Blog

CogniLynch: where cognition meets machine learning

I founded CogniLynch to make AI reliable and useful for the teams who depend on it. This post covers what the company does, how it grows out of my research, and why I'm steering my work toward the point where cognitive science and machine learning meet.

What CogniLynch is

CogniLynch is a custom AI development company. We build and ship production AI systems for teams that need them to work in practice, not just in a demo:

• LLM fine-tuning, retrieval-augmented generation (RAG), and enterprise agents.
• Production ML delivery and MLOps: deployment, monitoring, pipeline orchestration, and scalable inference.
• Vision, multimodal, and retrieval systems: embeddings, semantic search, knowledge graphs, OCR, and video analytics.
• AI cybersecurity and ethical jailbreaking.

Alongside client work, we run applied research in medical AI, NLP and vision, and trustworthy and explainable AI.

Built on research

Several CogniLynch products grow directly out of my papers:

• On-device LLM runtime builds on BitRL, my work on reinforcement learning with 1-bit quantized language models for edge deployment (IEEE FET 2026).
• Federated fine-tuning platform builds on Federated Fine-Tuning Frameworks for Privacy-Preserving Distributed Large Language Intelligence (Scientific Reports, in press).
• MRI tumor segmentation service builds on my augmented 3D U-Net for multimodal MRI brain tumor segmentation (ICADCML 2025).
• Grid battery control platform builds on FedGrid-RL, my federated deep RL framework for private, personalized battery control (under review).

Security R&D

Our security line is still in development, and it mirrors the research I'm doing now:

• Inference timing auditor, from Batch Echoes: token-length timing leakage in continuous-batching LLM servers such as vLLM and SGLang.
• Honeypot robustness testbed, from Deception Collapse: how adversarial LLM agents evade deception-based defenses.
• Autonomous security evaluation agent, from Phantom Profiler: an LLM-augmented attack agent formalized as a POMDP.

Why cognition

The name is deliberate. I want to work where cognitive science and machine learning meet: systems that model how people attend, learn, and remember, and that adapt to them.

I'm already doing this. I lead AI research and development for TheraMuse at Moodifai, an adaptive music therapy platform for people with Down syndrome, ADHD, and dementia. It personalizes sessions in real time with Q-Learning and Linear Thompson Sampling over 32 clinically informed parameters. TheraMuse is supervised by Dr. Mustak Ibn Ayub (University of Oxford) and has secured $3 million in venture funding.

What's next

Next, I want to take CogniLynch further into cognitive AI: adaptive systems for cognitive health, learning, and attention, built with the same care for privacy, robustness, and explainability that runs through my research.

Get in touch

If you work on cognitive AI, trustworthy ML, or AI security and want to collaborate, email me at ashiqsazid494@gmail.com, or reach CogniLynch at info@cognilynch.com. You can also find us at cognilynch.com and on LinkedIn.

Six research ideas I want to work on

A couple of papers I read this month, plus problems I keep coming back to, have turned into a list of research directions I want to pursue. Some already have experiments running; others are still at the brainstorming stage. Here they are one by one, with the venue I'm aiming each at.

1. Quantum circuits designed by large language models

This one started with a new paper from Anthropic and Harvard researchers, Discovering state-of-the-art quantum circuits with large language models. An LLM proposes circuit designs, an independent evaluator checks every candidate for correctness and scores its cost, and the search keeps the best. In seven of the eight tasks, the discovered circuits cut both computational cost and qubit requirements compared with the state of the art, and the search turned up a new architecture for modular inversion.

What excites me is the pattern as much as the result: a generative model paired with a verifier it can't fool. Quantum circuit design is a field where correctness can be checked exactly, which makes it a natural testbed for AI-assisted discovery. I'm planning to work in this domain, aiming for ICLR.

2. System One models for AI agents

Psychologists describe two modes of thinking: System 1 is fast and automatic, System 2 is slow and deliberate. Most LLM agents today run everything through System 2, calling a large model for every small decision.

JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs takes a different route. Smaller decision models handle routine choices, and the LLM steps in only when they're unsure. The authors release AgentDecisionBench, the JevAgentHarness framework, and AgentDecision-4B, and keep task success rates high while cutting cost.

I'm brainstorming what a good "System One" model for agents should look like: when it should act on its own, when it should hand control back, and how it connects to the cognitive side of my work at CogniLynch. Target: ICML.

3. Batch Echoes: timing side channels in LLM inference servers

Inference servers such as vLLM and SGLang use continuous batching: requests from different users are scheduled together to keep the GPU busy. That sharing has a price. If the time it takes to serve one request depends on how many tokens another user's request produces, timing becomes a side channel.

Batch Echoes measures this token-length timing leakage in continuous-batching schedulers. I'm running pilot experiments on RTX 4090 infrastructure, and the target is USENIX Security.

4. FedGrid-RL: private battery control for grid-interactive buildings

Buildings with batteries can shift when they draw power to help the grid, but the data needed to control them well is private. FedGrid-RL uses federated deep reinforcement learning with a shared dynamics encoder: each building gets battery control personalized to it, and the buildings learn together without pooling their raw data.

5. Deception Collapse: when LLM agents beat honeypots

Honeypots and other deception-based defenses work by watching how an attacker behaves. An attacker driven by an LLM can learn to suppress the behavioral signals those systems look for. Deception Collapse studies how adversarial LLM agents evade detection in deception-based security environments, and what that means for the robustness of LLM honeypots. Target: IEEE Transactions on Information Forensics and Security.

6. Phantom Profiler: an autonomous adversarial LLM agent

To test defenses against AI attackers, you need a principled AI attacker. Phantom Profiler formalizes an LLM-augmented attack agent as a POMDP/Semi-MDP, so it can adapt as it probes a system and exploit vulnerabilities on its own. The goal is a rigorous tool for autonomous security evaluation. Target: NDSS.

Want to work on one of these?

If any of these overlap with what you're doing, I'd like to hear from you at ashiqsazid494@gmail.com. Several of them also feed into what we're building at CogniLynch.