kendine mühendis
Software Engineer · İzmir

Software engineerbuilding AI systemsthat leave the notebook.

Software Engineer at Probel Yazılım, building submodules for a hospital information system. Independent AI/ML builder the rest of the time.

I work end to end — from a backprop loop in NumPy to a production Electron + WebSocket app. The part I care about most is the space in between: making models reliable, useful, and shippable.

now · TR--:--:--
focus · shipping GIDEON safety gate
press⌘Kfor commands
pipelinescroll to trace →
NumPy backpropPyTorch modelRAG pipelinePython backendWebSocketElectron UINumPy backpropPyTorch modelRAG pipelinePython backendWebSocketElectron UINumPy backpropPyTorch modelRAG pipelinePython backendWebSocketElectron UI
gideon@local~--:--:--
02/Story
01 / Math

Gradients, matrices, and a backprop loop written by hand.

I started close to the metal: gradients, matrix operations, loss curves, and backpropagation written by hand.

forward passlossbackward passweight update
backprop.py
W = np.random.randn(n, m)b = np.zeros(m)z = X @ W + ba = relu(z)dz = da * (z > 0)dW = X.T @ dz W -= lr * dW
loss↓ converging
02 / Models

CNNs, CIFAR-10, and the discipline of debugging models.

Then came CNNs, CIFAR-10 experiments, PyTorch, TensorFlow, and the practical discipline of debugging models instead of only training them.

conv blocksoverfit probeaugmentationeval
cnn · cifar-10
conv
conv
pool
conv
fc
softmax
train acc
val loss
batches
03 / Systems

The bridge from model behavior to software behavior.

The interesting part became the bridge: turning model behavior into software behavior.

API nodesbackend servicesWebSocketUI panel
topology
clientapiwsworkermodelstore
panel
connectedidle
$ run --verify
04 / Product

Dependable, observable, and safe to use.

A model is only useful when the surrounding system makes it dependable, observable, and safe to use.

safety gateverificationlocal assistant UIapproved / blocked / verified
safety gate
intent parsedapproved
safety gatehold
command scopedverified
outcome checkedverified
assistant ui
user: delete the temp files
gate: 3 files · confirm?
approveblock
verified · 3 removed
scroll to advance
§ III · 03Currentlyin role · ongoing
03Currently

Software Engineer · İzmir

Software Engineer at Probel Yazılım, building submodules for a hospital information system. Independent AI/ML builder the rest of the time.

in role · ongoing

Every state change is a durable five-step unit — not a best-effort write.

Domain
Hospital Information System / HBYS
Software that touches real clinical workflows.
Focus
Submodules, backend logic, database-driven workflows
Module-level ownership inside a larger enterprise system.
Environment
Enterprise software, healthcare operations, production constraints
Where 'mostly works' is not a category.
record · flowtxn
01BEGIN
02INSERT INTO patient_record (…)
03UPDATE module_state SET …
04AUDIT log_write(…)
05COMMIT
·By the numbers

Constraints, not vanity metrics.

Each number below is a design constraint a project actually had to satisfy — not a growth dashboard.

0
shipped / building
projects across ML, web & enterprise
01
0
GIDEON pipeline steps
wake word → verified result
02
0%
destructive actions gated
nothing destructive runs unchecked
03
0
cloud deps for local STT
audio never leaves the machine
04
0+
languages in daily use
SQL · C# · Python · TS · C/C++ · Java
05
0
POS in daily production
paying client · live restaurant
06
§ IV · 04Selected workfour cases · 2023–24
04Selected work
§

Four projects that trace the line from math to shipped software.

Not a portfolio of demos — each entry below carries a real constraint, a real problem, and a real outcome. Click any card to open the case.

A note on selection: these four were chosen because each forced a different failure mode — a model that wouldn't generalize, a system that couldn't stay local, a workflow that couldn't fail silently. See fig. 4–7.

01

GIDEON

ongoing

Wake-word desktop AI assistant

2025100%actions gated0cloud deps

A graduation project built as a local-first desktop assistant. Speech is transcribed locally with faster-whisper. The assistant can edit files, run commands, research the web, and verify its own outcomes behind a safety gate for destructive actions.

Role
End-to-end engineer
Outcome
Graduation project · local-first action system with safety gate
Stack
PythonElectronWebSocketfaster-whisperAgentic workflow
Interesting problem

Designing an assistant that can act on the local machine without becoming careless.

02

Turkish RAG Platform

ongoing

Retrieval-augmented generation with a Turkish NLP pipeline

2025+38%chunk recall~1.2slatency

Built around document ingestion, preprocessing, chunking, embedding, retrieval, and answer generation. The focus is not only asking questions over documents, but making Turkish text retrieval behave reliably.

Role
Builder
Outcome
Working retrieval pipeline tuned for Turkish text
Stack
PythonFastAPIEmbeddingsVector storeLangChain
Interesting problem

Turkish retrieval is not just English retrieval translated — morphology changes chunking and matching.

03

Restaurant POS

shipped

A production POS system shipped to a paying client

202499%+uptime120+tables / shift

Designed for real restaurant operations, not a demo scenario. The work included practical UI flows, data persistence, order handling, and client-facing reliability.

Role
Full-stack
Outcome
Shipped to a paying client · in daily use
Stack
C#.NETSQL ServerDesktop UI
Interesting problem

Real restaurant flow under live service pressure — tables, splits, voids.

04

Deep Learning From Scratch

learning

Backpropagation by hand in NumPy, then CNNs on CIFAR-10

20248+layers~78%acc

A first-principles learning project focused on understanding the math before relying on frameworks. Backprop written from scratch, then convolutional networks trained on CIFAR-10.

Role
Self-directed study
Outcome
Working backprop + CNN training pipeline from primitives
Stack
NumPyPyTorchCIFAR-10Matplotlib
Interesting problem

Understand gradients, loss surfaces and update dynamics before the framework hides them.

§ V · 05GIDEON · deep dive8 nodes · local-first
05GIDEON · deep dive
§

GIDEON is not a chatbot. It is a local action system.

The project explores what happens when an assistant is allowed to move from conversation into controlled execution.

  1. 01
    Wake word
    listening
  2. 02
    Local speech transcription
    transcribing
  3. 03
    Agent reasoning
    reasoning
  4. 04
    Tool selection
    planning
  5. 05
    Safety gate
    gate: hold
  6. 06
    Command / file / research action
    executing
  7. 07
    Outcome verification
    verifying
  8. 08
    User-facing result
    done
step 01 / 08listening

Wake word

Listening stays local. The assistant wakes on a small keyword model without sending audio anywhere.

scroll to advance the flow
gideon~listening
↳ flow running
principles

Local-first where possible

Audio, transcription and reasoning stay on-device. The cloud is optional, not load-bearing.

Destructive actions require a safety gate

Risk is classified before anything runs. High-risk actions wait for explicit confirmation.

Actions are verified, not assumed

After execution, the assistant checks the result. Failure is feedback, not silence.

UI and backend over local WebSocket

The Electron shell and the Python brain talk over a local socket — no HTTP round-trips per action.

system idea
Wake word
Local transcription
Intent routing
Action planning
Safety gate
Execution
Verification
§ VI · 06Stack8 categories · what I reach for
06Stack
§

Tools I use across database-heavy systems, backend services, AI/ML workflows, and product interfaces.

Not a badge wall — a working inventory. Hover any item to see where it actually shows up.

01
Database
Relational systems, healthcare & POS data, query tuning.
02
Data
Schema discipline and access paths.
03
Backend
Services that survive production traffic.
04
Web
Interfaces that feel honest to use.
05
AI / ML
From primitives to trained models.
06
LLM
Making language models usable.
07
Infrastructure
Shipping and operating.
08
Languages
What I reach for first.
ctxhover an item to read its context
§ VII · 07Notesfive short essays
07Notes
§

Short technical notes on building models, systems, and the layer between them.

Not long-form essays — working notes. The kind of thing I write to make sure I actually understood something.

On the form: these notes are deliberately short. A note exists when an idea survives being written down — not when it reaches a publishable length. See fig. 8.

shelf
mlfirst-principlesnumpy

Why I wrote backprop before using the framework

Why handwriting the chain rule once makes every framework debugging session shorter.

Frameworks hide the part that actually matters: where the gradient comes from and why the update works.

kendine mühendis
Software Engineer · İzmir10 min
gideonagentssafety

What makes a local AI assistant dangerous — and useful

Capability isn't the risk. Classification is.

The moment an assistant can run a command, the question stops being 'can it?' and becomes 'should it, and who checks?'

kendine mühendis
Software Engineer · İzmir11 min
ragnlpturkish

Turkish retrieval is not just English retrieval translated

Chunking is downstream of morphology. Retrieval is downstream of chunking.

Agglutination changes where a chunk ends. Naive splitting breaks the very thing retrieval depends on.

kendine mühendis
Software Engineer · İzmir13 min
probelhbysreliability

What hospital software teaches about reliability

Constraints become instincts once a misplaced decimal is an incident.

When the data is a patient record, 'mostly works' is not a category. Constraints become instincts.

kendine mühendis
Software Engineer · İzmir12 min
engineeringproduction

The boring parts that make AI systems shippable

Logging, retries, idempotency, timeouts — the layer most prototypes skip and every production system needs.

Logging, retries, idempotency, timeouts. The unglamorous layer that decides whether a model is a feature or a liability.

kendine mühendis
Software Engineer · İzmir12 min
engineeringreliabilityresilience

Designing for failure

Timeouts, retries, circuit breakers, idempotency, graceful degradation — the patterns that decide whether a system degrades or dies.

Failure is not an exception. It is the default state of a distributed system. The question is how it fails.

kendine mühendis
Software Engineer · İzmir13 min
engineeringabstractiondesign

The cost of abstraction

When abstraction helps, and when it hides what you need to see.

A good abstraction hides detail you do not need. A bad abstraction hides detail you do need. The difference is everything.

kendine mühendis
Software Engineer · İzmir12 min
llmagentstrends

Context engineering for LLMs

The 2025 shift from prompt engineering to context engineering — retrieval, tool use, and agent loops.

The interesting work in 2025 is not writing a clever prompt. It is engineering the context the model sees at inference time.

kendine mühendis
Software Engineer · İzmir14 min
llmedgeprivacytrends

Local-first AI and the edge

On-device models, privacy, latency, and the quiet shift away from cloud-only AI.

Cloud-only AI made sense when models were large and GPUs were rare. Both premises are changing.

kendine mühendis
Software Engineer · İzmir13 min
aiimagebeginner2026

From zero to your first AI image — a 2026 beginner's guide

You have never written a prompt. You do not know what a model is. By the end of this you will have generated images you are proud of.

No design background, no software, no jargon. Just you, a browser, and the strongest image models of 2026. Step by step, from the very first click.

kendine mühendis
Software Engineer · İzmir14 min
aidesignbeginner2026

Designing without Photoshop — a 2026 beginner's guide to AI design

Posters, logos, social media posts, product mockups — all without opening Photoshop, all without a design background.

Photoshop is a professional tool that takes years to learn. AI design tools in 2026 let you skip it. Here is the complete beginner's path.

kendine mühendis
Software Engineer · İzmir13 min
aivideobeginner2026

From zero to your first AI video — a 2026 beginner's guide

You have never edited a video. By the end of this guide, you will have generated a video from a single sentence.

In 2026, you describe a video in words and the model creates it — sound, motion, and all. No camera, no editing software, no experience needed. Here is the complete path.

kendine mühendis
Software Engineer · İzmir14 min
aiwebsitebeginner2026

Building a website without coding — a 2026 beginner's guide

You do not know HTML, CSS, or JavaScript. You do not need to. In 2026, you describe your website and AI builds it.

No code, no templates, no drag-and-drop. Just describe what you want in words and watch a real, deployable website appear. The complete beginner's path, from idea to live site.

kendine mühendis
Software Engineer · İzmir14 min
aiprompt-engineeringprofessional2026

Prompt engineering, professionally — a 2026 practitioner's guide

Prompts are not magic words. They are a specification language. Here is how a 2026 engineer writes prompts that are reliable, reusable, and model-aware.

From single-shot wishes to structured, evaluated, model-routed prompts. The complete prompt-engineering skill, the way working AI engineers actually practice it in 2026.

kendine mühendis
Software Engineer · İzmir18 min
aiecommercesupabasekirofullstack2026

Build a real e-commerce store from scratch — Supabase + Kiro, a 2026 beginner's guide

No coding background. By the end you will have a live online store with a Postgres database, secure auth, a product catalog, a cart, and a checkout — built spec-by-spec with Amazon Kiro and Supabase.

Frontend and backend, together. A complete, production-shaped e-commerce site built the way a 2026 engineer actually builds one — with an agentic IDE, a Postgres backend you control, and prompts written like specs.

kendine mühendis
Software Engineer · İzmir22 min
§ IX · 09Colophonon the making · rev 0.3.2
Colophon
A note on the making of this folio.

This page is a single, hand-set folio — written, drawn, and tuned end to end. No template, no analytics, no tracker. Just the path from model to product, set in ink on warm paper.

Typeset
SF Pro — Display & Text
Set on a warm paper ground at 1.6 line-height, with tabular numerals throughout.
Structure
Single folio, eight chapters
Hero · Story · Role · Work · GIDEON · Stack · Notes · Contact — pinned scenes via GSAP ScrollTrigger.
Accent
Coral / clay (#A84C28)
One deliberate accent — used for active states, pull-quotes, marginalia, and the scroll hairline.
Build
rev 0.2.1 · build a1f3c
Next.js 16 · TypeScript · Tailwind v4 · shadcn/ui. Hand-built, no templates.
Words
body text in <main>
Read time
~200 wpm, technical prose
Footnotes
marginalia cited inline
Figures
fig. N references in footnotes
Set in ink
An imprint of this folio.
imprint · hand-set
typeSF Pro Text & Display
paperwarm cotton · 80 lb
folios— bound · 1 sheet
read time≈ — min · 200 wpm
editionfirst · 2024
revision0.3.2 · build a1f3c
composed bythe author · kendine · Istanbul

Pressed in browser, not on a press. The grain you see is a fractal-noise SVG overlay fixed to the page background; the ink you read is the system SF Pro stack at -0.011em. No images of paper were used.

Footnotes & Figures
An appendix of marginalia cited in this folio.
  1. fig. 3

    Record-flow transaction — a five-step auditable unit: BEGIN → INSERT → UPDATE → AUDIT → COMMIT. Every transaction that changes state leaves an audit trail by construction, not by convention.

    cited in · Role
  2. fig. 4–7

    On the discipline of selection. The four projects below are not the most recent — they are the ones whose lessons survived being written down. Failure modes recorded, not buried.

    cited in · Work
  3. fig. 8

    On the form of the working note. A note exists when an idea survives being written down — not when it reaches a publishable length. The five notes above are deliberately short.

    cited in · Notes
Glossary
12 terms cited in this folio.
RAGRetrieval-Augmented Generation
Feeding a model retrieved context from a corpus before it answers, so it grounds claims in source documents.
MCPModel Context Protocol
An open protocol that standardizes how models discover and call external tools, files, and data sources.
HISHospital Information System
The operational software backbone of a hospital — patient records, billing, clinical workflows, audit trails.
RLSRow-Level Security
A database-level policy that restricts which rows a session can read or write, enforcing tenant isolation at the data layer.
CNNConvolutional Neural Network
A neural architecture built from convolutional filters — historically the workhorse for vision before transformers arrived.
ASRAutomatic Speech Recognition
Transcribing spoken audio to text. Locally, usually a streaming model running on-device for latency and privacy.
TTSText-to-Speech
Synthesizing natural-sounding speech from text — the reverse direction of ASR.
LoRALow-Rank Adaptation
Fine-tuning a model by training small low-rank adapter matrices instead of the full weight set — cheap, swappable.
FAISSFacebook AI Similarity Search
A library for fast dense-vector similarity search at scale — the index layer of most retrieval systems.
WSWebSocket
A persistent bidirectional TCP connection — the transport layer for real-time streaming between client and server.
SQLStructured Query Language
The lingua franca of relational databases. Used here for audit, transactions, and row-level policies.
TRTurkish (language)
Agglutinative, morphologically rich — retrieval and tokenization require language-specific handling, not just English defaults.
Take this folio offline

A clean manuscript PDF — no chrome, no tracking, ink on paper. Uses the same print stylesheet the page already ships.

§ VIII · 08Contactcorrespondence · open
08Contact

Building useful AI systems, one reliable interface at a time.

Based in İzmir. Open to software engineering, AI engineering, backend, and product-focused ML opportunities.

— end of correspondence —Yours, in pursuit of reliable interfaces.
··