MANAS-MFG

MANAS-MFG is a helpful AI assistant for engineering and manufacturing. It computes internally (arithmetic, formulas, math), recalls its own knowledge, thinks step-by-step, explains algorithms, writes code, and chats — all without needing external tools.

Author: @k_k_kaundal on X

What it does

• 💬 Natural conversation and question answering
• ➕ In-weights arithmetic — solves math step-by-step internally (no calculator)
• 📐 Applies formulas — percent, algebra, geometry, calculus, physics, engineering (feed rate, F=ma, Ohm's law, torque, …)
• 🧠 Thinks & recalls — reasons step-by-step and recalls what it knows (and says "I don't know" when it doesn't)
• 🏭 Manufacturing knowledge — dimensions, tolerances, fits (H7/g6), GD&T, threads, machining, G-code
• 🧮 Algorithms — explains and applies DSA (sorting, searching, complexity)
• 🐍 Writes code — Python (and more)
• ✅ Self-corrects — catches and fixes its own mistakes

Files

• manas-mfg.pt — the model weights (PyTorch checkpoint, 430M params)
• tokenizer.model / tokenizer.vocab — the SentencePiece tokenizer

Usage

import torch
from huggingface_hub import hf_hub_download

# 1) download weights + tokenizer
model_path = hf_hub_download("kaundalvip/manas-mfg", "manas-mfg.pt")
tok_path   = hf_hub_download("kaundalvip/manas-mfg", "tokenizer.model")

# 2) load the checkpoint
ckpt = torch.load(model_path, map_location="cpu")
# ckpt["model"] -> the model weights (state_dict)
# ckpt["cfg"]   -> the model configuration

# 3) build your model from ckpt["cfg"], load ckpt["model"] weights,
#    and generate text with the tokenizer (SentencePiece).

Examples

You: What is 931 * 49?
MANAS: 931*40 = 37240, 931*9 = 8379, sum = 45619. 931 * 49 = 45619.

You: A mass of 10 kg accelerates at 3 m/s^2. Force?
MANAS: Using F = m x a: the answer is 30 N.

You: RPM=8000, chip load 0.05 mm, 2 flutes. Feed rate?
MANAS: feed = RPM x chip load x flutes = 8000 x 0.05 x 2 = 800 mm/min.

You: What kind of fit is H7/g6?
MANAS: H7/g6 is a clearance/running fit: hole basis H7, shaft g6.

You: Explain binary search and its complexity.
MANAS: Binary search halves a sorted array each step. O(log n) time.

Notes

• 430M parameters, lightweight and efficient — runs on a laptop.
• Best at short, direct questions in engineering, math, code, and everyday conversation.

Connect

Built by K K Kaundal — follow on X: @k_k_kaundal