Hi, Iâm Halil.
Turkish NLP models, privacy tools and AI workflows built around your own data.
Teams
Choose the right task, build a pilot and measure the result.
Read the roadmap
KVKKBSEBY
Mask and classify Turkish personal data with models that run on your own machines.
See the models
Engineering
Understand agent activity, permissions and context costs from session logs.
See the tools
Tell me what your team does today and where it gets stuck.
Four small models for Turkish text: three for personal data, one for prompt injection. Run them locally; explore their weights, benchmarks and demos.
Masking, classification and the benchmarks behind them.
Evaluation sets and demo
halilneed/turkish-kvkk-classification-benchmark halilneed/turkish-bseby-classification-benchmark halilneed/turkish-prompt-injection-false-alarm-benchmark halilneed/turkish-pii-detection-demo halilneed.agency/masking-toolThe PII model reads an instruction and a text, and writes the text back with the requested fields replaced by labels. The same weights mask everything, only the fields you name, or everything except them.
Metindeki tÞm kiÅisel verileri uygun etiketlerle maskele.
mÞÅteri AyÅe YÄąlmaz tc 12345678901 tel 0532 111 22 33 e-posta demo@example.com.
mÞÅteri [AD] tc [TCKN] tel [TEL] e-posta [EMAIL].
A 270M model improved from 0.740 to 0.944 exact match with targeted synthetic data and averaged weights. Read the training notes for the recipe and failure cases.
Open-source tools, mostly for people who work with coding agents.
6 Claude Code pluginsMIT
agent-blackbox, skillbench, scar, gardener, painradar and devpersona. Five of them read the session logs already on your disk and report what the agent did, with no network calls. painradar reads public complaints on the web.
All six plugins
Node.json npm
A local dashboard that reads session transcripts, auto-memory and instruction files to show what your agents know, flag context at risk, and move context between sessions, machines and agents.
GitHub
MCP server
Pulls live data from Reddit, GitHub Trending and Hacker News and recommends tools for solo developers.
GitHub
Agent skill19 rule files
Teaches a coding agent to write correct HyperFrames (HTML-to-video) compositions.
GitHub
Browser quiz3 languages
Five prompt scenarios, scored on clarity, creativity, technical depth and strategy. In Turkish, English and Spanish.
GitHub
Ready-made kits for legal, compliance and AI security teams. About the legal and compliance kit.
Kit$19
8 task cards, a data rule, 6 ready prompts, an Excel workbook and an editable Word AI usage policy, mapped to the KVKK guide of 24 Nov 2025.
Gumroad · TÞrkçe sÞrÞm
Test pack$39
Ready-to-run attack prompts for LLM apps and agents (direct, indirect, encoded and hidden-Unicode cases), a scoring sheet and a remediation checklist.
Gumroad
Four interactive AI guides for teams, available in English and Turkish with PDFs.
Guide 14 stepsPDF
From one task to a pilot you can measure: the right first task, scores by slice, shadow mode.
Try it
Guide 24 stepsPDF
Rules or a model, six kinds of model and four questions that choose between them.
Try it
Guide 34 stepsKVKK
What to classify, mask and keep in-house before text reaches a model.
Try it
I write on Medium about what the logs and benchmarks show.
in Turkish
How a 270M Turkish PII masking model went from 0.740 to 0.882 on a public 1,000-row benchmark with 24,000 targeted examples, and the two slices where it got worse.
Read on Medium
One team cut its daily incoming requests from 100 to 35 without buying an AI tool, by charting its own data first.
Read on Medium
Across 128 sessions there were 376 activations. Six skills did all the work, and 1,447 lines of instructions were never opened by any agent.
Read on Medium
In twelve projects the median was about 6,000 tokens per request, and one rule turned out to be working because it never fired.
Read on Medium
Claude Code writes every session to your disk as JSONL. Mine showed 35 tool calls per prompt, 49 denials and other things that did not flatter me.
Read on Medium
I work in data analysis, product management and software in Istanbul. I build Turkish privacy models and open-source tools for coding agents.
The models, tools and articles here share benchmarks, evidence and known limitations.
