Priyanshu Joshi

AI/ML Engineer · Deep Learning, LLMs & Applied ML

Ahmedabad, Gujarat, India

@priyanshu_joshi

LinkedIn GitHub

0 followers

🎓 B.Tech in Artificial Intelligence & Machine Learning at Parul University · Graduating 2029

About

AI/ML engineer focused on building and training models, including a 125M-parameter transformer trained from scratch on a single 6 GB GPU. Has also shipped offline RAG, agentic systems, and applied machine learning projects using local LLMs and classical ML methods.

Education

2029

B.Tech

Parul University

Artificial Intelligence and Machine Learning · 2025 – 2029

Artificial Intelligence and Machine Learning · 2025 – 2029

Skills

1 to 2 years

LLMs FastAPI Node.js Random Forest

6 to 12 months

RAG Pandas OpenCV Tf-Idf Naive Bayes Transformers Linear Regression Feature Engineering KNN (K-Nearest Neighbors) Logistic Regression ( in progress )

3 to 6 months

CUDA NumPy K-Means WebSockets scikit-learn Deep Learning

<3 months

Mlp PyTorch Electron Convolutional Neural Networks

Tools / apps / platforms

6 to 12 months

Ollama

Projects

Knoc8 Desktop Agent — Voice-Driven Agentic AI ↗

Python, ESP32

Built a wake-word agent that records from an ESP32 mic, transcribes locally, plans multi-step tasks on Windows, and reports back by voice.
nigixmosha — LLM Director for Full-Cast Audiobooks ↗

Ollama Cloud, TypeScript, TTS engines

Built a pipeline where an LLM reads a chapter, infers character persona and emits structured speaker, emotion and voice assignments, with deterministic fallback reading.
Context-Aware Autocorrect — Edit-Tagging Transformer ↗

Python, PyTorch, SentencePiece

Designed a sequence-tagging alternative to generative correction using one edit tag per token, with a composite embedding of token, character n-gram and word-shape features.
CrimeGPT — Fully Offline RAG Legal Assistant ↗

Ollama, Qwen 2.5, nomic-embed-text, Electron

Built a zero-network-call retrieval pipeline with local model selection, embeddings over a legal corpus, query rewriting, vector similarity search, citation-backed answer synthesis, and JSON vector storage.
CodeSeek — 125M-Parameter Transformer, Trained From Scratch ↗

Python, PyTorch, CUDA

Built and trained a 12-layer, 768-dim, 12-head decoder-only Transformer from random initialisation, implementing RoPE, SwiGLU, pre-norm LayerNorm and weight tying. Fit 125M parameters on a 6 GB GPU using gradient checkpointing, FP16 AMP and gradient accumulation; trained a custom 16K-vocabulary BPE tokenizer.

🏆 Achievements & awards

AI Summit Buildothon Finalist · 2026
Smart India Hackathon Finalist · 2026

MDoNER
Kanad S.H.I.E.L.D. 2026 · 2026

Ahmedabad Cyber Crime Branch
57 Public Repositories

Explore the Broxer community

💬 Q&A Ask about roles, salaries and interviews — and help others. 👥 Circles Join discussions on skills, companies and industries. 📰 Feed Follow updates and insights from the community.