Short answer: my Junagadh lab ships production MCP agents past midnight on a ₹6K VPS with Valkey, n8n, Postgres, and a 90-day ledger — action P95 1.8 seconds, blended task cost ₹4.10, two clients live in Surat and Rajkot. The timestamps below are real: 06:00, 13:00, 18:00, and 01:00. Failures included.
I am Deepak Bagada. I work from a two-room setup in Junagadh, Gujarat — one room for servers and screens, one for sleep I sometimes skip. No co-founders. No funding. Just a fiber line, a ledger habit, and clients who found me through proof, not ads. This is what shipping actually looks like from here.
06:00 — ledger review with morning tea
The day starts with yesterday's numbers, not code. My nightly script prints tasks, P95 per model, and total rupee burn. September 16 read: 1,184 tasks, blended ₹4.10, one alert — Fable cache hit rate dipped to 0.71 overnight.
I traced it before breakfast. A Rajkot catalog prompt had a date string injected at the top, breaking prefix stability. Same bug class as my September 5 incident. Moved the date to the bottom. Hit rate recovered to 0.83 by 07:30.
Boring work. This is the work. Agents rot through prompt drift, not dramatic outages. The ledger catches drift while it is cheap.
My standard stack for this life is documented in my AI agent development service. Finished client systems are in my project log.
The lab stack: ₹6K and a ledger
| Layer | What runs | Monthly cost | Why this choice |
|---|---|---|---|
| VPS, 8 vCPU | Hetzner Mumbai box | ₹6,000 | Single box, full control |
| Postgres plus pgvector | Memory, checkpoints, ledger | Included | One database, three jobs |
| Valkey | Tool cache, sessions | Included | Faster than round trips |
| n8n | Cron checks, alerts | Included | No-code watchdog |
| Model APIs | Flash, Fable, Astra, Spark | ~₹98,000 at volume | Blended ₹4.10 per task |
| Domain plus backups | S3 snapshots nightly | ₹900 | Sleep insurance |
Total fixed base: under ₹7K monthly before API usage. A metro office pantry costs more. The money goes into model spend and backups, not rent.
Short version. Cheap base. Expensive discipline. The ledger is the product.
13:00 — Surat showroom demo over a hotspot
Afternoon brought the Surat textile showroom demo — evening-routine agent, lock plus lights plus thermostat, driven from the owner's phone. Their office Wi-Fi died ten minutes before the call. We ran the whole demo over my phone hotspot. Jio saved the deal.
The agent listed 14 devices across 5 rooms, ran the routine in 1.8 seconds P95, and answered "what happened while we were out" from history. The owner asked one sharp question: what stops it from unlocking the gate by mistake? I showed him the human-confirm tap and the OPA deny row in the ledger. He signed at ₹85K before the call ended.
Demo checklist I now follow, learned the hard way:
#!/usr/bin/env bash
set -euo pipefail
# pre-demo gate — abort the call if any check fails
curl --fail --silent https://bridge.example.com/api/health
python3 /tmp/demo_smoke.py
echo "P95 last hour:"
grep "$(date +%Y-%m-%d)" model_ledger.jsonl | python3 /tmp/p95_check.py
# /tmp/demo_smoke.py — 60-second pre-demo verification
import json, urllib.request
checks = [
("health", "https://bridge.example.com/api/health"),
("structure", "https://bridge.example.com/api/structure"),
]
for name, url in checks:
req = urllib.request.Request(url, headers={"Authorization": "Bearer DEMO_KEY"})
code = urllib.request.urlopen(req, timeout=10).status
print(name + "=" + str(code))
print("smoke done")
{
"demo_kit": "hotspot pluseshow",
"backup_net": "phone hotspot, tested",
"abort_if": "any smoke check fails",
"rule": "never debug live in front of a client"
}
For the security layer behind that demo answer, see my MCP security checklist. For model cost math I quoted on the call, my frontier routing guide has the table.
18:00 — Rajkot apartment routine goes live
Evening was the Rajkot service-apartment cutover: overnight doorbell plus motion summaries in Gujarati to WhatsApp at 07:00. Real residents. Real doorbell. No staging net.
Cutover plan: old manual process ran parallel for three nights while the agent posted shadow summaries to my phone only. Night one matched. Night two matched. Night three caught an event the manager missed — a 03:12 gate entry with no logbook line. The manager went pale, then smiled. Go-live approved on the spot.
Build cost ₹55K. Time saved: 45 minutes daily for the manager. Payback at his salary rate: under two months. Numbers like that sell better than any feature list.
01:00 — the deploy that failed, then shipped
Midnight was supposed to be quiet. A billing-bot dependency bump turned red in CI at 00:20. Tool-list schema changed upstream; my adapter cache served the old list; three refund tests failed. Same stale-cache species as my ₹18,400 lesson, smaller teeth.
I flushed Valkey keys, dropped TTL to 120 seconds for the night, re-ran green at 00:55, deployed at 01:00. Watched the ledger for 40 minutes. Cache hit rate 0.81. P95 2.7s. Slept at 02:00.
Exact lines from that night: schema_mismatch tools=14 expected=15, then flush OK, then deploy 01:00 green.
No heroics. Just procedure. Flush, tighten, verify, ship, watch, sleep. Midnight deploys feel routine when the ledger watches with you.
Failures I keep on the wall
Two more, briefly, because founder stories that skip failures are ads:
The ₹18,400 refund lesson. Stale tool cache served refund_full for six hours after the server added refund_partial. I covered the difference. Now every server deploy flushes cache keys automatically. I wrote up the full migration separately.
The 11-hour paused graph. A bulk price approval interrupt keyed resume on session cookies under a stateless core. Nothing resumed until 06:00. Now resumes key on thread IDs in the payload. Approval P95 is 1.1 seconds.
Total tuition: ₹18,400 plus one lost night. Degrees cost more and teach less about caching.
To hire the lab behind these stories, contact me here.
Production Trade-offs: the honest limits of a one-person lab
Do not expect instant replies at 15:00. I am often heads-down in client work afternoons. I answer mornings and nights, and emergencies get a call. If you need a 10-person support desk with SLAs in legal language, hire the agency. You will pay triple and get the desk.
Do not ask for ten parallel projects. I run two client builds max at once. That is the capacity where quality holds. Anyone solo promising five parallel launches is promising five delays.
Do not skip the trial. My ₹15K five-day trial exists because both sides learn cheap. Full-commit contracts with strangers fail. Trials with deliverables work.
Do not confuse cheap base with cheap work. The ₹6K VPS is frugality, not low value. Builds are ₹55K to ₹85K because the engineering is senior, measured, and ledger-backed. The server is cheap. The judgment is not.
Frequently Asked Questions
What does your Junagadh AI lab actually run on?
One ₹6K Hetzner VPS in Mumbai with Postgres plus pgvector, Valkey, n8n, and Docker, serving Laravel APIs and Next.js frontends with LangChain MCP agents on top. Blended model cost is ₹4.10 per task across 1,200 daily tasks, action P95 is 1.8 seconds, and every call lands in a 90-day JSONL ledger with a nightly cost report.
How do clients in Surat and Rajkot work with a Junagadh developer?
Video calls for demos, WhatsApp for daily updates, screen-shared ledger reviews weekly, and on-site visits for go-lives within Gujarat. The Surat showroom signed over a hotspot demo and the Rajkot apartment cut over after a three-night shadow run — both without a single metro meeting.
What are your biggest production failures so far?
A stale tool cache that issued full refunds instead of partial ones for six hours, costing me ₹18,400 out of pocket, and a stateless-core resume bug that left a price-approval graph paused 11 hours. Both are fixed with deploy-hook cache flushes and payload-keyed thread resumes, and both are documented publicly with exact errors.
How do I start a project with your lab?
Start with the ₹15K five-day paid trial: a working demo against your data, one human approval gate, a P95 log, and a cost-per-task sheet, all credited against the build on hire. Then two weekly milestones at 40-40-20 payments. I take two client builds at a time maximum.
Bottom Line
06:00 ledger, 13:00 hotspot demo, 18:00 live cutover, 01:00 green deploy: my Junagadh lab ships agents on a ₹6K VPS with 1.8s P95 and a 90-day ledger — and the two failures on the wall are why clients trust the third deploy.