This book teaches you to build and adapt language models for Indian languages. This appendix answers the next question: who pays for that, and when is a small fine-tuned model the right answer for them?
The ten cases at a glance¶
| # | Business | Languages | Task | Requests / month | API / month | Self-host / month | Verdict (example prices) |
|---|---|---|---|---|---|---|---|
| 1 | Health insurer | Hindi, Hinglish | policy and claims support, escalation | 300,000 | $1,155 | $734 | self-host saves ~$420 |
| 2 | Gulf bank / remittance house | Roman Urdu, Urdu | intent JSON, fraud first | 500,000 | $1,148 | $734 | self-host saves ~$410 |
| 3 | E-commerce | Tamil, Telugu, Kannada | ticket triage | 1,500,000 | $4,238 | $734 | self-host saves ~$3,500 |
| 4 | Call centre | Hinglish | score every call transcript overnight | 200,000 | $1,600 | $734 | self-host saves ~$870 |
| 5 | Agri-input company | Hindi, Marathi | farmer advisory on WhatsApp | 400,000 | $2,008 | $734 | self-host saves ~$1,270 |
| 6 | Government helpline | Hindi + state languages | scheme eligibility, hand-offs | 1,000,000 | $5,206 | $923 (2 boxes) | self-host saves ~$4,280 |
| 7 | Ed-tech | Telugu, English | homework doubt solving | 800,000 | $6,120 | $923 (2 boxes) | self-host saves ~$5,200 |
| 8 | UK hospital | Urdu, Punjabi, Tamil | discharge instructions, nurse-reviewed | 6,000 | $118 | $734 | API cheaper; self-host only if records can’t leave |
| 9 | Law firm | Hindi | court-order summaries | 3,000 | $128 | $859 | API cheaper; self-host only if documents can’t leave |
| 10 | Devotional app | Sanskrit, Hindi, English | Gita verse citations | 20,000 | $56 | $734 | API cheaper at this volume |
The pattern behind the table¶
A small fine-tuned model tends to win when most of these are true:
The task is narrow and repeated. The same few kinds of question, thousands of times a day.
The output has a fixed shape. A JSON intent, a priority, a two-line answer, a tag like
[ESCALATE].Volume is high, typically above a hundred thousand requests a month at example prices.
The language is expensive in tokens on English-centric tokenizers, as Tamil, Telugu, Kannada and Devanagari often are (Step 3).
The data is sensitive: health, money, citizens’ records, call recordings.
A big API tends to win when the work is rare, long and open-ended: a few thousand long documents a month, each needing broad knowledge and careful reasoning. Say so when it is true. A consultant who admits where the API wins is believed about the rest.
How an engagement usually runs¶
The same loop as Season 2, sold in stages:
| Stage | What you deliver | Season 2 tools |
|---|---|---|
| 1. Assessment (1–2 weeks) | a 100–300 item test set from the client’s real data; scores for a frontier API and two or three open models; a cost estimate at their volume | csv_to_seed.py, evaluate.py, report.py, cost_calculator.py, fertility.py |
| 2. Pilot (2–6 weeks) | a LoRA adapter trained on their data, with a before/after table on the same test set, per language | expand_with_teacher.py, train_lora.py, run_episode.sh |
| 3. Deployment | the model served on their hardware (a GPU server, a laptop or a small device), throughput measured, a gateway in front, and a plan to keep the test set growing | Season 2 Episodes 5–7: merge_adapter.py, export_gguf.sh, serve_vllm.sh, load_test.py, gateway.py, fit_calculator.py, kiosk.py |
| 4. Knowledge base (optional) | search and cited answers over the client’s own documents, in every language their customers use | Season 2 Episode 8: build_index.py, eval_retrieval.py, rag.py |
Stage 1 is the easiest to sell, because the client learns something useful even if the answer is “keep using the API”.