Researcher and Engineer

On November 11, 2025, I set out to build something like PyTorch — but for agents. Not a wrapper. Not another "magic" framework. A ground-up rethink of what agentic infrastructure should actually be.

Eight months later, I had a cognitive architecture, six published preprints, and a formal proof that routing safety and machine learning are the same mechanism.

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.

Six papers, one architecture, traced end to end: detect the disagreement, universalise it across any model, learn from it without labels, remember it across encoder upgrades, embody it in a deployed robot, persist until the problem is resolved.

Multi-Stream Routing Without Modal Fusion and the Safety-Learning Equivalence.

Introduces the Safety-Learning Equivalence Theorem: the invariants that make a routing system safe are mathematically identical to the invariants that make inter-model disagreement a valid, label-free training curriculum. Safety and learning are not a trade-off. They are the same mechanism.

A sufficient and extensible cognitive contract for autonomous multi-model systems.

Proves that ten Abstract Base Classes constitute a complete, formally-specified cognitive contract for any autonomous multi-model system. Establishes domain isomorphism across biomedical imaging, quantitative finance, and geophysical hazard assessment — structurally unrelated domains governed by identical routing invariants.

Routing flags as gradient signals: self-curating curriculum, LoRA adapter integration, and annotation-free continual learning.

Proves that a routing flag — the moment two independent streams disagree — is a free, automatically-labelled training example. Acting on those flags closes the training loop with no human annotation, building a self-curating D_hard curriculum that trains LoRA adapters directly from routing disagreement.

World-grounded difficulty representations for encoder-invariant and predictive continual learning.

What's hard is a fact about the world, not about your current encoder. Storing hard cases as raw pairs means the memory survives a model upgrade — swap the encoder, keep the knowledge. Difficulty identity holds stable across encoder families (Spearman ρ = 0.672).

Label-free detection via concept-space routing in deployed robotic systems.

When the floor turns to ice, a robot's live embedding drifts away from its reference. That drift is detectable from geometry alone — no labels, no rules, no reward signal. 95% violation detection the moment friction changes, validated on a physical robotic policy, not just simulation math.

Proprioceptive error resolution with scope-bounded invariant signal tracking.

Detecting a violation isn't adapting — it's logging. The signal that detects a physics violation is the same signal that verifies whether the fix worked. Closes the six-paper adaptation loop: a system that detects physical surprise, learns from it without labels, refines its response until resolution, remembers what worked, and escalates when the problem exceeds its scope.

The PyTorch for Agents.

Most agent frameworks are black boxes. When they fail in production, you get a stack trace and no idea what happened, why, or how much it cost. Lár is built on a different premise: 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.

• 23 EU AI Act Requirements — exercised as live runtime nodes, not documentation: LethalTrifectaGuard, FundamentalRightsImpactNode, AuthorityLedger, CredentialVault, HumanJuryNode, ComplianceManifestGenerator, and more. HMAC-signed audit artefacts on every run.
• Fractal Agency — AdaptiveNode composes new graph sub-topologies at runtime, guarded by a deterministic TopologyValidator. Self-modification is an auditable event, not a security risk.
• The Numbers — 1% LLM + 99% code. 0.08s latency vs 60s+ in standard frameworks. Deterministic by design.

A complete cognitive contract for autonomous systems.

Ten Abstract Base Classes. 33 formal invariants. The complete framework — from perception to memory to learning — specified mathematically.

• AbstractDivergenceRouter (V1–V6) — routes by inter-stream geometric divergence, not content. V4 Content Blindness is not optional: the routing function never sees what it's routing. This is what makes the Safety-Learning Equivalence possible.
• Domain Isomorphism — crystal lattice fault localisation, geological risk coordinates, network failure nodes, medical image–report disagreement, financial signal disagreement. One architecture. No domain-specific modification.
• The Self-Improving Loop — routing flags disagreement → disagreement builds D_hard curriculum → curriculum trains LoRA adapters → adapters improve routing. Closed-loop. No human annotation in the critical path.

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 — Hot (Working), Warm (Semantic), Cold (Episodic)
• The Dream Loop — consolidates raw interaction logs during idle periods, injects the "Last Dream" on waking. Catastrophic forgetting solved architecturally, not by prompt tricks.

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

MCP BioForensics — Natural-language querying over 17,000+ ClinicalTrials.gov records via MCP.

MCP Forensic Toolkit — AI-enabled digital forensics: log triage, SHA-256 integrity, audit-grade reports.

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. Safety and learning are not a trade-off. They are the same mechanism. The routing invariants that prevent catastrophic decisions are identical to the invariants that identify the most valuable training examples. This is not a design choice. It is a theorem.

Lár → deterministic execution (Glass Box) Lár-JEPA → ten-ABC cognitive contract (The Nervous System) DMN → persistent bicameral memory (The Hippocampus) D_hard loop → self-curating curriculum (The Immune System) Execution → Cognition → Memory → Learning.**

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