01 AI and LLMs

LLM development company for AI features that work from your own data

Tailwebs is an LLM development company that builds large language model features into products: chatbots that answer from your own content, AI extraction from documents and data, and assistants inside apps. Our shipped AI work includes the NITI Aayog Frontier Tech Repository portal with its NFTR Assistant, AI extraction behind Vidhi's Crime and Punishment database, and our own TailGPT.

Talk to us about your AI featureBook a call
AI work for

02 Shipped

Generative AI development we've built

Four pieces of AI work: two for clients and two of our own products.

  1. The NFTR Assistant chat panel open on the NITI Frontier Tech Repository home page
    01

    NFTR Assistant

    Built into the NITI Aayog Frontier Tech Repository portal, which we built: a chatbot that answers from the knowledge base, cites its sources and reports unanswered questions to admins. Shows: answers grounded in sources, citations, and a log of the gaps.

  2. The Crime and Punishment database home page
    02

    Crime and Punishment

    For Vidhi Centre for Legal Policy: a scraping engine plus AI extracts punishment data and tags law sections, for a database of 7,305 crimes across 370 central laws that Tailwebs built and maintains. Shows: extraction at scale. Live at crimeandpunishment.in.

  3. The TailGPT website, tailgpt.tailwebs.com
    03

    TailGPT, our own product

    AI agents for your website, built from your own content. Shows: running an AI product ourselves. See TailGPT.

  4. 04

    Tailwebs Operator, our own product

    An engine that automates business processes through AI assistants: it interviews you, then builds them. Each assistant sorts actions into auto, draft or ask. Shows: AI that does the work. See Tailwebs Operator.

We also run Kaipo, our own conversational AI app, and 200+ AI agents work in our own delivery process alongside 50+ people.

See what clients say in our Clutch reviews.

The same methods cover search across reports, summaries of long documents and assistants inside existing apps. To scope one as a project, see AI development, AI chatbot development or AI automation.

03 In production

What does an AI product need once it is live?

The right model for each job, a memory that finds what matters, and guardrails on what goes in and out. These come from AI products we run ourselves.

  1. 01

    Multi-model routing

    On a consumer AI app we run, different models are routed to different jobs, with fallbacks.

  2. 02

    Long-term memory

    The same app keeps long-term memory of what a user has shared.

  3. 03

    Guardrails

    On TailGPT, answers come only from the client's content.

04 Approach

LLM integration: the cheapest approach that works, first

For answers or extraction from your own documents there are three approaches, from the simplest to the most involved. Each step up adds build and running cost.

  1. A well-built prompt

    Fits when
    The content is small and fixed, such as one policy, a product manual or a page of FAQs, so it can travel with every question.
    Move up when
    The content outgrows the prompt, or changes often enough that keeping it there gets slow and costly.
    What it adds
    The least to build. Running cost grows with the length of the content sent with each question.

  2. Retrieval over your documents

    Fits when
    The content is large or changes often: a library of reports, a knowledge base, years of records. Only the passages that match each question are fetched and sent with it.
    Move up when
    Answers to your own questions still fall short, even when the right passages are found.
    What it adds
    An index of your documents to build and keep fresh, and a search step on every question.

  3. Fine-tuning

    Fits when
    Only when testing on your own questions shows the first two rungs fall short, for example for a fixed output format or a specialist vocabulary.
    Move up when
    This is the top rung. If it still falls short, the scope is revisited with you before more is spent.
    What it adds
    Training data to prepare, a training run, and a retrain when the content or the base model changes.

The same rule picks the model

The smallest model that answers your questions well, with a larger model only for the questions that need it.

We start on the cheapest rung that can do the job and move up only when it falls short. The NFTR Assistant, which answers from a knowledge base and cites its sources, is on the retrieval rung. An LLM feature sits inside the product it serves, whether that backend is built in Python, Ruby on Rails or PHP.

05 Evaluation

Tested on your own questions, before it goes live

Quality is checked on your own questions, not promised. Our own products show how: TailGPT checks each cited fact against its source page and says when the site doesn't cover something, and every assistant Tailwebs Operator builds passes red-team probes before it goes live.

  1. Start from your questions

    The feature is tested on real questions from your organisation, chosen with your team.

    For example The questions your users already ask, each with the answer your team would give.

  2. Check answers against sources

    Each answer is checked against the source it cites, so a wrong answer shows up before your users see it.

    For example TailGPT checks each cited fact against its source page, and says when the site doesn't cover a question.

  3. Probe it before launch

    We test before release, the way we do on our own products.

    For example Every assistant Tailwebs Operator builds passes red-team probes before it goes live.

  4. Learn from live questions

    After launch, the questions it couldn't answer show your team where the content has gaps.

    For example The NFTR Assistant reports unanswered questions to admins, and TailGPT shows a 7-day view of what the site couldn't answer.

Live questions show the gaps

06 Guardrails

Guardrails: what stops wrong or unsafe answers

Answers drawn only from your sources, an honest "I don't know", and a person in the loop where a mistake would cost you.

  1. 01

    Grounding with citations

    Answers come from your sources and show them, so a reader can check where each one came from, as on the NFTR Assistant.

  2. 02

    Saying "I don't know"

    When your sources don't cover a question, the assistant says so and logs it for your team. The NFTR Assistant reports unanswered questions to admins, and TailGPT says when the site doesn't cover something.

  3. 03

    A person in the loop where it matters

    Where a mistake would cost money or reputation, the output can wait for a person to approve it before anyone relies on it.

  4. 04

    Defence against prompt injection

    TailGPT, our own product, refuses prompt-injection and off-topic requests.

  5. 05

    Personal data kept in check

    Personal data stays out of prompts and logs unless the feature needs it. TailGPT purges old chat transcripts on a retention schedule.

  6. 06

    You choose what runs on its own

    Every assistant Tailwebs Operator builds sorts actions into auto, draft or ask, and passes red-team probes before going live. On an export outreach engine we are building for a client, a person approves every message before it is sent.

For legal, policy or research documents, see how we build for policy and research organisations.

07 Running cost · draft

Cost per conversation, worked out before launch

The monthly bill is worked out before launch: the tokens in a conversation, times the model's price, times your expected volume, plus retrieval and hosting.

What the bill is built fromDraft
  1. Tokens per conversationThe question, the passages sent with it, and the answer.
  2. The model's pricePer token, from the provider's own price list at the time.
  3. Expected volumeConversations a month, at launch and a year on.
  4. Retrieval and hostingThe search index and the servers that run the feature.
What it comes to
  • Low use, under 5,000 conversations a month: about ₹3,000–15,000 a month in AI usage.
  • Medium use: about ₹15,000–60,000 a month.
  • High use: ₹60,000+ a month.

Levers that lower it

Simple questions routed to a smaller model, repeated answers cached, and context kept short.

AI usage is billed from your own provider account where possible. The full guide is what an AI app costs to build and run.

08 Your data

Where your data lives

Five points decide where your data goes: the region, the accounts, training, the server and the law.

Working with policy or public-sector data? These points are the ones to settle first, before any document is shared.

Where it is stored and processed
In the region your rules require, agreed with you before any document is shared.
Accounts
In accounts you own where possible.
Training on your data
Whether the model provider may train on your data is checked against its terms for each project.
Running on your own server
Where your rules need it, we can run the system on your own server or cloud account, agreed before the build.
Data protection law
We work under India's DPDP Act 2023, and GDPR where it applies.
Hiring instead?

Need an engineer for LLM work in your own team? Hire a dedicated developer

09 FAQ

Questions about LLM development.

What does an LLM feature cost to run each month?
In our draft bands, AI usage costs about ₹3,000–15,000 a month for low usage (under 5,000 conversations) and ₹15,000–60,000 for medium usage, depending on the model, context size and caching.
How much does it cost to build an LLM feature?
In our draft bands, adding an AI feature to an existing app costs ₹2–8 lakh, and an AI chatbot on your own data ₹3–10 lakh to build.
What data do I need before building an AI feature?
The documents or records the feature should work from, access to them, and real example questions with good answers, so we can check the feature against them.
How do you test AI features before release?
Before launch we test the feature on real questions from your data. TailGPT, for example, checks each cited fact against its source page, and every Tailwebs Operator assistant must pass red-team probes before it goes live.
How do you stop the AI making things up?
It answers only from your sources, shows its citations and says when it doesn't know, logging the question for your team, as the NFTR Assistant does.
Which AI models do you work with?
We pick the model that fits each job, weighing cost, accuracy and where your data is allowed to go. On a consumer AI app we run, different jobs go to different models, with fallbacks.