Mustafa DOGAN

• I establish what a system actually does before changing it, and I write that down. Undocumented behaviour becomes a named contract in the repository that the dependent code cites.
• I automate against a platform's own interfaces rather than clicking through its console. Deploys are scripts, a fix to a live flow is an idempotent patch, and production is mirrored to disk so drift can be diffed.
• I build the safety rail before the bulk action. Mandatory preview, a drift check that halts if the target set moved more than 10 per cent since preview, hard run caps, per-record audit rows and a matching rollback script.
• I treat rules and scenarios as data so one executor serves many cases, rather than writing a script per case.
• Where a language model is involved, the deterministic layer stays in charge. A generated query does not run until it passes a cost estimate, and safety verdicts come from rule-based gates rather than from the model.
• I choose the read-only variant when the mutating one would be easier. An accuracy audit stops before it places anything.
• I write the memo that argues why the obvious approach will not work, before building the one that does.
• I record negative results. Where a decode was incomplete or a run produced untrustworthy output, that is written down next to the code.
• I build an escape hatch out of any reverse-engineered path, so a sanctioned API can replace it later without touching the callers.
• I make runs reproducible and auditable: query text, row counts, bytes billed, prompt and model versions kept for later inspection.

158 on LinkedIn · most recent first

• Gemini for DevelopersGoogle DeepMind · May 2026ID 10APGA9A4OPT
• Exploring AI with Google AI StudioGoogle DeepMind · May 2026ID SCTZ55P5J1JS
• Getting Started with Google Gemini APIGoogle DeepMind · May 2026ID I03582Z7VBMH
• Optimization Methods in Business AnalyticsMIT OpenCourseWare · Apr 2026from LinkedIn
• Computational Neuroscience with NeurobloxMIT OpenCourseWare · Apr 2026from LinkedIn
• Understanding the Enterprise Systems EnvironmentSAP · Dec 2025ID CAIEK88031MM
• SAP Professional FundamentalsSAP · Dec 2025ID YA4CHNYOQ5Q5
• Microsoft Business AnalystMicrosoft · May 2025ID PKL37O0PD75D
• Power Platform in Business AnalysisMicrosoft · May 2025ID C2KGRIB2F41N
• Requirements Gathering in Business AnalysisMicrosoft · May 2025ID J0TMPGMUQFF9

• Dialler rule engine

  Query compilation

  Campaign and contact-list management for contact-centre tenants. An operator writes conditions over the live columns, either in a visual builder or as a SQL WHERE clause, previews the matching set, and then applies bulk actions to it: priority, urgency, agent assignment, disabling contacts. Every run is audited per contact, and a rule can run on a schedule instead of by hand.

  The condition tree is compiled into the platform's own filter parameters, and only the leaves of a top-level AND chain are pushed down, so a wrong push can never widen the set the action runs against; the remainder is evaluated locally. Which parameters are supported was not assumed: the API ignores unknown parameters silently, so one counts as supported only after a deliberately non-matching value is shown to shrink the result count.

  Next.js · TypeScript · Postgres · Drizzle · OAuth2 + PKCE · On-device LLM

• Agentic Report Studio

  Second place, AI hackathon

  Reporting product for tenants whose requirements fall outside the platform's standard reports. Users describe what they need and get a governed query against the warehouse. Second place at the AloTech AI Hackathon, July 2026, and in production afterwards with real tenants.

  No generated query executes until it passes a dry-run cost estimate against configured limits, so an expensive mistake is refused rather than billed. 67 ordered migrations, with separate unit, API, end-to-end and security suites.

  Next.js · TypeScript · BigQuery · Postgres · Tool-calling agent

• Reporting client for iOS and Android

  Mobile delivery

  Phone client for the reporting product, built to match the web app screen for screen rather than offer a reduced version of it: conversational reporting with streamed answers, saved reports, and dashboards with cross-filtering and drill-down. Routes are deep-linkable and mirror the web paths, and the interface ships in Turkish and English.

  Sign-in reuses the platform's existing OAuth flow inside an in-app web view and lifts the resulting session cookie into the native cookie store, so the phone client talks to the same backend with no second authentication path to keep in step. The 2,122 interface strings were taken from the web app verbatim rather than translated again, which is what keeps the two clients saying the same thing.

  Flutter · Dart · Riverpod · go_router · Material 3

• Curyy

  A product owned end to end

  Native Android product I build and run myself, aimed at the Philippines: it logs what you eat and gives allergy, calorie and budget guidance in a floating overlay. With per-app consent it reads the screens of allow-listed food delivery apps as text, so an order becomes a logged meal without anyone typing it in. It writes nutrition back to the phone's health platform, tracks spending against a budget, and exposes its own MCP server so an assistant can read the diary and log a meal.

  Text, never screenshots, and only from apps the user has allow-listed one at a time — a tool that reads your screen has to be narrower than it is capable of being, or nobody should install it. The screen classifier has a parity test so the on-device path and the reference path cannot drift apart, and MCP tokens are read-only unless a write scope is granted explicitly.

  Kotlin · Android · Jetpack Compose · Health Connect · MCP · Postgres

• CRM to contact-centre bridge

  Integration middleware

  Lets a sales team push leads and deals out of their CRM straight into a dialling campaign, with their own field mapping, triggered from a workflow action, a button on the record, or automatically when the record changes. One codebase with the CRM side and the contact-centre side both behind provider interfaces, so a second CRM or a second platform is an implementation rather than a fork.

  A CRM app cannot hold a secret. It is configuration, interface extensions running in the vendor's sandbox, and HTTP callbacks the vendor posts to, and the CRM never talks to the contact-centre platform directly. This backend is the part that can be trusted with credentials: it brokers the OAuth, holds each customer's platform credentials encrypted, receives the callbacks and makes the calls. Getting that boundary right was the design decision.

  TypeScript · Turborepo · Next.js · Postgres · OAuth2 · HubSpot

• May reklamo

  A product shipped end to end

  Public complaint platform for Philippine brands, government offices and public authorities, in English and Tagalog. Complaints and replies are written in a rich-text editor, supporting files are attached as screenshots or documents, and sign-in is with Google. Built and run outside work, on my own.

  Bilingual throughout rather than translated at the edges, because a complaint written in Tagalog has to read as the primary version and not as a secondary copy of an English one.

  Next.js · TypeScript · Postgres · Prisma · Vercel Blob · TipTap

• Call search and bulk recording export

  Multi-tenant product delivery

  Multi-tenant tool for finding calls with the full filter set an operations team actually uses, then downloading the recordings — one at a time or an entire result set — or delivering them to an FTP, FTPS or SFTP target. Any tenant onboards itself: type the tenant address and the branding switches to match; sign in with the tenant's own single sign-on, then add your own API credentials.

  Nobody's credentials are shared and none are held centrally. Each person brings their own, so what they can see is exactly what their credential allows, and there is no account in the middle that would quietly hold more access than the person using it. Sign-in is the tenant's own OAuth authorization code flow with PKCE, so the session key never reaches the browser.

  Next.js · TypeScript · Postgres · OAuth2 + PKCE · SFTP

• Softphone shell for Android

  Native and web integration

  Android application that hosts the browser softphone agents already use, so a phone can be the working device. The call session stays alive in the background, the microphone is wired through, incoming calls raise native notifications, and dialling a number from another app hands off into the softphone.

  The shell implements no telephony at all. Anything that would need a new native release every time the web softphone changes belongs in the web app, so the native side stays a policy-driven host and call events cross one documented message bridge.

  Flutter · Dart · Android · WebView · WebRTC

• Demo builder

  Pre-sales tooling

  Tool for showing a click-to-connect feature in context. A partner names the page they want it on; the tool takes a static copy of that public homepage, injects the script, and publishes it under a demo URL with a banner linking back to the original site, so a conversation can happen over the partner's own page instead of a generic sandbox.

  Only public homepages are copied and every published page carries the banner and the link back, so a demo cannot be mistaken for the real site.

  Next.js · TypeScript · Postgres · NextAuth

• Browser test platform

  Test automation

  Test platform for an ordering integration. Scenarios are stored and edited centrally rather than checked in as scripts, run in real browsers on demand, watched live while they run, and their artefacts filed automatically. Access is by role, and a new account has to be approved before it can run anything.

  Run output is streamed to whoever is watching, and each run keeps its own artefacts, so a failure stays diagnosable after the fact instead of only being reproducible.

  Next.js · Playwright · Postgres · Google OAuth · Server-Sent Events

• API Explorer

  Reverse engineering, as a tool

  Browser extension that watches the requests a web application makes while you use it and turns them into a searchable collection: every endpoint, its payloads and its headers, with any of them replayable. Published on the Chrome Web Store. It can hand a captured endpoint to an AI provider for analysis, using a key you supply yourself.

  This is the method I use, turned into a tool: rather than reading a system's documentation and hoping it is current, watch what it actually does and keep the record. The AI step takes your own key on purpose — a developer tool that captures live traffic has no business also shipping a service that sees it.

  TypeScript · Chrome MV3 · DevTools protocol

• Workflow-as-code toolkit

  Infrastructure as code

  Command-line tooling for the automation platform, so workflows are created, listed, renamed, audited and cleaned up from a script instead of by hand in the editor. Executions can be pulled back down for inspection when something failed overnight.

  140 workflows is past the point where the editor is the right interface. Renaming them to a convention, finding the archived ones and reading an execution's actual payload are all things you want to do to the whole set at once, which means the platform's API rather than its UI.

  TypeScript · Node.js · n8n REST API

• On-call schedule and alerts

  Operations tooling

  Application that puts the on-call rota and the open alerts behind the same sign-in the platform already uses, so whoever is on shift can see who is on call and what is firing without a second account and a second tool.

  Sign-in is the platform's own OAuth authorization code flow with PKCE, so no separate credential is created for the tool and access follows the account that already exists.

  Next.js · TypeScript · OAuth2 + PKCE · Opsgenie API

• Call flow documentation and build automation

  Configuration as code

  A written reference for every one of the 52 step types a call flow can contain — what each does, which fields it requires and which are optional — and scripts that build and export flows against it instead of clicking them together in the console.

  The field list came from working through each step type deliberately rather than from documentation, because a step that silently accepts a missing optional field is the kind of thing that only shows up in production. Once the reference existed, building flows from it was the easy part.

  Node.js · JSON Schema · Markdown

• Model Context Tool Inspector

  Agent tooling at the edge of the platform

  Browser extension for inspecting, monitoring and running the tools a page exposes to an agent through Chrome's experimental in-page tool API. Published on the Chrome Web Store. It needs a flag turned on because the API it targets is not shipped yet.

  Built against a browser API still behind a flag, which means the interface can move under it. That is the reason for building it: when this becomes the way pages talk to agents, the useful thing to have is the tool that shows you what a page is actually offering.

  TypeScript · Chrome MV3 · WebMCP

• Self-hosted automation platform

  Workflow automation

  Two n8n instances, development and production, carrying 140 workflows with 62 running: bridges between the contact-centre platform and CRMs, scheduled reconciliation, warehouse reporting, and 106 webhook endpoints other systems call into.

  272 of the 929 nodes are hand-written code, which is the honest measure of this platform's limit: the triggers, the branching and the scheduling are worth having, and the shaping in between usually has no node that does it. Development and production are separate instances so a half-finished flow cannot answer a live webhook.

  n8n · Docker Compose · Postgres

• RAG Lab

  Explaining a system by building it

  Teaching tool that walks through a retrieval pipeline one stage at a time — ingestion, chunking, embeddings, sparse index, retrieval, reranking, context packing, generation, validation — and puts BM25, vector, hybrid and hybrid-with-reranker side by side so the difference is visible. Explanations come in a beginner and a technical register.

  The embeddings and scores are deterministic mocks and the tool says so on the page. A teaching simulator that quietly used real model output would be unrepeatable between two people looking at it, and someone would eventually cite its numbers as a benchmark.

  TypeScript · Next.js · KaTeX

• Digital business card

  Front-end delivery

  A single page that hands over contact details, with a QR code for the certification list. Written by hand with no framework and no build step.

  Three files served as they are. Anything that needs a build to show a phone number is over-built.

  HTML · CSS · JavaScript

node counts read from the instances, not estimated · customers not named

• n8n operations assistant in Google Chat

  47 nodes

  Chat message, plus internal webhooks for the slow jobs

  Engineers run the automation platform by talking to it in the chat client they already have open. Ask in plain language and it lists the workflows, summarises one, reports which ones errored in the last day, explains why, and proposes a fix. Accept the fix and it patches the workflow, deploys it, activates it and offers to replay the payload that failed so you can see it pass. It also writes new workflows: it interviews you for the specification, researches the nodes it will need, and builds the flow.

  A planner model picks one of 10 tools rather than improvising, so every answer comes from a real call and not from the model's memory. What it refuses is deliberate: it never deletes a workflow, never prints a secret into a chat thread — credentials are referred to by name only — and works against one instance at a time. Anything slow runs on an internal webhook and posts its result back, because a chat client will not wait for a build.

  Google Chat · LLM planner · n8n REST API · Code · Webhook

• Caller context popup for the agent

  45 nodes

  Webhook on the inbound call, answered with HTML

  When a call arrives, the caller's number is looked up in the CRM and the agent gets a rendered panel: who is calling, their open deals, their notes, tasks and activities, their company, their stores, and their previous calls on the platform. If the number is not a known contact it falls back to matching the account instead of showing nothing. A second endpoint lets the agent write a note straight back into the CRM from that panel.

  Six CRM lookups and the call history run in parallel and are merged in five stages, because sequentially they would not return before the agent has to speak. The panel is built as HTML inside the workflow and returned in the response, so there is no front end to host for it. The token refresh happens in the same run rather than being cached somewhere that could go stale mid-call.

  Webhook · Zoho Bigin · Merge · Code · HTML response

• Trial administration as a single workflow

  51 nodes

  Webhook for the actions, hourly schedule for the reminders

  The whole administrative side of a trial programme: adding and removing tenants, registering users and approving them, keeping the notification list, reading usage totals out of the warehouse, and writing an audit line for each action. Registration is not automatic — an approval email goes out, and the user is only mailed once someone has approved them. An hourly pass works out which trials are due a reminder, checks their usage against the threshold and mails whoever needs mailing, then deactivates the reminder so it cannot fire twice.

  One webhook and a switch carry about 20 distinct actions, which is why 30 of the 51 nodes are code. The reminder branch does not trust its own schedule: it re-checks usage against the threshold before it sends, so a tenant who became active in the meantime does not get chased.

  Webhook · Schedule · BigQuery · Gmail · Code · Switch

• An authenticated dashboard with no application behind it

  25 nodes

  Three webhooks: the page, sign-in, and the OAuth callback

  A campaign dashboard that a person signs into and reads. The workflow serves the sign-in page, runs the OAuth handshake, sets the session cookie, queries the warehouse and renders the dashboard — as HTML for a browser or as JSON for anything else asking. An unauthenticated request gets a redirect or a 401 depending on what it asked for.

  There is no application server, no deployment and nothing to keep alive: a workflow is the entire service, sign-in included. It branches on what the caller wants back rather than assuming a browser, which is the only reason the same three endpoints can serve a page and an API at once.

  Webhook · OAuth2 · Session cookie · BigQuery · HTML response

• 10 endpoints for a conference showcase

  40 nodes

  10 webhooks, 10 responses

  The back end for a live demonstration stand: event logging, RSVPs, an outbound SMS through a gateway, a mocked payment step, an account lookup, a rendered report, spreadsheet setup, a dialler log — and a conversational turn against a hosted voice agent so visitors could talk to it.

  A spreadsheet is the whole database, which for something running for the length of an event is the right amount of infrastructure. The payment step is mocked on purpose and labelled as such, so nobody demonstrating it has to wonder whether it is charging anyone.

  Webhook · Google Sheets · SMS gateway · Voice agent API · Code

• Meetings reconciled against the CRM calendar

  20 nodes

  Webhook, on demand

  Recorded sales meetings are matched against what the CRM calendar says happened. Transcripts come from the recording service, events and the user list come from the CRM, and the workflow decides per meeting whether the matching event needs updating or does not exist yet and has to be created.

  Three sources are merged in sequence before anything is decided, because the match needs all three at once — the meeting, the calendar entry and the user who owns it. Deciding first and fetching later would have produced an event assigned to the wrong person.

  Webhook · GraphQL · Zoho CRM · Merge · Code

• Store records between a CRM and the platform

  15 nodes

  Webhook for the calls, hourly cron for the push

  A small API over store records: read one, search, create or update, delete, and the same operations in bulk for whoever is moving a whole list. An hourly job pushes what has changed into the contact-centre platform.

  Two paths on purpose. Callers get the immediate answer; the scheduled push is what makes the bridge trustworthy, because a bridge that only reacts to events silently keeps whatever it dropped. Bulk operations are separate endpoints rather than a loop over the single one, so a partial failure is reported per record.

  Webhook · Schedule · Switch · Code

• Scheduled report feed in XML over FTP

  15 nodes

  Schedule every 30 minutes, manual run kept

  Three reporting queries — queue performance, agent queue activity and agent system time — run against the warehouse every half hour, are converted to XML and delivered to an FTP drop, because XML over FTP was what the system on the other end could accept.

  The consumer's format decided the design, not ours. Each report is generated and delivered independently, so one failing query does not hold up the other two, and the manual trigger stays next to the schedule because the first thing anyone wants after a missed window is to run it again by hand.

  Schedule · BigQuery · XML · FTP

• CCaaS / VoIP
• SaaS platforms
• CRM / ERP
• Zoho CRM
• HubSpot
• REST APIs
• SQL
• n8n
• MCP / WebMCP
• Google Apps Script
• JIRA
• Selenium
• Python
• SAP
• Power Platform
• Cursor
• Vercel
• HTML / CSS / JS
• BigQuery
• Chrome extensions (MV3)
• Flutter
• Docker