Skip to content
View Gjusev's full-sized avatar
🏠
Working from home
🏠
Working from home

Highlights

  • Pro

Block or report Gjusev

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
Gjusev/README.md

AI product & forward deployed engineer · founder of mokka

I put AI into the systems companies already run, and publish the numbers that show whether it works.

Hugging Face · Kaggle · LinkedIn · mokka

3

products
in production

1,108

PyPI downloads
4 packages

676

downloads of the
VoxEntity-EU benchmark

21

likes on Hugging Face
8 models, datasets, demos
Live from PyPI and Hugging Face, checked daily · last change 2026-11-04

In production

  • Datenmaske · offline GDPR redaction of PDFs for law firms and authorities. Recall 75% → 85% on a 1,428-entity benchmark.
  • Klarbescheid · explains German administrative notices in plain language and flags likely errors. Unknown domain to shipped product in 6 weeks.
  • Nexary · knowledge platform inside a client's own network: hybrid search, reranking, OCR. Portfolio export.

Models & research

What it is Measured
MK-SH-V 260M visual decision model: a choice, or a refusal below a validated gate 0.839 on real held-out screenshots vs 0.731 for a ~15× larger model; trails it on receipts
MK-SH-1 140M text decision model with calibrated abstention 97.9% on numeric thresholds; calibration error 0.073 → 0.016
laya-triage support-ticket routing, 77 intents → 8 departments 89.8% on the full 3,080-ticket BANKING77 test
VoxEntity-EU does speech recognition catch the exact phone number? DE/FR best ASR gets 37.1% of slots exact, against a 90.8% ceiling
gate-calibrated act-or-ask gate for tool-using agents, on tau-bench wrong writes held at 1.7% (α 0.02), but no win over simple baselines

Each repo carries the sample sizes, confidence intervals and the point where the number stops holding.

Tools

  • nitelos · did the agent really do what it says? Checks the evidence, not the agent's report
  • voice-evals · replay calls, probe live voice agents, fail CI on a regression
  • note-lm · local research notebook where every claim keeps its source passage
  • hard-mode · skill-issue · i-didnt-write-this · tools for working with coding agents
More projects

Apps · HeizPro KI German voice agent for heating-service leads · Maqrivo meal planning with CP-SAT · Somatriq self-hosted wearable data · Studio Command Center · Falschinfo-Radar · Nexary website

Decision-model tooling on laya and Cloudflare Clef · laya-router · clef-router · laya-compactor · clef-compactor · laya-evals · clef-evals · laya-evals-gate · laya-phishield · laya-warden

Regression kits · locale-invoice-check invoice extraction in EN/DE/ES · vision-input-check vision APIs across image re-encodes


Open to AI engineering, full-stack product and forward deployed roles · EU or relocation · from November 2026 · LinkedIn

4+ years of professional software engineering; mokka runs alongside a full-time role building integrations for enterprise clients.
Python · PyTorch · TypeScript · Next.js · FastAPI · LangGraph · Tauri · PostgreSQL · Qdrant · Docker · Playwright

Pinned Loading

  1. MK-SH-V MK-SH-V Public

    MK-SH-V: a 260M visual decision model. Calibrated probabilities, a choice, and a refusal below a validated gate, over document pages and screenshots.

    Python 5 1

  2. MK-SH-1 MK-SH-1 Public

    Small local decision models with calibrated confidence and abstention. Early development.

    Python 1

  3. laya-router laya-router Public

    OpenAI-compatible proxy that routes prompts to cheap or frontier LLMs using a local System 1 decision model (laya)

    Python 2

  4. clef-router clef-router Public

    Route prompts between cheap and frontier LLMs with Cloudflare's Clef decision model. OpenAI-compatible proxy + library.

    Python 2 1

  5. nexary-platform nexary-platform Public

    On-premise enterprise AI platform: multi-provider LLM chat, hybrid RAG (dense + BM25 + cross-encoder rerank), SYSLOG-auditable — runs inside customer networks

    TypeScript 1

  6. nitelos nitelos Public

    Nitelos — verify the outcome, not the attempt (Python SDK)

    Python 1