Ekjot Singh
Parsing the Thermodynamics of Intelligence.
I am a Systems Architect & Researcher building at the frontier of Deterministic AGI. Currently bridging the gap between biological intuition and digital reasoning at Metanthropic.
Selected Artifacts
01Chromabase
Owner • Private
A powerful CLI utility that encodes massive datasets and files into resilient video formats for infinite cloud storage bypass.
SIM8086
Owner • Public
A modular cross-platform hardware emulator and assembler for the Intel 8086 microprocessor.
Technical Arsenal
02Systems & Low-Level
- Rust
- Go
- C
- C++
- Assembly
- Bash
Database & Real-Time Engines
- PostgreSQL
- libpq
- PostgREST
- Elixir
- WebSockets
- Vector Databases
AI & Applied Research
- Python
- PyTorch
- Llama Models
- OpenRouter API
- Dataset Distillation
Infrastructure & Orchestration
- Docker
- Kubernetes
- Terraform (HCL)
- Ansible
- Linux Kernel
Security & Cryptography
- Web Crypto API
- JWT
- Row Level Security (RLS)
- Uncrypto
Application Architecture
- TypeScript
- Next.js
- React
- Node.js
- Bun
- Tailwind CSS
- Qt 6
Education
03Vellore Institute of Technology
Specialized in Artificial Intelligence and Systems Engineering.
Certifications
04Algorithms Specialization
Machine Learning Specialization
The Bits and Bytes of Computer Networking
Experience
04TealBase
Engineering an open-source Firebase alternative utilizing PostgreSQL. Architecting the core database engine, real-time WebSocket broadcasting (realtime-js), and secure JWT-based auth systems with Row Level Security. Managing open-source community contributions and infrastructure scaling.
Metanthropic Lab
Architecting a frontier research institution dedicated to building verifiable, interpretable, and safe Artificial General Intelligence. Leading development on cross-model generalization methodologies like Linear Gradient Matching for dataset distillation.
Publications
05The Fragility of Guardrails: Cognitive Jamming and Repetition Collapse in Safety-Steered LLMs
A mechanistic audit of LLM residual streams using Sparse Autoencoders (SAEs). We demonstrate that aggressive safety-steering vectors often interfere with latent world-modeling circuits, triggering 'Cognitive Jamming'—a failure mode where models spiral into repetition rather than grounded refusal.
Dataset Distillation for the Pre-Training Era
Introducing Linear Gradient Matching (LGM), a novel method to condense massive datasets into a single synthetic image per class. We reveal shared 'Platonic' representations across foundation models (CLIP, DINO-v2) and demonstrate how distilled data acts as a lens to diagnose model robustness.
The Notebook
04"The Notebook" is a collection of rough thoughts, research notes, and architectural decisions. It serves as a public log of my attempt to reverse-engineer intelligence.
View All Entries →The Infinite Pixel
Jan 29, 2026The CSI 'Enhance' button is finally real. How I built a local-first AI upscaler to recover lost details without sending your photos to the cloud.
Code at the Speed of Thought
Jan 24, 2026We are moving from 'writing code' to 'architecting logic'. Introducing AhhCode, the agent that turns the terminal into a pair programmer.
The Cloud is Just Someone Else's Computer
Jan 21, 2026Why rent an H100 when you have five laptops gathering dust? Building VGD to create a decentralized home compute cluster.