Skip to content
View axdithyaxo's full-sized avatar

Block or report axdithyaxo

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.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

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

Report abuse
axdithyaxo/README.md

Aadithya Vishnu Sajeev

I see patterns in the noise before the noise knows it's a pattern.

I'm from Kerala. I live in Dublin. I build things because I can't not build them.

On November 11, 2024, I wanted to build something like PyTorch — but for agents. Not a wrapper. Not another LangChain. A ground-up rethink of what agentic infrastructure should actually be.

109 days later I had Lár, Lár-JEPA, DMN, and a compliance architecture mapped to the EU AI Act. I didn't plan most of it. The ideas emerged as I followed the logic of the problem wherever it led.

That's how I build everything.


The Work

Lár — The Glass-Box Agent Engine

The PyTorch for Agents.

Most agent frameworks are black boxes. When they fail in production you get a 100-line stack trace and no idea what happened, why, or how much it cost. I built Lár because I think that's the wrong foundation for serious AI systems.

Lár is a deterministic, define-by-run graph execution engine. Every node, every state change, every decision is logged to a forensic flight recorder. Built-in HMAC cryptographic audit trails. EU AI Act Article 12/13/14 compliant by architecture. Air-gap capable. FDA 21 CFR Part 11 ready.

Not magic. Not wrappers. Just pure, debuggable Python you actually own.

The numbers: 1% LLM + 99% code hybrid architecture. 0.08s vs 64s latency. $0.00 vs $3.60/run vs standard frameworks. LangGraph hits recursion limit at step 25. Lár ran 10,000+ steps without a single error.

Lár-JEPA — Post-LLM Orchestration

The nervous system for world models.

Every major agent framework assumes text in, text out. That assumption breaks when your model outputs a 768-dimensional tensor representing the abstract state of a physical environment.

Lár-JEPA is the deterministic execution spine for Predictive World Models — routing high-dimensional latent tensors without text prompting. The industry is building the brain. This is the nervous system.

DMN — Bicameral Memory Architecture

An organism, not a tool.

Standard agents have amnesia. DMN implements a biologically-inspired Default Mode Network — a background daemon that watches execution logs, sleeps, dreams, and consolidates short-term memory into long-term semantic understanding via ChromaDB. Solves catastrophic forgetting without retraining. Gets smarter while you're away.

Metacognition — Dynamic Self-Modifying Graphs

Agents that rewrite their own execution topology at runtime. Safely.

DynamicNode generates JSON graph specs — including BatchNodes for parallel execution — validated by a deterministic TopologyValidator. Self-modification as an auditable event, not a jailbreak risk. The AI proposes. The code decides.


Other Work

BreakHis Classifier — ResNet-50 breast cancer classifier on histopathology data. 0.96 F1-score, 0.98 AUC.

MCP Forensic Toolkit — AI-enabled digital forensics via Model Context Protocol.

MCP BioForensics — Clinical trial data exploration with hybrid retrieval and natural-language querying.


The Philosophy

The industry is building the Brain. I'm building the Nervous System.

Never use an LLM to police another LLM. Use code. An approval is not a flag. It is a cryptographic signature of a specific state. Self-modifying code is only dangerous in a black box. In a glass box it's just evolution with an audit trail.


The Stack

Lár           → deterministic execution
Lár-JEPA      → world model orchestration
DMN           → persistent bicameral memory
Metacognition → safe self-modification

Execution spine → world modelling → persistent memory → self-awareness.

A complete cognitive architecture. Built from scratch. In public. Under Apache 2.0.


Background

MSc Data Analytics, Dublin City University. Kerala → Dublin. Future: CTO, SnathAI.


axdithya@gmail.com · LinkedIn · snath.ai · docs.snath.ai


"Apna time aayega."

Popular repositories Loading

  1. mcp-forensic-toolkit mcp-forensic-toolkit Public

    Secure AI-enabled digital forensics toolkit built with Model Context Protocol

    Python 13 3

  2. axdithyaxo axdithyaxo Public

    Config files for my GitHub profile.

    1

  3. BreakHis_Classifier BreakHis_Classifier Public

    Built a high-accuracy ResNet-50 breast-cancer classifier using the BreaKHis histopathology dataset (40x–400x magnification), achieving 0.96 F1-score and 0.98 AUC. Developed a Streamlit diagnostic a…

    Python 1

  4. mcp-rag-toolkit mcp-rag-toolkit Public

    A Production Ready local Retrieval-Augmented Generation (RAG) system using FAISS, SQL, and Claude

    Python

  5. mcp-bioforensics mcp-bioforensics Public

    MCP BioForensics is an AI-driven platform for clinical-trial data exploration with hybrid retrieval and natural-language querying.

    Python

  6. docs docs Public

    MDX