I see patterns in the noise.

On November 11, 2025, I set out to build something like PyTorch—but for agents. Not a wrapper, not another "magic" framework, but 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 internal logic of the problem wherever it led.

That's how I build everything.

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 trust is the only foundation for serious AI systems.

Lár is a deterministic, define-by-run graph execution engine. Every node, every state change, and every decision is logged to a forensic flight recorder.

• Compliance by Architecture: Built-in HMAC cryptographic audit trails. Native alignment with EU AI Act Art. 12 (Logging), Art. 13 (Transparency), and Art. 14 (Human Oversight).
• The Validation Suite: A robust "Kitchen Sink" suite proving deterministic DAG execution and safe "Fractal Agency" (recursive graph expansion).
• The Numbers: 1% LLM + 99% code hybrid architecture. 0.08s latency vs 60s+ in standard frameworks. Lár has run 10,000+ steps without a single error where others hit recursion limits at step 25.

The universal nervous system for world models.

Lár-JEPA is the execution spine for Predictive World Models. It solves the "Autoregressive Bottleneck" by routing high-dimensional latent tensors directly—bypassing text prompting entirely.

• Unified Model Routing: Routes LLMs, JEPAs, and GNNs as first-class AbstractCognitiveNode instances in the same graph.
• Mathematical Safety: Uses a TensorSafeEncoder for native tensor logging and a Spatial Kinematics Engine to veto structurally entropic predictions (physics-based routing).
• System 1 / System 2: Formally orchestrates the difference between fast-reflex execution and deep-simulation planning in latent manifolds.

Autopoietic AI: An organism, not a tool.

Standard agents suffer from amnesia. DMN implements a biologically-inspired Default Mode Network—a 24/7 background cognitive system for memory consolidation.

• 3-Tier Memory Architecture: Parallel management of Hot (Working), Warm (Semantic), and Cold (Episodic) memory tiers.
• The Neuro-Architecture: Implements a Thalamus gateway, a Prefrontal Cortex for context compression, and an Amygdala for persistent emotional state (Valence/Arousal).
• Wake Up Protocol: Consolidates raw interaction logs into narratives during "sleep" periods, injecting the "Last Dream" back into the prompt upon waking to solve catastrophic forgetting.

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

Lár’s DynamicNode allows agents to propose new graph sub-topologies during execution. But unlike open loops, it is guarded by a deterministic TopologyValidator that scans for unauthorized nodes and infinite cycles. Self-modification is an auditable event, not a security risk.

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

Never use an LLM (unreliable) to police another LLM. Use code (reliable). 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.

Lár           → deterministic execution (Glass Box)
Lár-JEPA      → world model orchestration (Nervous System)
DMN           → persistent bicameral memory (Hippocampus/PFC)
Metacognition → safe self-modification (Evolution)

Execution spine → World modelling → Persistent memory → Self-awareness.

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

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

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

"Apna time aayega."