🧠Building neuro-symbolic AI Alone... Help is welcome

Hi DEV,

I'm Marco. Four months ago I started building OrKa, a modular, YAML-defined orchestration engine for agentic reasoning. Today it’s not just a prototype, it’s a functional system with live memory, explainable flows, and a full visual builder. It’s also way too big for one person to maintain.

I’m doing this solo. One brain, one laptop, one repo. OrKa has grown into a cognitive infrastructure stack: structured memory layers, Redis/Kafka queues, traceable agent chains, a visual UI, and benchmarked orchestration loops. Version 0.8.0 is out. It works. But it’s a grind, and I need help.

Why OrKa Exists: To Kill Black-Box AI

LangChain, Flowise, AutoGen, they’re building chains and calling it cognition. But try tracing memory across agents, understanding why a step was taken, or observing reasoning in real time, you can’t.

OrKa is different.

• Modular agents with independent logic and memory
• YAML-defined cognition flows, not spaghetti scripts
• Live Redis/Kafka trace logging for every agent decision
• Observable UI + terminal-based TUI
• 6-layer memory model, with decay and scoped storage
• Confidence-weighted routing to simulate dynamic reasoning

This isn’t “prompt chaining.” It’s cognitive orchestration, explainable, testable, and local.

What Works in OrKa 0.8.0

• ✅ Fork/Join agent execution with dynamic paths
• ✅ Confidence scores per agent + agreement synthesis
• ✅ Redis and Kafka-compatible logging
• ✅ Visual OrKaUI with YAML sync
• ✅ Full local + API runtime support
• ✅ ServiceNodes (RAG, MemoryWriter, Embedding Fetchers)
• ✅ Benchmarked: 1000 loops, 7.6s avg latency, 0.00011¢ per run on DeepSeek 32B

But Here’s What Still Sucks

This is real open source. Here’s where I’m drowning:

• 🧠 Memory Scope: Working, but v0.8.0 needs a simpler Redis GET/SET fallback
• 💀 Code Bloat: The Orchestrator class is still too big. We need clean modular separation.
• 🧩 UI Gaps: Memory nodes are stubbed, not visual. Trace replay only half-works.
• 📄 Docs: The readme’s outdated. The guides are scattered. /examples is growing but still thin.
• 📣 Awareness: No Reddit, no X. ~200 GitHub stars, ~10 PyPI installs/day.
• 🧍‍♂️ Solo Dev Hell: I’m writing infra, fixing YAML bugs, fielding Discord questions, and raising 3 kids.

This project won't survive on vibes.

What You Can Do (Today)

🧠 Build SimpleMemory

• File: /src/memory/simple.py
• Use Redis SET/GET, drop Kafka overhead
• Improve fallback logic for small LLMs

🔍 Clean the Codebase

• Break /src/orchestrator.py into smaller modules
• Add ruff, mypy, enforce stricter lint
• Improve logging separation by agent vs. orchestrator

🧱 Contribute Nodes or Agents

• Add a PlannerAgent
• Extend RAGNode with Pinecone/Chroma support
• Add MemoryVisualizerNode to the UI

✍️ Write Real Docs

• Add /docs/architecture.md
• Polish /examples/fact_checker.yaml
• Write /quickstart.md for non-engineers

🛰 Spread It

• Post your workflow on X, Reddit, Discord
• Share a custom YAML orchestration
• Fork it and use it in a hackathon

Why This Matters

AI shouldn’t be a black box. If it can’t be traced, it shouldn’t be trusted.

OrKa is an attempt to build cognitive systems that are:

• Transparent
• Local-first
• Composable
• Deterministic

It’s not a LangChain clone. It’s not a chatbot wrapper. It’s a runtime for modular cognition. You define the graph, OrKa executes it, traces it, and explains it.

But I can’t do this alone.

The Ask

Don’t star the repo, run it.

pip install orka-reasoning

orchestrator:
  id: fact_checker
  strategy: sequential
agents:
  - id: validator
    type: binary
    prompt: Is this statement factual?

from orka import Orchestrator
orc = Orchestrator("fact_checker.yaml")
orc.run({"input": "The moon is made of cheese"})

Want to fix docs? Build a node? Just hang out?

I’m in the server. Ping me. Fork the repo.

Let’s build the cognitive runtime we all wish existed.

Repo: https://github.com/marcosomma/orka-reasoning