.dev/projects/logg

Co-Founder, CEO and CTO · 2017, 2018

Co-founded a private journaling and life-tracking app, featured by Apple and on Product Hunt.

Tags: Swift, Node.js, Objective-C, Electron, MongoDB, Consumer apps

#### The idea

Logg is a life logger: record activities and moments, track them in unlimited categories, and get reminders, statistics and search, without needing any other tracking app.

Logg ran on iPhone, on the desktop through an Electron app, and on the web. We published it on the App Store and Google Play.

#### The key decision

Privacy by default. People's logs belonged to them, which set Logg apart from social apps that shared everything. Logg started with some social features and later became a fully private diary.

#### What I built

As technical founder, I built the native iOS app and the Electron desktop app, worked on the backend and APIs, introduced MongoDB, and designed the subscription features.

I built the core platform on cloud infrastructure, with Node.js, MongoDB, Redis and Firebase behind the Swift iOS app, and wrote the feed algorithms for the early social features. As co-founder I also ran more than 10 investor meetings and over 50 product demos.

[Image: The iOS app, from its App Store listing.]

[Image: Logg Desktop, the Electron app.]

#### Recognition

Apple featured Logg as App of the Week in August 2018, and Product Hunt made it App of the Day in July 2018. By my own estimate it reached about 25,000 users.

The App Store featured it in the US, in Turkey and in several Asian countries.

### DIGIEGGS: my first company

https://hakanbaybas.dev/projects/digieggs

Founder · 2014, 2015, 2016, 2017, 2018

Founded a software company in Izmir: 15 clients, 20+ projects and our own products, with a team I trained from junior level.

Tags: Swift, React Native, React, Node.js, C#, Pharma

#### Why I started it

As an SAP consultant, I kept meeting businesses that needed custom software SAP work could not provide. I also wanted to own products, not just deliver projects, and I had found a specific gap in pharma.

I founded DIGIEGGS in Izmir in 2014. We built bespoke web and mobile apps for clients on a project basis, and our own products alongside them.

#### Products

Pharma sales teams worked from paper and spreadsheets, with no cloud or mobile app to track performance. That became Cosmegg, a CRM for pharmaceutical companies, sold directly and through our consulting clients. Our products that earned revenue included Cosmegg, Advegg (MDM for pharmacies), Convegg (congress and fair management) and Thermoegg (an IoT temperature and humidity tracker). Convegg and Thermoegg were sold to new owners. Cosmegg and Advegg were ended when their revenue no longer justified them.

[Image: Two of our own products: Cosmegg and Convegg.]

#### Team and clients

I hired junior developers and trained them, and the team grew to between 10 and 20 people. Many went on to lead teams; one grew from junior to CTO. We served 15 clients, including Yaşar Holding, where we built web and mobile apps, an admin platform and SAP integration across several businesses.

Other clients included İnci Holding, Nurol Technology, Eczacıbaşı and Decktopus in the US. Work came through referrals and existing relationships, direct proposals, and inbound interest from our own products. On the Yaşar project I connected the SAP world with the new software, ran the work across several teams and made the trade-offs visible to the client. For the toughest clients I built core parts myself.

[Image: A few of the mobile apps from our portfolio.]

#### What I changed along the way

Over five years, scope changed, deadlines slipped and things broke in production. I answered with written scope and change control, smaller milestones with earlier demos, and contracts with clear limits on support and revisions.

The people are what I am proudest of. Many of the juniors I hired went on to lead teams, several founded their own companies, and some still work with me.

## Stack

Python, TypeScript, LLM evaluation, LangChain, LangGraph, Langflow, Embeddings, Vector search, Weaviate, pgvector, Fine-tuning, Learning to rank, MCP, OpenAI API, AWS, Cloudflare Containers, Durable Objects, Docker, Redis, DynamoDB, PostgreSQL, Node.js, Next.js, React, GraphQL

## FAQ

### What AI work do you do today?

I design and build an AI shopping assistant end to end at Rezolve Ai, from the design doc to production. I own the AI/ML behind it: catalog-aware conversation, retrieval and vector search, intent matching, attribute extraction, vision models for product images, and model routing. I also built a knowledge layer that lets merchants author and publish expert shopping knowledge for the assistant and an API.

### How do you know an AI change is actually better?

I measure it before it ships. Every change goes through an evaluation: latency comparisons, recognition tests and blind side-by-side conversation judges. If the evaluation doesn't say it is better, it doesn't ship, however good the demo looked.

### Do you use AI coding agents?

Yes. Coding agents made writing code easy for everyone, but they don't tell you what to build. They won't tell you whether it scales, how many users it can serve at once, how fast it responds, what the customer experience needs, or whether anyone will pay for it. That judgement comes from 19 years of shipping products that had to sell and earn revenue. It is what I bring on top of the agent, and it is why the code I ship with them holds up.

### Do you train and fine-tune models, or only call APIs?

Both. Model work is part of my role at Rezolve Ai: fine-tuning embedding, extraction and vision models on real labels, and learning to rank results from shopper signals. Model routing is part of it too: picking the right model for each step instead of one model for everything.

### Have you built agent systems?

Yes. At Writer I was tech lead for Agent Deployment, the platform that runs customers' AI agents inside Writer's Agent Builder. I architected per-tenant execution on Cloudflare Containers and Durable Objects, with MCP OAuth integrations for tools. Before building it I researched multi-agent orchestration, agent frameworks such as the AWS Strands SDK, and MCP as the standard for tool use.

### How long have you worked with LLMs?

Since 2023. At Commented I used the OpenAI API for summaries and tone suggestions. At Barkley AI in 2024 I used the Azure and OpenAI SDKs to generate summaries and action points from meeting videos. Then came agent infrastructure at Writer from 2025, and applied ML at Rezolve Ai from 2026.

### How do you take an AI feature from idea to production?

I start with who will use it and how we will know it works, and write that into a design doc. Then I build it across every layer it touches: the service, the data pipeline, shared libraries and deployment config. It ships only when an evaluation says it is better.

### How do you work with a customer's engineering team?

I listen before I build. I start with how their engineers actually work and what success looks like for them, find the use cases that would make the biggest difference, and prove them with demos. When something they need is missing, I tend to build it myself, and I bring what I learn back into the roadmap. I have done this in more than 200 sales, demo and customer success meetings at Commented, in client meetings at Barkley AI, with the customer's team at Upsolve, and with customer engineering and solution architects during Writer's rollout.

### How do you keep AI systems secure?

By designing for the worst case from the start. On Writer's Agent Deployment, each customer's agents ran in isolation, so no two customers shared a runtime. Customer secrets were encrypted at rest and decrypted only on demand, uploaded files were scanned with ClamAV before an agent could touch them, and Redis rate limiting protected the platform. I got SecOps, DevOps and the customer teams to agree on one design before we built it.

### Which languages, models and tools do you use?

Python and TypeScript day to day. On the AI side: embeddings, vector search, fine-tuning, learning to rank, LLM evaluation, MCP, and the OpenAI, Anthropic and Azure OpenAI APIs. On the platform side: AWS (I am an AWS Certified Solutions Architect, Associate), Cloudflare Containers and Durable Objects, Docker, Redis, PostgreSQL and DynamoDB.

### Can you lead an AI team, or are you an individual contributor?

Both, and I do them at the same time. At Writer I was tech lead and team lead for a team of 4. At Decktopus AI I led the front-end team and still made 168 commits myself, and the team delivered 90% of planned sprints on time. Earlier I led development teams of up to 12 people across time zones. Today at Rezolve Ai I am hands-on as an individual contributor.

### What kind of work are you looking for?

Hands-on AI engineering where product judgement matters: building applied AI that real users depend on, and leading the people who build it with me. My background as a founder and team lead is what I bring to that work, not where I am heading back to.

### Where are you based, and can you work in the UK?

I live in London and I am a British citizen, so I can work in the UK without sponsorship.

### How do you approach an unfamiliar problem?

In four steps. First, I clarify what success looks like for the business. Second, I talk to the people closest to the problem: customers, users and the team. Third, I map the system and its constraints. Only then do I list what is missing.

### How do you decide what to cut when time or money is tight?

By customer impact. I keep whatever changes the customer's outcome most and cut what only adds polish.

### How do you handle disagreement?

I look for the root cause of the objection, because it often points to a real constraint I missed. Then I write the options and their trade-offs down, so the decision is clear and not personal.

### Do you mentor engineers?

Yes. I give direct, specific feedback, let people own their work and explain the bigger picture behind decisions. One developer I hired as a junior grew to senior, team lead, manager and then CTO. Many others went from junior to senior, and some from senior to team lead.

### What have you founded?

Two companies: DIGIEGGS, a software company in Izmir (2014 to 2018), and AlmondSpring Software Solutions in London (2019 to 2023). I also co-founded two startups: Logg, a private journal app, and Commented, a collaboration tool for software teams and their clients.

### Which languages do you speak?

Turkish (native) and English (full professional).

## Contact

- Email: baybashakan@gmail.com
- LinkedIn: https://www.linkedin.com/in/hakanbaybas
- GitHub: https://github.com/bybash
- CV (PDF): https://hakanbaybas.dev/cv.pdf
