1-day current streak·7-day longest streak
Maanas Tyagi LLM Engineer · AI Intern at Sstudize · HuggingFace Contributor Built a Python-based tool to replace a $3,000 reconciliation software. Built a language model from scratch — no…
Maanas Tyagi
LLM Engineer · AI Intern at Sstudize · HuggingFace Contributor
Built a Python-based tool to replace a $3,000 reconciliation software.
Built a language model from scratch — no library shortcuts, just PyTorch and stubbornness.
Leading the team at Sstudize — building LLM applications and multi-agent systems for students.
Actively working on research projects.
📖 Currently learning: AI Systems Design
---
Projects
🧠 small-llm-from-scratch
GPT-style LLM built entirely from raw PyTorch — no transformers library, no pre-built attention. Implemented autograd, causal self-attention, Grouped-Query Attention, and a full training loop. Pretrained on TinyStories (~112M tokens) on a free T4 GPU. Live demo on HF Spaces.PyTorch Python Gradio HuggingFace
📊 RAG Eval System
Built specifically to practice the part most RAG projects skip — evaluation. Found and fixed a real dedup bug in LangChain's MultiQueryRetriever, then ran ablations proving the "obvious" hybrid retrieval default actually loses to semantic-only search (Recall@5 0.85 vs 0.82).LangChain Chroma RAGAS Python
🔬 Lumen Research
Local-first AI paper reader. Upload a 100-page PDF, chat with it, highlight anything for ELI5 / math / step-by-step breakdowns. Zero cloud dependency via Ollama.FastAPI React FAISS Ollama RAG
🔄 Corrective RAG
Full CRAG pipeline: retrieve → grade → rewrite → generate. Falls back to web search when the vector store fails. Graceful, not fragile.LangGraph LangChain FAISS Python
📦 RAG with Summary Embeddings
Multi-vector RAG: embed summaries for retrieval, store original chunks separately. Better semantic search without losing information — fixes the core trade-off in naive RAG.LangChain ChromaDB HuggingFace Python
🧩 Concept Mapper
Type a topic. Get a personalized learning path as an interactive 3D knowledge graph. Click any node for resources and next steps.Three.js Flask OpenRouter
⏱️ Claude Usage Meter
Windows tray widget showing live Claude.ai session limits + reset countdowns. Polls the same internal endpoint Claude's own UI uses. One-click installer.Electron JavaScript
💬 Todoist Agent
WhatsApp bot that manages your Todoist tasks via natural language. "Add call Mom at 5pm tomorrow" — done. Deployed on Railway.Python DeepSeek Twilio Railway
🗺️ Groove
GTA fog-of-war for the real world. Map starts dark, reveals permanently as you walk or drive through it. Custom lat/lng quantization grid in pure Dart — because H3 broke on Android FFI.Flutter Dart Android
✍️ Human-Write
A field guide for eliminating AI writing patterns — 10 categories, 40+ specific tells, real failure examples, and the fix for each.Resource Prompt Engineering Writing
---
Research
📄 Front-Loading Bias in Large Language Models · ResearchGate Preprint
Started as a personal observation: LLMs write less per bullet as a response grows longer. Formalized into an empirical study across GPT, Gemini, and Claude.
📝 Instruction Fidelity Decay in SLM Agents · *Under review*
Tested how small language model agents degrade under semantic noise — format compliance breaks before semantic recall does.
---
-
Concept_Mapper ★ PINNED
Concept Mapper is an AI-powered web application that takes a topic you want to learn, asks you a few clarifying questions, and generates a personalized learning path visualized as an interactive 3D knowledge graph. Click any node to see what to do, what to study, and where to find the best resources.
HTML ★ 2 3mo agoExplain → -
Corrective-Rag ★ PINNED
A personal Second Brain built with a Corrective RAG (CRAG) pipeline: multi-query retrieval, grading, optional web search, cross-encoder reranking, and a Streamlit UI for querying Markdown notes.
Python ★ 1 3mo agoExplain → -
Lumen-Research ★ PINNED
A local-first AI tool to deeply understand research papers — with RAG-powered chat, math explanations, deep dive threads, and zero cloud dependency.
TypeScript ★ 1 3mo agoExplain → -
Research-Digest-Bot ★ PINNED
No description.
Python ★ 0 3mo agoExplain → -
human-write ★ PINNED
A complete field guide for eliminating AI writing patterns — every tell, every fix, in one file.
★ 3 1mo agoExplain → -
todoist-agent ★ PINNED
Personal Todoist agent powered by DeepSeek via OpenRouter - WhatsApp bot via Twilio, deployable on Railway
Python ★ 0 1mo agoExplain → -
claude-usage-meter
Windows tray app showing your claude.ai session (5h) and weekly usage limits
JavaScript ★ 2 11d agoExplain → -
Tool_Calling_SLM_Research
No description.
Python ★ 1 11d agoExplain → -
transformers ⑂
🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
Python ★ 1 4mo agoExplain → -
maanas1234
No description.
★ 0 1h agoExplain → -
agentfm
No description.
Python ★ 0 1h agoExplain → -
portfolio
No description.
JavaScript ★ 0 3h agoExplain → -
small-llm-from-scratch
A small language model built entirely from scratch in PyTorch — no transformers library, no pre-built attention layers. Every piece (tensors → autograd → attention → grouped-query attention → transformer block → full model → training loop → generation) was implemented by hand, then pretrained on GPU and deployed as a live demo.
Jupyter Notebook ★ 0 3h agoExplain → -
freellmapi ⑂
OpenAI-compatible proxy that stacks the free tiers of 28 LLM providers (~4B tokens/month) behind one /v1 endpoint — plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.
★ 0 2d agoExplain → -
Witness
No description.
TypeScript ★ 0 13d agoExplain → -
notes-mcp
No description.
Python ★ 0 17d agoExplain → -
waku-agent ⑂
Waku Waku! Waku agent is your personal AI agent, on your own laptop, in code you can read in an afternoon — harness + loop + memory + eval
★ 0 16d agoExplain → -
RAG-Eval-System
A hybrid-retrieval RAG system built specifically to practice the part most RAG projects skip: rigorous evaluation. The pipeline (ingest → retrieve → generate) is intentionally simple — the actual point of this repo is the eval harness and the decisions it drove.
Python ★ 0 20d agoExplain → -
tau ⑂
a minimalist agent that teaches you to create coding agents
★ 0 23d agoExplain → -
hesoyam-finance-tracker
Automated personal finance tracker: Flutter Android app + React web dashboard + Supabase
Dart ★ 0 25d agoExplain → -
groove
A real-life fog of war exploration map — GTA-style map reveal for the real world
Dart ★ 0 1mo agoExplain → -
Making-an-Agent-from-scratch
No description.
Python ★ 0 1mo agoExplain → -
ReviewerAgent5000
No description.
Python ★ 0 1mo agoExplain → -
Agent-Reliability
No description.
Python ★ 0 1mo agoExplain → -
My_Harness
No description.
Python ★ 0 1mo agoExplain → -
Speculative-Decoding
No description.
Jupyter Notebook ★ 0 2mo agoExplain → -
Reverse-Proxy-using-Nginx
No description.
★ 0 2mo agoExplain → -
Fine-Tuning-HuggingFace
No description.
Jupyter Notebook ★ 0 2mo agoExplain → -
Running_simple_docker_dashboard
No description.
★ 0 2mo agoExplain → -
1-bit-LLM
No description.
Jupyter Notebook ★ 0 2mo agoExplain → -
building-llm-applications ⑂
No description.
★ 0 8mo agoExplain → -
RAG-WITH-SUMMARY-EMBEDDINGS
No description.
Python ★ 0 3mo agoExplain → -
Naive-RAG-using-LangChain
A simple yet powerful implementation of a Retrieval-Augmented Generation (RAG) pipeline built using LangChain, designed to help you understand how LLMs can interact with custom data.
Python ★ 0 3mo agoExplain → -
Embeddings_Visualizer
No description.
Python ★ 0 3mo agoExplain → -
Ongaku
No description.
Python ★ 0 3mo agoExplain → -
Auto_theft_Project
No description.
Jupyter Notebook ★ 0 5mo agoExplain → -
Super_Store_PowerBi_Analysis
No description.
★ 0 5mo agoExplain → -
Netflix_Data_Analysis
No description.
Jupyter Notebook ★ 0 6mo agoExplain → -
sql-music-store-analysis
No description.
★ 0 6mo agoExplain → -
Celestia_Hackathon
Hackathon Stuff LOL
Python ★ 0 10mo agoExplain → -
Machine-Learning-Projects
All my Machine Learning Projects
Jupyter Notebook ★ 0 10mo agoExplain → -
Air-India-Flight-
It will remove all Air India flights from any flight booking platform
JavaScript ★ 0 1y agoExplain →
No repos match these filters.