An orchestrated fleet of specialized SLMs/LLMs running on consumer hardware — zero cloud dependencies.
DeepThink AIOS is an enterprise-grade, fully local multi-agent AI Operating System that routes user queries across specialized neural pipelines for software engineering, theoretical mathematical reasoning, machine learning forecasting tournaments, 2-volume university textbook authoring, 3D physical semiconductor layout synthesis (from 180nm planar to 2nm GAAFET), and hardware-accelerated benchmarking — all running locally with dynamic hardware scaling from Intel iGPUs to NVIDIA H100s.
| Specification | Minimum (Lightweight LLMs) | Recommended (Full Swarm Fleet) |
|---|---|---|
| System RAM | 8 GB RAM (using 1.5B–3B quants) | 16 GB – 32 GB RAM (for 7B–9B quants) |
| GPU VRAM | Integrated iGPU / 2–4 GB VRAM | 8 GB – 16 GB+ VRAM (Vulkan / CUDA / Metal) |
| Storage | 10 GB free disk space | 30 GB SSD space for full local GGUF fleet |
| OS | Linux (Ubuntu/Debian), macOS (Apple Silicon), Windows 10/11 | Linux / Kaggle Cloud VM / macOS |
-
🎓 Master 2-Volume Study Engine (
🎓 Study) — Pedagogical textbook synthesis powered by DeepSeek-R1. Authors 15–20 page master reference books with centered display KaTeX formulas ($$ ... $$), embedded Mermaid architectural flowcharts, pedagogical alert callouts (> [!TIP],> [!IMPORTANT]), a 1-Page High-Yield Formula Cheat-Sheet, multi-dimensional comparison tables, a 10-Problem Solved Question Bank, and a Standardized University Mock Exam. Supports direct PDF/Slide ingestion. -
🔮 Maximum Power Prediction Engine (
🔮 Predict) — High-precision time-series forecasting across Financial Markets, Climate/Weather, Energy & Battery SOH decay, and Cloud Telemetry. Features a 14-Signal Alpha Feature Space ($\text{RSI}_{14}$ ,$\text{MACD}$ Histogram, Bollinger Bandwidth, Fourier Harmonics, News Sentiment Decay), an 8-Algorithm Tournament with Bayesian Softmax Inverse-Loss Stacking ($\beta = 3.5$ ), and Conformal Prediction Probabilistic Uncertainty Bands ($80%$ &$95%$ corridors in Plotly). -
🔬 Scientific Semiconductor EDA & 3D Physical Die Visualizer — Synthesizes synthesizable Verilog/SystemVerilog HDL and SPICE netlists across all process nodes (180nm Planar to 2nm RibbonFET / GAA Nanosheets). Integrates real transistor physics (
$L_g = 12\text{nm}$ ,$W_{\text{eff}} = 240\text{nm}$ ,$\text{HfO}_2$ dielectric, Subthreshold Swing$S = 65\text{mV/dec}$ , BSPDN$\Delta V = 11.8\text{mV}$ ), semi-transparent low-k$\text{SiCOH}$ ($k=2.2$ ) glass, and an interactive Alpha-Power Law live clock simulation toolbar ([ ▶ Run ],[ ⏸ Pause ],[ ⏭ Step ]). -
⚡ Zero-Hallucination Mathematical Reasoning Engine — Solves complex calculus, differential geometry, and theoretical physics proofs with pure 2-stage theoretical KaTeX derivations or Program-Aided Language (PAL) SymPy CAS sandbox execution where exact code computations override conversational text guesses. Verifies general relativistic proofs using Kretschmann Curvature Scalar Invariants (
$K = \frac{48G^2M^2}{c^4 r^6}$ ) and asymptotic boundary limits (Minkowski & Newtonian). -
💻 Production-Grade Autonomous Coding Pipeline — Multi-phase software engineering with Big-O complexity optimization (
$O(N)$ / $O(N \log N)$), automated AST Static Analysis Linting (SAST), strict type annotations, Google-style docstrings, C++17 shared mutex concurrency (std::shared_mutex, lock-free SPSC queues), and multi-language execution sandboxes. -
📊 Benchmark Studio & Telemetry Dashboard — Parallel evaluation across 11 standard suites (HumanEval, MBPP, GSM8K, MATH, GPQA, AIME, MuSR, MMLU-Pro, SWE-bench Lite, SWE-bench Pro, SearchQA) with real-time scoring vs GPT-4o and Claude 3.5 Sonnet baselines, live throughput (
$\text{tok/s}$ ), and JSON report exports. -
🌐 100% Keyless Multi-Tier Web Search (
🌐 Search&🔬 Extreme) — Scrapes live financial quotes, real-time weather, and multi-source academic publications with deep synthesis without API keys. - ⚡ Elastic VRAM Management (EVM) & DMA — Zero-cost dynamic model hot-swapping between System RAM and GPU VRAM with CPU-to-GPU cache promotion.
| Pipeline | Example Prompt to Try in the UI | Key Output Artifacts |
|---|---|---|
| 🔬 Chip Design | Design a 2nm GAAFET TPU with an 8x8 Systolic Array of Bfloat16 PEs, Backside Power Delivery (BSPDN), synthesizable Verilog, and 3D silicon layout. |
Synthesizable RTL, self-checking testbench, and interactive 3D WebGL silicon die with live clock stepping |
| 🔮 Prediction | Predict the price trajectory of Bitcoin (BTC-USD) over the next 15 days based on macro momentum and FinBERT news sentiment. |
14-Signal Alpha Features, 8-Model Bayesian Softmax Tournament, and 80%/95% Conformal Uncertainty Plotly curve |
| ⚡ Reasoning | Derive the Schwarzschild metric from Einstein's field equations R_uv = 0, computing all Christoffel symbols, Newtonian limit, and Kretschmann invariant. |
Publication-grade KaTeX derivation ($$ ... $$) with Kretschmann scalar invariant |
| 💻 Coding | Implement a high-throughput thread-safe LRU Cache in Python with TTL expiration, O(1) ops, complete type hints, docstrings, and unit tests. |
Production-grade Python module, AST linted, with self-testing test harness passing in local sandbox |
| 🎓 Study | Teach me Transformer Attention Mechanism (Self-Attention, Multi-Head, KV-Cache) from first principles as an exhaustive graduate textbook. |
2-Volume Master Treatise with Mermaid architecture diagram, display math, 1-page formula cheat-sheet, 10 solved problems & mock exam |
| System Role | Model Name | HuggingFace Repo ID | GGUF Filename & Projector | Quants |
|---|---|---|---|---|
| Master Router | Phi-3.5-Mini / Llama-3.2 | bartowski/Phi-3.5-mini-instruct-GGUF |
Phi-3.5-mini-instruct-Q6_K.gguf |
Q6_K / Q4_K |
| Agentic Coder | Qwen2.5-Coder / Ornith | deepreinforce-ai/Ornith-1.0-9B-GGUF |
ornith-1.0-9b-Q6_K.gguf |
Q6_K / Q4_K |
| Reasoning Engine | DeepSeek-R1 Distill | unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF |
DeepSeek-R1-Distill-Qwen-7B-Q6_K.gguf |
Q6_K / Q4_K |
| Syntax Linter | VibeThinker 3B | prithivMLmods/VibeThinker-3B-GGUF |
VibeThinker-3B.Q6_K.gguf |
Q6_K / Q4_K |
| Vision & OCR | Qwen-2.5-VL / Qwen3-VL | unsloth/Qwen2.5-VL-7B-Instruct-GGUF |
Qwen2.5-VL-7B-Instruct-UD-Q6_K_XL.gguf + mmproj-BF16.gguf |
Q6_K / Q4_K / Q8_0 |
flowchart TD
%% ── TOP-LEVEL INGESTION ──
USER([User Prompt / Image / PDF]) --> MODE_CHECK{"Pipeline Mode Selected?"}
MODE_CHECK -->|🎓 Study| STUDY_PIPE["Study Pipeline: 2-Volume Master Curriculum + Mermaid + 10 Problems + Exam"]
MODE_CHECK -->|🔮 Predict| PREDICT_PIPE["Predict Pipeline: 14-Signal Alpha Features + 8-Model Bayesian Tournament"]
MODE_CHECK -->|🔬 Extreme| EXTREME_PIPE["Extreme WebSearch: Multi-Source Academic Survey"]
MODE_CHECK -->|🌐 Search| SEARCH_PIPE["Simple Search: Live Real-Time Web Data"]
MODE_CHECK -->|📊 Benchmark| BENCH_PIPE["Benchmark Studio: 11-Suite Parallel Worker Evaluation"]
MODE_CHECK -->|Auto / Prompt| ROUTER["Fast-Path & Router Intent Classifier"]
%% ── Intent Classification Branches ──
ROUTER --> PATH_CODING["1. CODING (AST Linter + O(N) Complexity + Type Hints)"]
ROUTER --> PATH_REASONING["2. REASONING (SymPy CAS Grounding & Kretschmann Scalar)"]
ROUTER --> PATH_CHIP["3. CHIP DESIGN (2nm GAAFET, BSPDN & 3D Live Clock)"]
ROUTER --> PATH_VISION["4. VISION & OCR"]
ROUTER --> PATH_SIMPLE["5. DIRECT / CONVERSATIONAL"]
%% ── Execution Pathways ──
STUDY_PIPE --> STUDY_OUT["Master Textbook + 1-Page Cheat Sheet + 10 Solved Problems + Mock Exam"]
PREDICT_PIPE --> PREDICT_OUT["8-Model Bayesian Stacking + 80%/95% Conformal Fan Chart"]
BENCH_PIPE --> BENCH_OUT["Real-Time Throughput / Accuracy Telemetry vs Baselines"]
PATH_CODING --> CODE_SB{"Execution Sandbox"} --> CODE_PASS["Verified Working Polyglot Code"]
PATH_REASONING --> PAL_SB{"SymPy / Math Sandbox"} --> PAL_PASS["Verified KaTeX Proof ($$ ... $$)"]
PATH_CHIP --> EDA_SB{"Icarus / Yosys / SPICE"} --> CHIP_OUT["Verilog Module + Interactive 3D Die Visualizer"]
git clone https://github.com/Arpit104147/DeepThink-AIOS.git
cd DeepThink-AIOS
# Launch servers (Backend on :8000, Web UI on :5173)
./start.shOpen http://localhost:5173 in your browser.
Paste and run this complete Python script inside a single Kaggle Notebook cell:
# =========================================================================
# 🚀 DEEPTHINK-AIOS: KAGGLE BACKEND + CLOUDFLARE TUNNEL (CONTINUOUS RUNNER)
# =========================================================================
import os, subprocess, time, re, sys
# 1. Clone or Auto-Update Repository
if os.path.exists("/kaggle/working/DeepThink-AIOS"):
os.chdir("/kaggle/working/DeepThink-AIOS")
subprocess.run(["git", "pull", "origin", "main"], check=True)
else:
os.chdir("/kaggle/working")
subprocess.run(["git", "clone", "https://github.com/Arpit104147/DeepThink-AIOS.git"], check=True)
os.chdir("/kaggle/working/DeepThink-AIOS")
# 1.5 Install System EDA Chip Design Tools (Icarus Verilog, Yosys, NGSPICE, KLayout)
print("🔬 Installing System EDA Chip Design Tools (iverilog, yosys, ngspice, klayout)...", flush=True)
subprocess.run(["apt-get", "update", "-y", "-q"], check=False)
subprocess.run(["apt-get", "install", "-y", "-q", "iverilog", "yosys", "ngspice", "klayout"], check=False)
# 2. Install dependencies & Pre-compile CUDA llama-cpp-python for Kaggle GPU
print("⚡ Installing requirements & pre-compiling CUDA llama-cpp-python for Kaggle GPU...", flush=True)
subprocess.run([sys.executable, "-m", "pip", "install", "-r", "requirements.txt", "-q"], check=True)
try:
import torch
if torch.cuda.is_available():
print("🔥 Pre-installing CUDA-accelerated llama-cpp-python for Kaggle GPU...", flush=True)
env = os.environ.copy()
env["CMAKE_ARGS"] = "-DGGML_CUDA=on"
env["FORCE_CMAKE"] = "1"
subprocess.run([
sys.executable, "-m", "pip", "install",
"llama-cpp-python", "--force-reinstall", "--no-cache-dir", "-q"
], env=env, check=False)
except Exception as e:
print(f"⚠️ CUDA setup note: {e}")
# 3. Download Cloudflare Tunnel binary
subprocess.run(["wget", "-q", "https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64", "-O", "/tmp/cloudflared"], check=True)
subprocess.run(["chmod", "+x", "/tmp/cloudflared"], check=True)
# 4. Launch FastAPI Backend
print("⏳ Launching FastAPI Backend on Port 8000...", flush=True)
backend_proc = subprocess.Popen([sys.executable, "-m", "uvicorn", "backend.app:app", "--host", "0.0.0.0", "--port", "8000"])
# 5. Create Cloudflare Tunnel
print("🌐 Creating Secure Cloudflare Tunnel...", flush=True)
tunnel_proc = subprocess.Popen(
["/tmp/cloudflared", "tunnel", "--url", "http://localhost:8000"],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True
)
time.sleep(4)
# 6. Extract & Print Public URL
public_url = None
for line in tunnel_proc.stdout:
match = re.search(r"https://[a-zA-Z0-9-]+\.trycloudflare\.com", line)
if match:
public_url = match.group(0)
print("\n" + "="*72, flush=True)
print("🎉 YOUR KAGGLE BACKEND PUBLIC URL:", flush=True)
print(f"👉 {public_url}", flush=True)
print("="*72, flush=True)
print("📌 COPY the URL above and paste it into your local Laptop Frontend!", flush=True)
print("="*72 + "\n", flush=True)
break
# 7. Continuous Heartbeat to keep Kaggle alive overnight
start_time = time.time()
print("⚡ Backend is ACTIVE & serving requests continuously...", flush=True)
try:
while True:
time.sleep(120)
elapsed_min = int((time.time() - start_time) // 60)
print(f"💓 [HEARTBEAT - {elapsed_min}m elapsed] DeepThink-AIOS Backend Running | URL: {public_url}", flush=True)
except KeyboardInterrupt:
print("Stopping server...", flush=True)
backend_proc.terminate()
tunnel_proc.terminate()Copy the printed https://xxxx.trycloudflare.com URL, open http://localhost:5173 in your local browser, click ⚙️ Settings, and paste the URL into Server URL.
- Backend: FastAPI, Uvicorn, Python 3.10+, PyTorch, Vulkan SDK,
llama-cpp-python, ChromaDB, PyPDF, Scikit-Learn, Icarus Verilog, Yosys, NGSPICE, SymPy, NumPy, Pandas - Frontend: React 18, Vite, KaTeX Mathematical Typography, Plotly.js, Three.js / WebGL, Vanilla CSS (Glassmorphism)
- Hardware Acceleration: Vulkan Compute (NVIDIA, AMD, Intel iGPU/dGPU), NVIDIA CUDA, Apple Metal, Multi-Core CPU Fallback
MIT License — see LICENSE for details.