Skip to content
View ayushsyntax's full-sized avatar

Block or report ayushsyntax

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 is supported. This note will only be visible to you.
Report abuse

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

Report abuse
ayushsyntax/README.md

AYUSHSYNTAX

software should feel inevitable.

small models, useful agents, and systems that respect the problems they solve.


LinkedIn   X   Email   Portfolio



╭───────────────────────────────────────────────────────────────╮
│                                                               │
│   ayush@void:~$ whoami                                        │
│                                                               │
│   NAME       Ayush Kumar                                      │
│   ROLE       AI Engineer                                      │
│   EDUCATION  B.Tech AI & Data Science, 2027                   │
│                                                               │
│   FOCUS      ML Engineering · LLM Engineering                 │
│              Agentic AI · MLOps                               │
│                                                               │
│   STACK      Python · PyTorch · TensorFlow                    │
│              LangGraph · FastAPI · Docker                     │
│              AWS · MLflow · DVC                               │
│                                                               │
│   STATUS     training models. building agents.                │
│              shipping things that actually work.              │
│                                                               │
╰───────────────────────────────────────────────────────────────╯

🔬 current work

projects:
  - name: GPT From Scratch
    tech: PyTorch
    note: decoder-only transformer,
          first principles

  - name: Agentic Research Assistant
    tech: LangGraph · RAG
    note: multi-agent autonomous
          research pipeline

  - name: Brain Tumor MRI Classifier
    tech: TensorFlow · FastAPI
    note: production medical imaging
          classification system

  - name: Production RAG Systems
    tech: LangChain · FAISS
    note: scalable retrieval-augmented
          generation at scale

📡 signal

learning:
  - LLM Internals
    # attention, KV caches, quantization
  - Multi-Agent Systems
    # orchestration, tool calling, memory
  - Distributed Training
    # FSDP, DeepSpeed, gradient checkpointing

reading:
  - "Attention Is All You Need"
  - "Scaling Laws for Neural LMs"
  - "ReAct: Reasoning + Acting"

philosophy: >
  if it doesn't solve a real problem,
  delete it.
  the best model is the one
  that's actually deployed.

────────────────────────────────────────────────────────────────
  midnight. rain. code. deploy. repeat.
────────────────────────────────────────────────────────────────

◉ projects

GPT FROM SCRATCH PyTorch

Decoder-only transformer built from first principles.

multi-head attention causal masking BPE tokenizer mixed precision training inference pipeline

AGENTIC RESEARCH ASSISTANT LangGraph

Multi-agent system for autonomous research workflows.

supervisor-worker arch persistent memory vector retrieval dynamic tool selection RAG pipeline

BRAIN TUMOR MRI CLASSIFIER TensorFlow

Production-grade medical image classification system.

EfficientNet transfer learning FastAPI endpoint DICOM support Docker deployment

PRODUCTION RAG SYSTEMS LangChain

Scalable retrieval-augmented generation pipeline.

hybrid search dense + sparse retrieval context-aware chunking reranking RAGAS evaluation


⚙ working set


core    Python  C++  SQL  Bash

ml · dl    PyTorch  TensorFlow  scikit-learn  HuggingFace  OpenCV

llm · agents    LangChain  LangGraph  OpenAI  FAISS

infra · ops    Docker  AWS  FastAPI  MLflow  DVC  Git  Linux



────────────────────────────────────────────────────────────────
  "I'd rather be dry, but at least I'm alive."  ── the nbhd
────────────────────────────────────────────────────────────────

◉ metrics

GitHub Stats    Top Languages



GitHub Streak

╭───────────────────────────────────────────────────────────────╮
│                                                               │
│   ayush@void:~$ cat /etc/motd                                 │
│                                                               │
│   → ship fast. ship often. ship things that matter.           │
│   → read papers. write code. repeat.                          │
│   → the best model is the one that's actually deployed.       │
│   → if it can't run in prod, it doesn't exist.                │
│                                                               │
│   ayush@void:~$ echo $STATUS                                  │
│   > open to collaborations & opportunities_                   │
│                                                               │
╰───────────────────────────────────────────────────────────────╯

"Sweater Weather" is just lo-fi for engineers.



built with obsession · deployed with intent

Pinned Loading

  1. GPT-from-Scratch-in-PyTorch GPT-from-Scratch-in-PyTorch Public

    A small GPT built from scratch in PyTorch, trained on Shakespeare. Focused on learning and understanding transformers. Includes training scripts, evaluation, and a demo interface.

    Python

  2. Regression_ML-End-to-End Regression_ML-End-to-End Public

    Implements the full ML lifecycle: time-aware feature engineering, XGBoost modeling with Optuna + MLflow, FastAPI inference, Streamlit dashboard, Dockerized services, CI/CD via GitHub Actions, and d…

    Jupyter Notebook

  3. Agentic-Research-Assistant Agentic-Research-Assistant Public

    Production-grade agentic research assistant built with LangGraph. Features persistent memory, PDF RAG, tool calling for web/news/research, and full observability via LangSmith. Demonstrates modern …

    Python

  4. DL-Clf-ENDtoEND DL-Clf-ENDtoEND Public

    End-to-end production-grade binary brain tumor detection from MRI scans using EfficientNetV2-S transfer learning. Achieves 98.19% accuracy & 99.17% tumor recall on 1,600-image hold-out. Full MLOps:…

    Python 1