KONJO — Know, Outline, Nail, Justify, Optimize.

Konjo · ቆንጆ — Beautiful. · 根性 — Fighting spirit. · 康宙 — Health of the universe. · खोजो — Search and discover.

Make it Konjo — Plan, build, test, ship, rest, repeat.

Konjo is the word you reach for when a tool does exactly what it needs to do — nothing more, nothing less.

We build the open infrastructure layer for AI. Not wrappers, not prompt chains — the deep systems that power them: compression, inference, vector engines, memory architectures, compliance, and the research that pushes the whole stack forward. Vertically integrated, benchmark-driven, built in public.

┌──────────────────────────────────────────────────────────────┐
│                        Orchestration                          │
│              lopi (Agent Orchestrator) · kyro (RAG)          │
├──────────────────────────────────────────────────────────────┤
│                       Inference Layer                         │
│            squish (Local LLMs) · kairu (Speculative)         │
├──────────────────────────────────────────────────────────────┤
│                       Retrieval Layer                         │
│             vectro (Vectors) · kohaku (Memory)               │
├──────────────────────────────────────────────────────────────┤
│                  Research & Compliance                        │
│    drex (Architecture) · toki (Robustness) · miru (XAI)     │
│                    squash (EU AI Act)                         │
└──────────────────────────────────────────────────────────────┘

Compress local LLMs once. Run them forever.

OpenAI + Ollama drop-in for Apple Silicon with statistically identical accuracy and 54× faster cold starts. INT8/INT4/AQLM quantization, MLX-native, sub-second load times.

quantization apple-silicon mlx local-llm openai-api ollama

Speculative decoding engine for HuggingFace models.

EAGLE-style drafting, dynamic early exit, and token budget control with a live performance dashboard. Designed to squeeze every token out of your inference budget.

speculative-decoding huggingface inference latency throughput

Lightning-fast embedding quantization and ANN indexing.

12M+ vec/s throughput — 4.85× faster than FAISS C++ — with drastically reduced memory footprint. Sub-1ms encode, HNSW (recall@10=0.920), NF4/INT8/PQ quantization, AutoQuantize, pipeline CLI, WASM bindings.

rust mojo quantization vector-search embeddings rag hnsw

Neural episodic memory engine using HDC hypervectors.

Stores long-term context and retrieves via associative recall. A persistent memory layer that goes beyond RAG — designed for LLMs that need to remember across sessions and tasks.

episodic-memory hdc hyperdimensional-computing rag-alternative llm

High-performance orchestrator for Claude Code agents.

Concurrent agents in git-isolated branches with retry loops, priority queuing, SQLite memory, pattern mining, scheduled tasks, and a live TUI + web dashboard. Remote control via Telegram.

rust claude-code agent-orchestrator async-rust tokio tui telegram-bot

Production RAG pipeline.

Hybrid retrieval (dense + BM25 + RRF), reranking, and RAGAS evals baked in. Local-first with Squish + Vectro. Plugs into OpenAI, Anthropic, or runs fully offline.

rag hybrid-search bm25 reranking ragas qdrant local-llm

Automated EU AI Act compliance for AI/ML teams.

Annex IV documentation, SBOMs (CycloneDX/SPDX), policy checks, and Sigstore-signed audit records inside your CI/CD pipeline. August 2, 2026 enforcement deadline — squash gets you there.

eu-ai-act compliance-automation sbom cyclonedx sigstore cicd nist-ai-rmf

AI Compliance Gate for GitHub Actions.

EU AI Act · NIST AI RMF · ISO 42001. One step in CI/CD. Pairs with squash.

Homebrew tap for squash.

brew install konjoai/squash/squash-ai — the fastest path to EU AI Act compliance on macOS.

homebrew macos squash tap

Next-generation neural architecture research.

The transformer is a brilliant hack scaled past its limits. DREX is what comes next — tiered memory, sparse execution, and a learned controller that knows what to remember. Mamba SSM + ESN reservoir + HDC encoding + NoProp memory + RL controller + KAN readout.

transformer-alternative mamba tiered-memory sparse-attention cognitive-architecture

Adversarial fine-tuning lab for small models (1B–3B).

Generate red-team attacks, harden via LoRA, ship datasets and robust weights. Investigates whether robustness truly generalizes or just overfits under pressure.

red-teaming adversarial-ai lora fine-tuning alignment model-robustness

Multimodal reasoning tracer.

Answers questions over images and documents while visualizing step-by-step attention and reasoning chains. Built for explainable vision-language AI — see what the model sees and why.

vlm multimodal xai attention vision-language-models explainability

Konjo AI design system + portfolio homepage monorepo.

Packages: ui · Apps: web

We build primarily in Rust and Python, integrating Mojo and C++ where the metal demands it.

1. CLI-first, pipe-friendly — Every tool must play seamlessly with Unix workflows and CI/CD pipelines.
2. Show the math — Real benchmark numbers over vibes. Latency (p50/p95/p99) and accuracy are measured constantly.
3. Open specs, zero lock-in — Built on open formats. Swap any Konjo layer for something else; the system allows it.
4. Ship fast, iterate in public — Ugly v1s beat perfect local branches every time.

We're a small, focused team building in the open. If you've written a custom benchmark script because nothing composable existed, had an agent forget everything between sessions, or spent more time on the README than the code — you'll feel at home here.

Ways to contribute:

• ⭐ Star a project you find useful
• 🐛 Open an issue with a bug, idea, or frustration
• 🔨 Pick up a good first issue and send a PR
• 💬 Start or join a Discussion — we make decisions in public
• 📣 Build something with our tools and share it

We're actively looking for:

• Rust and C++ engineers interested in inference optimization
• ML researchers working on memory systems or efficient architectures
• Python developers building agent infrastructure
• Anyone who's felt the gaps we're filling

• 🌐 konjotech.com
• 𝕏 @wesley_scholl

All Konjo AI projects are open source — most are MIT licensed unless otherwise noted. Build on them, fork them, ship them.

Make it Konjo — Plan, build, test, ship, rest, repeat.