Four privacy-first SaaS products. 95 SEO pages.
Autonomous content + reply pipelines. All on a single Mac Mini M4
Pro in India. Here's the architecture, the stack, the cost, and
why it works.
Live snapshot
Products live
4
Services running
74
Uptime
3 days
RAM free
0.2 GB
Disk free
13Gi
Blog posts
4
Changelog entries
10
Memory episodes
2941
This snapshot regenerates on every page request. The numbers above
are the actual values reported by `vm_stat`, `df`, `launchctl list`,
and the SaaS registry SQLite — not synthetic placeholders.
Each of the four SaaS frontends is a Next.js 15 + React 19 app on
Vercel. None of them run any LLM logic in their own runtime. Every AI
call (Whisper transcription, Llama-3.1 generation, sentence embedding,
Playwright PDF render) is a fetch over the Cloudflare Tunnel to the
Mac Mini, which has the model files, the GPU access, the persistent
state.
Stack — exact versions
Component
Implementation
Hardware
Mac Mini M4 Pro · 64 GB unified · ARM64 · 460 GB SSD
OS
macOS 25 · launchd-managed daemons
Python
3.11 (FastAPI · Uvicorn · pydantic)
Node
22 LTS (Next.js 15 + React 19)
Local LLM
llama3.1:8b, llama3.1:70b
Speech-to-text
OpenAI Whisper (base / small)
Embeddings
sentence-transformers all-MiniLM-L6-v2 (22 MB)
Vector store
ChromaDB persistent (per-doc collection)
HTML→PDF
Playwright Chromium · Inter font
HTML→MP4
HyperFrames (HeyGen, Apache-2.0)
Pub-sub / queues
Redis local (no RDS)
SQL
SQLite (WAL mode) · per-feature DB
Egress
Cloudflare Tunnel → anantatrade.com
Frontend hosting
Vercel (free tier × 4 projects)
Billing
Razorpay live (INR + UPI native)
Email
Telegram bot for ops alerts (cheaper, faster)
Memory layer
5-tier (Redis L1, SQLite L2, Graph L3, Wiki L4, Character L5)
What runs autonomously
Daily 08:00 IST — morning briefing Telegram (jobs digest +
affiliate activity + Razorpay events + Mac Mini health)
Daily 09:00 IST — YouTube Shorts variant drip (industry +
Indian-language content, 6 days queued)
Daily 11:00 IST — blog→Bluesky thread mirror (3-post
thread per new article)
Daily 14:00 IST — blog→YouTube Short (HyperFrames key-
points reveal)
Daily 23:55 IST — auto-generate the day's changelog entry,
IndexNow ping
Every 30 min — Bluesky reply-bot polls 14+ tracked posts
for new comments, drafts replies via local Llama
Every 2 h — affiliate orchestrator picks deals, posts via
HyperFrames price-drop video
Sunday 18:00 IST — Worldmonitor weekly digest video →
Bluesky × 2 + YouTube
Comparing this Mac Mini stack to the equivalent provisioned in
AWS / OpenAI / Vercel Pro:
Component
Cloud price (US)
This stack (one-time + recurring)
4-product SaaS hosting
Vercel Pro 4 × $20/mo = $80/mo
Vercel free × 4 = $0
LLM API (Llama-3.1 70B equiv)
OpenAI GPT-4o · ~$0.50/M tokens · realistically $200-500/mo
at moderate use
Local Ollama on M4 Pro · electricity ~$3/mo
Speech-to-text (Whisper-equiv hours)
Deepgram $0.43/hr · 100 hr/mo = $43/mo
Local Whisper · $0
Vector DB (10k vectors, 4 products)
Pinecone Standard $70/mo
ChromaDB persistent · $0
Hardware amortisation
n/a
Mac Mini M4 Pro $1,599 / 36 mo = $44/mo
Egress + DNS
$15-30/mo
Cloudflare Tunnel free · $0
Email (transactional)
Resend / Mailgun $30+/mo
Telegram bot · $0
Total
~$438-758 /mo
~$47 /mo (mostly hardware amortisation)
Operating margin at ₹399/mo per Notes customer:
customer pays ~$5, this stack costs ~$0.50/customer-month
marginal compute. ~90% gross margin, sustainable from
customer #1.
Privacy — what we don't do
Resume + JD never sent to OpenAI / Anthropic / Google
(ATS Tailor uses local Llama-3.1 only)
Audio file deleted within minutes of processing
(ANANTA Notes — temp directory auto-cleanup)
Documents indexed only in a per-doc ChromaDB collection on
the Mac Mini (ANANTA Docs — no cloud vector DB)
Slide topic + outline never fed to a third-party LLM
(ANANTA Slides — local Llama outline + Playwright PDF)
Free-tier outputs auto-purged on a strict TTL
(1 hour for Slides decks, 24 hours for Docs)
No web analytics on the apex (no Google Analytics,
no Meta pixel, no third-party tracker)
Why this works for a solo operator
Three structural advantages over the typical SaaS founder:
One stack, four products. Same Llama-3.1, same Whisper,
same Playwright, same ChromaDB. Adding a 5th SaaS is mostly a
new prompt + frontend, not a new infra layer.
Indian-tier pricing on Razorpay. ₹299–2,400/mo plans are
impulsive purchases for the target buyer; the same value at
US-tier $20+/mo would have 10× higher friction.
Privacy as a moat. Cloud-LLM SaaSes literally cannot
match the "your data never leaves the box" claim. They built on
OpenAI/Anthropic; we built on hardware we own.
Try the products
ATS Resume Tailor
— Llama-3.1 rewrites your resume against any JD in 3 seconds