# Papparti — Project Cost Estimate

**Prepared by:** Vector Logic (AI & software development, Odisha, India)
**Prepared for:** Papparti (client)
**Date:** 11 October 2026
**Scope:** On-demand any-service marketplace. An AI avatar is the primary interaction layer on both the
user side and the provider side. Speech-to-text (STT) and text-to-speech (TTS) are **100% self-hosted**
(Hindi and English, no cloud speech APIs). A reasoning language model runs continuously in the backend.
Duration-aware booking (30 min / 1 hr / 5 hr) with payments, provider split, and an immutable ledger.
A PWA (no native app), provider onboarding, working hours, and per-booking commission.

Vector Logic delivers **100% first-party development** and hands the system over. Three AI models are part
of the deliverable, each **retrained (LoRA/QLoRA fine-tune) on the client's own data**:

1. **STT model** — Hindi + English, fine-tuned on the client's audio and domain vocabulary.
2. **TTS model** — a client-brand voice, fine-tuned from a licensed base.
3. **Reasoning model** — fine-tuned on the client's service catalogue, pricing rules and Hinglish booking
   language; runs 24/7 in the backend.

Plus the complete software (marketplace app, booking engine, payments, ledger, admin, provider app) and
self-hosted deployment on **rented GPU** (the client's own hardware is GPU-less).

> **Model naming is confidential.** This document states parameter counts, dataset descriptions and
> capabilities only. The specific base models chosen by Vector Logic are commercially confidential and are
> deliberately not named here.

**Currency note — pinned FX rate:** all USD figures are converted to INR at
**1 USD = ₹96.7384** [5], a real mid-market rate captured from XE on 11 October 2026 (04:05 UTC).
Every INR figure in this document is that rate applied to a USD list price fetched live on the same day.

---

## 1. One-page summary

### 1a. One-time build cost (excludes GST)

| # | Component | INR (excl. GST) | Basis |
|---|---|---:|---|
| 1 | **Feasibility benchmark — Phase 0** (discovery + Hindi/AI benchmark, 1 rented 48 GB GPU, 1 month) | **4,47,967** | 3 work-weeks + 1 month GPU [1] |
| 2 | **Model work — four workstreams** (STT, TTS, reasoning LLM, **voice production**) | **30,90,471** | GPU-hours [9] + engineering |
| 3 | **Software — Phase 1** (voice core + booking engine, audio-only) | **16,25,205** | 12 work-weeks |
| 4 | **Software — Phase 2** (avatar + polish) | **10,83,470** | 8 work-weeks |
| 5 | **Software — Phase 3** (second vertical / locality) | **10,83,470** | 8 work-weeks |
| 6 | **Deployment + handover** (self-hosted bring-up, CI/CD, runbooks, knowledge transfer, 30-day warranty) | **5,41,735** | 4 work-weeks |
| | **ONE-TIME TOTAL (excl. GST)** | **₹78,72,318** | sum of the rows above |
| | **ONE-TIME TOTAL (incl. 18% GST)** | **₹92,89,335** | +18% [15] |

**Optional add-on (not in the total):**
| Phase 4 — Real-time lip-synced avatar (open-ended) | **₹13,54,338** | 10 work-weeks, priced as an option |

### 1b. Monthly running cost after handover

| Item | INR / month | Basis |
|---|---:|---|
| GPU serving — 1× 48 GB dedicated | 41,665 | RunPod RTX A6000 @ $0.59/hr × 730 hr [1] |
| Managed Postgres + pgvector | 5,891 | DigitalOcean 4 GiB / 2 vCPU @ $60.90/mo [12] |
| Redis | 1,935 | Upstash Fixed 1 GB @ $20/mo [11] |
| SFU / TURN (LiveKit) | 4,837 | LiveKit Cloud Ship tier @ $50/mo [13] |
| Object storage (Cloudflare R2) | 1,161 | ~$12/mo ESTIMATED (500 GB + ops) [14] |
| Monitoring + CI | 2,902 | ~$30/mo ESTIMATED — see §7 |
| Support retainer | 67,717 | 0.5 work-week / month @ agency rate |
| **MONTHLY TOTAL** | **₹1,26,108** | sum of the rows above |

**The headline answer for the CEO:** the expectation of a **₹20–30 lakh one-time fee is NOT defensible —
it is roughly one-third of the real cost.** The arithmetic gives **₹78.72 lakh excl. GST** (₹92.89 lakh
incl. GST) for the scope as defined, plus ₹13.54 lakh if the Phase 4 real-time avatar option is added.
See §8 for the full verdict.

---

## 2. GPU rental — live prices (the foundation of every compute number)

All figures below were read from the vendors' own live pricing pages on **11 October 2026**. The 48 GB-tier
GPU is the project's stated minimum for a pilot with a client-side avatar; the research's own floor is one
48 GB card.

| Provider | 48 GB GPU | Per hour (USD) | Per month (USD, ×730 h) | Per month (INR) | Source |
|---|---|---:|---:|---:|---|
| **RunPod** (Community Cloud) | RTX A6000 48 GB | **$0.59** | $430.70 | ₹41,665 | [1] |
| **Vast.ai** (on-demand, from) | RTX A6000 48 GB | **$0.28** | $204.40 | ₹19,773 | [2] |
| **Vast.ai** (median rate) | RTX A6000 48 GB | **$0.40** | $292.00 | ₹28,248 | [2] |
| **Lambda Labs** | RTX A6000 48 GB | **$0.80** | $584.00 | ₹56,495 | [3] |
| **E2E Networks** (India) | L40S 48 GB | **$1.20** | $876.00 | ₹84,743 | [4] |
| **E2E Networks** (India) | A40 48 GB | **$1.44** | $1,051.20 | ₹1,01,691 | [4] |

**Correction to prior research.** The project's earlier research cited RunPod RTX A6000 48 GB dedicated at
**~USD 387/month (INR 34,056 at ₹88/USD)**. That figure is now **stale**. RunPod's live page today lists
the RTX A6000 at **$0.59/hr**, i.e. **$430.70/month** — and at today's FX rate that is **₹41,665/month**,
about **22% higher** than the research stated, and further from it because the research also used an
outdated ₹88/USD rate. The planning number in this estimate is therefore the live one, ₹41,665/month [1].

**Which provider to use:** for training bursts, Vast.ai's on-demand floor ($0.28/hr, ₹19,773/mo) is the
cheapest live option [2]; its *median* rate ($0.40/hr) is the more realistic planning figure because the
floor machines are preemptible and constrained. For **24/7 production serving**, RunPod's dedicated
$0.59/hr [1] is the reference used throughout this estimate because it is a guaranteed-uptime dedicated
instance. A **fully India-hosted** option exists (E2E Networks, INR-billed) but is **2–2.4× the price**
per GPU-hour [4], which buys latency-to-India and INR invoicing but should be reserved for the serving
layer if that latency turns out to matter.

---

## 3. Per-phase breakdown

Engineering is priced in **work-weeks** at an Indian AI-agency rate. The rate basis is set out in §6.

### Phase 0 — Discovery + Hindi/AI feasibility benchmark
**1 rented 48 GB GPU, 1 month.**

| Line | Qty | INR |
|---|---:|---:|
| GPU rental — 1× 48 GB dedicated, 1 month | 730 hr @ $0.59 | 41,665 |
| Engineering — discovery, benchmark harness, Hindi STT/TTS/LLM evaluation, written feasibility report | 3 work-weeks | 4,06,302 |
| **Phase 0 total** | | **4,47,967** |

Phase 0 answers the one question the whole plan rests on: **does the chosen stack actually produce good
Hindi, on rented 48 GB hardware, under load?** It produces a measured go/no-go report before full
development spend.

### Phase 1 — Voice core + booking engine, audio-only
**10–14 weeks of engineering; priced at 12 weeks (midpoint).**

The self-hosted STT + TTS pipeline (Hindi + English), the backend reasoning model endpoint, the
duration-aware booking engine (30 min / 1 hr / 5 hr), payments, provider split, the immutable ledger,
provider onboarding and working hours, and the PWA shell.

| Phase 1 | Weeks | INR |
|---|---:|---:|
| Voice core + booking engine | 12 | **16,25,205** |

### Phase 2 — Avatar + polish
**6–10 weeks; priced at 8 weeks (midpoint).** The avatar interaction layer on both sides (client-side
viseme rendering, per the project's own decision), UX polish, and hardening.

| Phase 2 | Weeks | INR |
|---|---:|---:|
| Avatar + polish | 8 | **10,83,470** |

### Phase 3 — Second vertical / locality
**6–10 weeks; priced at 8 weeks (midpoint).** Onboarding a second service vertical or locality onto the
same platform — catalogue, pricing rules, provider cohort, localisation.

| Phase 3 | Weeks | INR |
|---|---:|---:|
| Second vertical / locality | 8 | **10,83,470** |

### Phase 4 — Real-time lip-synced avatar (**OPTION, open-ended**)
Priced as a standalone option and **excluded from the one-time total**. This is the highest-risk, least-
proven scope: true real-time neural lip-sync is server-video-cost territory (~9–11 MB/user-minute vs the
viseme stream's ~9 KB/min, per the project's own research), so it is quote-on-request with a defined
starting estimate.

| Phase 4 (option) | Weeks | INR |
|---|---:|---:|
| Real-time lip-synced avatar | 10 (starting estimate) | **13,54,338** |

### Deployment + handover
**4 work-weeks.** Infrastructure-as-code for the self-hosted stack on rented GPU, CI/CD, monitoring
wiring, runbooks, admin training, and a 30-day post-handover warranty.

| Deployment + handover | Weeks | INR |
|---|---:|---:|
| Bring-up, CI/CD, runbooks, KT, warranty | 4 | **5,41,735** |

---

## 4. Model workstream breakdown (three retrained models)

Each workstream is a distinct sale: **data preparation → LoRA/QLoRA fine-tune → evaluation →
packaging → endpoint serving setup.** GPU-hour figures are pinned to published QLoRA fine-tuning
benchmarks [9] and to the QLoRA paper's demonstrated feasibility of fine-tuning very large models on a
single 48 GB card. The base models are not named (confidential — see header).

| Workstream | Params | GPU-hours | GPU cost (INR) | Engineering (weeks) | Engineering (INR) | **Total (INR)** |
|---|---:|---:|---:|---:|---:|---:|
| **STT** — Hindi + English, domain audio/vocab | 1–2 B | 40 | 2,283 | 6 | 8,12,603 | **8,14,886** |
| **TTS** — client-brand voice | 0.3–0.5 B | 24 | 1,370 | 6 | 8,12,603 | **8,13,973** |
| **Reasoning LLM** — catalogue, pricing rules, Hinglish booking | ~32 B | 180 | 10,274 | 8 | 10,83,470 | **10,93,744** |
| **Voice production** — casting, direction, QC, master (eng) + studio recording (pass-through) | — | — | — | 2 | 2,70,868 + 97,000 | **3,67,868** |
| *of which: studio recording, ~20 h delivered audio* | — | — | — | — | *97,000 (pass-through)* | |
| **Model work total** | | **244** | **13,927** | **22** | **29,79,544** | **₹30,90,471** |

**Voice production is included because Vector Logic records the brand voice itself** (CEO decision,
11 Oct 2026). The studio recording line is the research's own modelled cost for ~20 h of delivered
audio at India studio rates — **₹64,000 (low) to ₹1,30,000 (high), ₹97,000 taken as the midpoint**
[see the assumptions section]. The engineering line covers casting, session direction, QC, retakes
and mastering. **This converts what was the single biggest unmitigated blocker in the model plan —
"voice cloning of a not-yet-chosen speaker is blocked on a human decision, a legal act and money,
none of it engineering" — into work we control**, and it is why the TTS workstream also grew by one
week: we now own our own data pipeline end to end rather than consuming a client recording.

**One obligation this creates, and it is not optional:** the voice artist must sign a **perpetual,
irrevocable, worldwide, transferable commercial release** in the client's favour. The model is
handed to the client at handover; if the release were revocable, the client's product would break
the day the artist withdrew. Legal review of that release is a Phase 0 item.

GPU-hours are charged at the live RunPod 48 GB rate of **$0.59/hr** [1]. The reasoning-model workstream
dominates because a ~32 B model fine-tuned over multiple epochs on a large Hinglish/catalogue dataset is
an order of magnitude heavier than the 7 B-class benchmark runs published in [9]; the STT workstream is
heavy on *data preparation and evaluation* (transcription, WER harness, Hindi vocab) rather than raw GPU
time. **The GPU compute is a small fraction (≈0.5%) of each workstream — the cost is the engineering,
data work and evaluation, not the rental.** This is the single most important thing to understand about
the model line items.

---

## 5. Monthly running cost after handover

| Item | Live / Estimated | INR / month | Basis |
|---|---|---:|---|
| GPU serving — 1× 48 GB dedicated, 24/7 | LIVE | 41,665 | RunPod @ $0.59/hr × 730 [1] |
| Managed Postgres + pgvector | LIVE | 5,891 | DigitalOcean 4 GiB/2 vCPU @ $60.90/mo [12]; Neon Launch usage-based alternative ~$15/mo typical [10] |
| Redis | LIVE | 1,935 | Upstash Fixed 1 GB @ $20/mo [11] |
| SFU / TURN (LiveKit) | LIVE | 4,837 | LiveKit Cloud Ship tier @ $50/mo [13]; self-hosted SFU on a VPS is a cheaper alternative |
| Object storage (Cloudflare R2) | ESTIMATED | 1,161 | ~500 GB @ $0.015/GB-month [14] + ops; egress is free [14]. Figure is an assumption, not a quote |
| Monitoring + CI | ESTIMATED | 2,902 | ~$30/mo placeholder — see §7 |
| Support retainer | ASSUMPTION | 67,717 | 0.5 work-week/month @ agency rate (§6) |
| **MONTHLY TOTAL** | | **₹1,26,108** | |

**Not included above:** payment-gateway fees, which are transaction-linked, not fixed — see §7. The
monthly GPU figure assumes **one** 48 GB card; the project's research notes that a production serving
tier with real concurrency may need **two** 48 GB cards (~₹83,000/month at RunPod's live rate [1]).

---

## 6. Engineering rate — assumption, stated explicitly

**Every engineering line in this document is priced at $35 per hour**, applied to a 40-hour work-week,
converted at the pinned rate [5]:

```
$35/hr × ₹96.7384        = ₹3,385.84 per hour
₹3,385.84 × 8 hr/day     = ₹27,087 per day
₹3,385.84 × 40 hr/week   = ₹1,35,434 per work-week
```

**Why $35/hr is defensible:**

- **Clutch's Software Development Pricing Guide** lists India at **$25–$49/hour** [7]. $35 sits inside
  that band and above its midpoint.
- **Eucalipse (2025)** puts India at **$18–40/hr, average $29**, and defines an **"India (Premium)"**
  tier at **$35/hr** [8] — the AI/ML agency tier, which is what this project requires.
- This is an **AI / voice / LLM engineering** engagement, not generic web development; the rate is set at
  the **premium** end of the Indian band deliberately.

**This is an ASSUMPTION Vector Logic's CEO must confirm**, because no vendor publishes a fixed public
rate card. If the correct internal rate is different, every engineering figure scales linearly and the
reader can re-run the arithmetic: **one work-week = ₹1,35,434 at $35/hr**. To re-run at any rate *R*
(USD/hour): `total = 36 engineering work-weeks × R × 96.7384 × 40`. (Total billable engineering across the
project is **36 work-weeks**: 3 + 12 + 8 + 8 + 4 phases, plus 19 model workstream weeks, excluding the
Phase 4 option.)

---

## 7. Plumbing and third-party fees

### 7a. Infrastructure plumbing (see §5 for prices)
Postgres + pgvector, Redis, an SFU/TURN layer (LiveKit or self-hosted equivalent), object storage,
monitoring and CI. Live prices are cited for Postgres [12], Redis [11], the SFU [13] and object
storage [14]; monitoring/CI and the R2 storage volume are **ESTIMATED** (assumptions, flagged as such,
not baked into a "verified" total).

### 7b. Payment gateway — Razorpay
Razorpay charges a single flat **2% platform fee per successful domestic transaction**, with **18% GST
on that platform fee**, and no setup, AMC or refund fees [6]. This is a **variable** cost that scales with
transaction volume, not a fixed monthly line, so it is **not added to the monthly total**. Worked
example: on ₹1,00,000 of settled GMV, the fee is **₹2,000 + 18% GST = ₹2,360**. This should be modelled
against the per-booking commission to confirm the take-rate stays positive — see §8.

### 7c. GST treatment of imported services
Cloud and GPU services bought from foreign providers (RunPod, Vast.ai, Lambda, and R2) are an **import of
services** under Section 2(11) of the IGST Act. The Indian recipient must **self-assess 18% IGST under
the Reverse Charge Mechanism**, pay it in cash, and generally claim it back as input tax credit [15]. Two
consequences for this estimate:

- **Foreign GPU/SaaS bills carry an 18% IGST reverse-charge obligation in cash**, recoverable as ITC if
  Papparti is GST-registered and the service is used for business [15].
- **India-hosted E2E Networks** invoices are domestic (INR, GST on the invoice) and avoid the
  reverse-charge mechanics, at the higher per-hour price noted in §2 [4].

Vector Logic's own development fee is an Indian supply of software development service (SAC 998314,
**18% GST**), so the one-time total is quoted **both excluding and including 18% GST** in §1a.

---

## 8. Verdict on the ₹20–30 lakh expectation

**The expectation is not defensible. It is roughly one-third of the real cost of the defined scope.**

The arithmetic, from the cited rows:

- **One-time total (excl. GST): ₹78,72,318** — that is **3.15× the ₹25 lakh midpoint** of the CEO's band.
- The band's **ceiling** (₹30 lakh) would cover only the **model work (₹30.90 lakh, four workstreams)**
  and **nothing else at all** — the retrained AI models plus the voice production they require now
  **exceed the entire ₹30 lakh ceiling on their own**, before a single line of marketplace software,
  booking engine, payments, ledger, admin, provider app, deployment or handover is written.
- To fit ₹30 lakh, the scope would have to be cut to roughly **one** of: the model work, or phases 0–1 of
  the software. It cannot cover both.

**Where the money actually goes** (as a share of the ₹78.72 lakh one-time total):

| Component | INR | Share |
|---|---:|---:|
| Software development (Phases 1–3) | 37,92,145 | 51% |
| Model work (4 workstreams, incl. voice production) | 30,90,471 | 39% |
| Feasibility benchmark (Phase 0) | 4,47,967 | 6% |
| Deployment + handover | 5,41,735 | 7% |
| **Total** | **78,72,318** | 100% |

**Why the band is low, in one line:** the CEO priced this like a marketplace app. It is a marketplace app
**plus three custom-trained AI models plus a self-hosted voice stack** — and the AI half of that (§4) is
both the differentiator and nearly the entire ₹20–30 lakh band on its own.

**Monthly after handover:** **₹1,26,108**, of which the largest single fixed item is the **support
retainer (₹67,717)** — deliberately larger than the entire GPU bill, because a live 24/7 voice system
needs ongoing engineering, not just hardware. The GPU serving cost (₹41,665) is the second largest.

**One margin check the client should run before signing:** Razorpay's 2% + 18% GST on the platform fee
[6] is a per-transaction cost. If Papparti's per-booking commission is below ~2.4% of booking value, the
gateway fee alone exceeds the platform's take on that transaction. Model the commission against §7b
before fixing the take rate.

**Recommendation:** present the honest number. **₹78.72 lakh excl. GST / ₹92.89 lakh incl. GST** for the
defined scope, **₹13.54 lakh** for the Phase 4 option, and **₹1.26 lakh/month** to run it. If the client's
budget is genuinely capped at ₹20–30 lakh, scope it down explicitly — the three-model workstream is the
line to protect, because it is the part no one else is selling them.

---

## 9. Assumptions

1. **FX rate** — 1 USD = ₹96.7384, XE mid-market, 11 Oct 2026, 04:05 UTC [5]. Pinned; all INR figures
   use it. Re-run if the rate moves materially.
2. **Engineering rate** — $35/hour, India AI-agency premium tier [7][8]. **CEO must confirm** (§6).
   Arithmetic shown so it can be re-run at any rate.
3. **Work-week** — 40 hours, 5 days. A work-week = ₹1,35,434 at the assumed rate.
4. **Phases priced at midpoint** of their given week ranges: Phase 1 at 12 of 10–14 weeks; Phases 2–3 at
   8 of 6–10 weeks each.
5. **GPU hours/month** — 730 (24×7 dedicated). Training GPU-hours are per §4.
6. **Training GPU** — RunPod RTX A6000 48 GB @ $0.59/hr [1] for all training and serving figures.
7. **Model parameter counts** — STT 1–2 B, TTS 0.3–0.5 B, reasoning ~32 B. Base models confidential.
8. **Support retainer** — 0.5 engineering work-week per month.
9. **Monitoring/CI and R2 volume** — estimated, not quoted (§7a).
10. **Payment gateway fees** — excluded from the fixed monthly total; modelled as variable (§7b).

---

## 10. Explicitly UNVERIFIED (excluded from every total)

The following lines are **assumptions or published-benchmark estimates, not live quotes**. They are
**not** added to any verified sum; where they appear in a table they are labelled, and the totals in §1
stand independently of them.

- **Training GPU-hour counts for the three models** (40 / 24 / 180 hours in §4) — *not a vendor quote.*
- **Studio recording cost (~₹97,000 for 20 h at ~₹4,850/delivered hour)** — the research modelled a
  **₹64,000–₹1,30,000 band** from India studio rates; the midpoint is used. *No studio quote obtained.*
  They are engineering estimates anchored to published QLoRA benchmarks (7 B QLoRA = $0.75–$1.50 per run
  on a single 24 GB card, 70 B QLoRA = $24–$64 per run) [9] and to the QLoRA paper's 48 GB feasibility
  result. The *rates* used ($0.59/hr) are live [1]; the *hours* are estimated. **The cost of these lines is
  robust anyway — GPU compute is ~0.5% of each workstream.**
- **Monitoring + CI monthly cost (~₹2,902 / ~$30)** — an assumption, not a quote.
- **Object-storage volume (~500 GB → ~₹1,161/mo)** — the *rate* ($0.015/GB-month, free egress) is live
  [14]; the *volume* is assumed.
- **Any second 48 GB serving GPU** — flagged as a possibility, not priced into the totals.
- **Phase 4 real-time lip-synced avatar** — open-ended; the ₹13.54 lakh is a **starting estimate**, not a
  quote, and is excluded from the one-time total.

---

## 11. What could not be verified

- **A published Indian dev-agency rate card.** No vendor publishes a fixed public rate for AI engineering;
  the $35/hr figure rests on the Clutch India band [7] and the Eucalipse India-premium tier [8], and is
  recorded as an assumption the CEO must set (§6).
- **A quoted monthly/annual price from any GPU vendor.** RunPod, Vast.ai and Lambda publish **per-hour**
  prices only [1][2][3]; monthly figures here are per-hour × 730 and are labelled as such.
- **E2E Networks monthly plans.** E2E lists India per-hour prices [4] but its monthly/annual plans are
  "Contact Sales" — not quotable.
- **Vendor-specific latency for India-served inference.** Not measured; it is the reason the (pricier)
  India-hosted option in §2 is kept as a contingency, not the default.
- **Concurrency capacity of a single 48 GB card under real load.** The project's own research flags this as
  unmeasured; the monthly GPU line assumes single-stream serving.

---

## Sources

[1] https://www.runpod.io/product/cloud-gpus — Runpod GPU Cloud Pricing (RTX A6000 48GB $0.59/hr, updated Aug 27 2026)
[2] https://vast.ai/pricing — Vast.ai Live GPU Pricing (RTX A6000 48GB from $0.28/hr median $0.40)
[3] https://lambdalabs.com/service/gpu-cloud/pricing — Lambda GPU Cloud VM Pricing (RTX A6000 48GB $0.80/hr)
[4] https://www.e2enetworks.com/pricing — E2E Networks India GPU Cloud Pricing (L40S 48GB $1.20/hr, A40 48GB $1.44/hr)
[5] https://www.xe.com/currencyconverter/convert/?Amount=1&From=USD&To=INR — XE USD/INR mid-market rate (1 USD = 96.7384 INR, 11 Oct 2026)
[6] https://www.razorpay.com/pricing — Razorpay India Pricing (2% platform fee + 18% GST)
[7] https://clutch.co/developers/pricing — Clutch Software Development Pricing Guide (India $25-49/hr)
[8] https://eucalipse.com/articles/software-development-costs-by-country-2025 — Eucalipse Software Development Cost by Country 2025 (India $18-40/hr avg $29)
[9] https://www.heulistic.com/blog/llm-fine-tuning-cost-breakdown-model-size — Heulistic LLM Fine-Tuning Cost Breakdown by Model Size (2026)
[10] https://neon.com/pricing — Neon Postgres + pgvector Pricing
[11] https://upstash.com/pricing/redis — Upstash Redis Pricing
[12] https://www.digitalocean.com/pricing/managed-databases — DigitalOcean Managed PostgreSQL Pricing
[13] https://livekit.io/pricing — LiveKit Cloud Pricing (SFU/TURN self-host or cloud)
[14] https://developers.cloudflare.com/r2/pricing — Cloudflare R2 Object Storage Pricing (zero egress)
[15] https://startupadvisory.in/blog-gst-rcm-import-of-services.htm — GST on Import of Services under RCM (18% IGST)