Skip to content
View nilusitechtigris's full-sized avatar

Block or report nilusitechtigris

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
nilusitechtigris/README.md

Hi, I'm Deniz 👋

Software & AI engineer in Budapest. I test what AI actually does on real hardware, and publish the numbers, including when it fails.

By day I build AI agents, agent harnesses, MCP servers and data products. Before that: motor-control software, camera inspection, sensor calibration and test rigs. That's why I trust a measurement more than a demo.


🔥 What I'm building

Prometheus: bring the fire home

A native Mac app for running open AI models locally, with a harness that helps small models do more: sandboxed code, trusted search, checks in code, a 3-agent council and a coding agent. No account, no cloud.

Same 1.7B model, same 100 questions: 24 → 62 with the harness switched on.

The same question asked twice: the plain model answers 14, the harness answers 36, and code confirms 36

Silicon AI News: AI news with the evidence attached

Every story links its sources, company claims are labelled as claims, and corrections are public. Also available to AI assistants as a free MCP server.


🧪 Lab notes

One question per experiment. Every result comes with its raw data and a way to reproduce it.

When Question Answer Data
Nov 2026 How hard is my 100-task benchmark for a 27B 4-bit model? Qwen3.8-27B passed 79 of 100 tasks, all graded by code (no LLM judge). Ready for Qwen4. Qwen3.8_VS_Qwen4
Nov 2026 How fast are 20 open LLMs on a 24 GB Mac mini M4 Pro? tokens/s ≈ 220 ÷ model size in GB. MoE wins: gpt-oss:20b runs at 52.5 tok/s. benchmark
Nov 2026 Can a better harness beat a bigger model? Yes: a 1.7B model went from 24 to 62 / 100 and beat a model twice its size. Prometheus
May 2026 Does training a 1.7B base on DeepSeek R1 traces make it reason better? No. GSM8K fell 79% → 67% and answers got 4–8× slower. It learned the style of reasoning, not the skill. tiny-reasoning-qwen3

Generation speed versus memory use for 20 local LLMs on a Mac mini M4 Pro


🏆 Hackathons

  • 1st place: recipe-to-cart assistant for ALDI International IT Services (AI Klub hackathon)
  • 2nd place: zero-token knowledge graph for ABAP developers (SAP hackathon, Budapest)

🛠️ Stack

Python · TypeScript · Swift · ABAP · LangChain / LangGraph · MCP · Ollama · MLX · llama.cpp · SAP BTP & Joule Studio

📫 Elsewhere

LinkedIn · Silicon AI News · Happy to talk about local LLMs, agent harnesses and honest benchmarks.

Pinned Loading

  1. Prometheus Prometheus Public

    Native Mac app for running open LLMs locally (Ollama, MLX, llama.cpp, LM Studio), with a harness that makes small models do more: sandboxed code, trusted search, checks in code, a 3-agent council a…

    Swift 1