Private AI Platform: phase-wise plan

A developer sends a document or a support ticket to our API. Small models sort it, pull out the details, check them and suggest the next step. Personal data is hidden before anything is logged, and anything the platform is unsure about goes to a person.

I can build it in five phases between 20 November and 24 December. Our current servers can hold most of it, but none of them has a GPU, so we need to rent GPU servers for the models.

Manish Singh Parihar, 19 November 2026. This plan answers the specification of 16 November 2026.

Our servers today

I logged in to each server on 19 November and checked what it has.

ServerWhat it hasCan it run this?
31.97.203.241
Hostinger, main server
8 vCPU, 31 GB RAM with 24 GB free, 270 GB free disk, no GPU. It runs all our live apps and all our databases.It is big enough for the application side, but I would keep it out of this. It is our production server, and the spec asks for a clean install on Kubernetes.
72.61.241.215
Hostinger
I could not log in, because none of our keys work.It could hold the application side only. Hostinger servers have no GPU.
72.61.241.216
Hostinger, Woodle OS staging
I could not log in here either.It has the same limit as the one above.
139.59.36.22
DigitalOcean
The server's identity key has changed since we last connected, so I stopped and did not log in. It may have been rebuilt.It is too small, and we should first confirm who rebuilt it.

None of our servers has a GPU, and Hostinger cannot add one. The spec needs two GPUs with 24 GB each (for example two NVIDIA L4) and an application server with 8 vCPU and 32 GB RAM.

Where it would run

Now: development

I run the API, the queue, the database and the model gateway on my machine and a test server. The small models run on my Mac or on one GPU rented by the hour.

From week 4: reference

We rent two GPUs and one application server from one Indian GPU cloud, such as E2E Networks in Delhi or Chennai. The data stays in India, and both servers share a private network.

Last week: AWS test

We run the same release on AWS in Mumbai for two or three days to pass the AWS smoke test, then delete it.

Every model call goes through the gateway, so moving from my Mac to rented GPUs and then to AWS does not change the API.

Rough cost

ItemPriceMonthly
Two L4 GPUs, E2E NetworksAbout ₹50 to ₹70 per GPU per hourAbout ₹22,000 to ₹31,000 if they run only in office hours (10 hours a day, 22 days). It is about ₹72,000 to ₹1,00,000 if they run all month.
Application serverSame providerA few thousand rupees
AWS smoke testAbout $0.97 an hour per L4 server in MumbaiAbout $140 once, for three days on two servers

I checked these prices on 19 November from the providers' pages and price trackers. We should confirm them at checkout.

The five phases

Phase 0 · 20 to 21 November

Decisions and setup

  • I need your yes on the GPU budget and the provider, and I need to know whose AWS account we use.
  • I get access to the two Hostinger servers I could not log in to, or we leave them out.
  • I set up the code repository, the container builds and the list of pinned versions.

Done when we can rent a GPU by the hour and the repository builds.

Phase 1 · 24 to 28 November

The platform skeleton

  • The API checks each app's key and its permissions, limits how fast each app can call, and returns the error format from the spec.
  • Jobs can run in the background with a status to check, or straight away for small items. A repeated request returns the first result.
  • The model gateway sits in front of Qwen3.5-4B and Qwen3.5-9B, checks every answer against its expected shape, and records which model answered.
  • I generate the test data: 300 documents (invoices, KYC forms, claims and junk) and 400 tickets in English and Hinglish. I hold back 100 of each for acceptance.

Done when a test key can submit a job, check on it and get an answer from a model through the gateway.

Phase 2 · 1 to 5 December

Use case 1: document intake

  • The platform works out the document type, hides personal data, pulls out the fields, checks totals, dates and GSTIN and PAN formats, and picks the next step.
  • When a small model is unsure, the larger model tries, and then a person reviews it.
  • The evaluation runner scores every step against the test documents, so we see accuracy from the first week instead of at the end.

Done when the evaluation report shows a first score for every document step.

Phase 3 · 8 to 12 December

Use case 2: ticket triage, and the controls

  • The platform picks the team and the priority for each ticket, pulls out order numbers and emails, drafts a first reply and escalates urgent cases. It never sends a reply to a customer by itself.
  • I add signed callbacks, the feedback, usage and model-version endpoints, and deletion of results after 30 days.
  • Every request leaves a record that nobody can change. I add dashboards, logs and a key vault.
  • I build a test set of documents and tickets that carry hidden instructions, to check that nobody can trick the platform.

Done when every endpoint in the spec works and both use cases have a first score.

Phase 4 · 15 to 19 December

Reference install and tuning

  • I install the whole platform from code onto the rented servers with Kubernetes.
  • Only the API is reachable from outside. The model servers have no internet, and only the gateway can call them.
  • I tune the prompts and fallbacks until the quality targets pass on the held-back data.
  • I run 5 requests a second for 30 minutes to check speed.
  • I run three drills: swap in Gemma 4 for four steps, kill a model server during load, and show that a bad prompt is blocked before release.

Done when the quality, speed and security checks pass on our own servers.

Phase 5 · 22 to 24 December

AWS test, hand-over and acceptance

  • I run the same release on AWS and pass the smoke test and a quality spot check.
  • I scan the images and libraries for security issues and write the runbook.
  • I do a fresh install in one working day or less, then a full removal with a deletion log.
  • Someone who did not build it fixes five staged problems using only the runbook.

Done when the acceptance report shows every check from the spec with its result.

Risks

  1. Time. Five weeks is tight for one developer. If something slips, I would move the AWS test and the runbook drill to the first week of January before I would cut any quality check.
  2. The GPU decision. Each day after 21 November without a GPU pushes Phase 4 back by a day.
  3. Quality targets. Some targets are high for small models, such as 97% for document type, 98% for urgent tickets and 98% for personal data. The spec lets us revise a target after the first score with a written reason, and I will tell you in week 2 if I see a gap.
  4. Hindi and Hinglish. Small models do worse in these languages, so the test data must include enough of them to show the real accuracy.

What I need from you

  1. Can I rent two L4 GPUs and one application server from an Indian GPU cloud, at about ₹25,000 to ₹35,000 a month in office hours?
  2. Whose AWS account should the AWS test run in?
  3. Who will do the runbook drill in the last week?
  4. Who has the login for 72.61.241.215 and 72.61.241.216, and did anyone rebuild the DigitalOcean server recently?