Generative AI and LLM solutions
Tankar builds applications on large language models, from a chatbot that answers from your documents to agents that complete tasks in your systems. Every build includes retrieval over your data where accuracy matters, an evaluation set that scores answers before and after each change, and controls for cost, latency, prompt injection and data leakage.
Grounded, measured, controlled
An assistant that answers from your documents with citations, scored against an evaluation set on every change, with rate limits, redacted logs and approval steps where the output is acted on. That is what a production LLM application needs beyond a prompt.
What a build includes
Ingestion of your documents, retrieval with citations, prompts and tool calls, the interface or API, an evaluation set that your QA engineer owns, and monitoring of quality, cost and latency in production.
Typical engagements
What this service usually produces, as deliverables rather than adjectives.
- Retrieval-augmented generation (RAG) assistant that answers from policies, manuals or contracts with citations
- Document intelligence pipeline that extracts fields from invoices, tenders or applications into structured data
- Support or sales chatbot connected to your knowledge base and CRM, with hand-off to a person
- Copilot inside an existing product that drafts, summarises or classifies
- AI agent that completes multi-step tasks through your APIs with approval steps
- Integration of LLM features into an existing product with evaluation and monitoring
How it runs
The stages this service goes through, and what you see at each one.
Discovery
The questions users ask, the documents that hold the answers, the systems an agent would act on, and the accuracy and cost targets. Written estimate within 48 hours.
The written estimate follows within 48 hours of the scoped call.
Design
An evaluation set of real questions with expected answers, the retrieval and prompting approach, and the guardrails for safety and data handling.
Build
Ingestion, retrieval, prompts, tool use and the interface, scored against the evaluation set at every change. Weekly demo of the scores and the product.
Test
Accuracy, groundedness, cost per query and latency measured. Red-team tests for prompt injection, data leakage and unsafe output.
Launch
Rate limits, logging with redaction, monitoring of cost and quality, and a feedback loop from users into the evaluation set.
Run
Model and prompt updates scored before release, cost reviews, and a monthly report on usage, quality and spend.
- Team shape
- A pod of a project manager, an AI engineer, a backend engineer, a designer for the interface and a QA engineer who owns the evaluation set.
Use cases by industry
Where this kind of work has paid for itself, by sector. Each is a task with a measurable output, not a demo.
Government and public sector
- Extracting eligibility, deadlines and requirements from tender documents
- Classifying and summarising citizen applications and grievances
- An assistant that answers policy questions with citations to the source circular
Fintech and BFSI
- Document extraction for onboarding and KYC with human review
- Support assistant grounded in product terms and policies
- Summaries of long agreements and disclosures for internal review
Healthcare
- Intake and enquiry triage with hand-off to staff
- Drafting patient-facing content from approved clinical sources
Retail and e-commerce
- Product content generation from supplier data with review
- Support assistant connected to orders and returns
- Vendor onboarding help grounded in marketplace policies
Travel and hospitality
- Quote drafting from supplier inventory for agents
- Itinerary and policy questions answered from booking data
Manufacturing and chemicals
- Answering questions from safety data sheets and specifications
- Extracting fields from certificates of analysis and purchase orders
How we evaluate
- Answer accuracy and groundedness scored against an evaluation set of real questions with expected answers, before and after every change
- Cost per query in production, including retrieval, model calls and any review step
- Latency at the 95th percentile, including time to first token for streaming interfaces
- Safety tests for prompt injection, data leakage across users or tenants, and unsafe or off-policy output
What we do not recommend
We do not recommend an LLM for a task a rule or a small classifier solves reliably, fine-tuning as a first step, an assistant without a citation or review path where a wrong answer has legal or financial consequences, or an autonomous agent acting on production systems without approval steps. We also do not recommend building on a model whose data terms you have not read.
Stack for this service
The technologies this work is usually built on, and why each one is used here.
| Technology | Why we use it here |
|---|---|
| LLM APIs (OpenAI, Anthropic, Google, Azure OpenAI) | Hosted models through APIs, chosen per task on accuracy, cost and the data-residency options each provider offers. |
| Retrieval-augmented generation (RAG) | Answers grounded in your documents with citations, which is what makes an assistant trustworthy in a business setting. |
| Vector databases (pgvector, Qdrant, Pinecone) | pgvector inside PostgreSQL when the product already runs on it; a dedicated store when the corpus or query volume needs one. |
| AI agents and tool use | Structured tool calls into your APIs with approval steps, so an agent can act without being trusted blindly. |
| Evaluation harnesses | A fixed set of real questions scored on every change, so a prompt or model update cannot quietly make the product worse. |
| Python and FastAPI | Ingestion pipelines and model services in the language the AI ecosystem is built in. |
Selected work
Case studies where this service carried the engagement. Only outcomes with a source are shown.
Own productGovernment and public sectorTenderBazaar: tender discovery and alerts for Indian SMEs
A tender discovery platform that collects listings from central, state and municipal portals into one searchable feed with daily alerts.Services: SaaS product development, Mobile app development and 1 moreRead the case study
Anonymised clientRetail and e-commerceA multi-vendor marketplace with vendor self-service and compliant invoicing
A marketplace where vendors onboard themselves, list and price their own stock, and are paid out against a commission engine with GST-compliant invoices.Services: E-commerce and marketplace development, Custom software development and 1 moreRead the case study
Engagement models that fit
A proof of value with an evaluation set is usually fixed price; production assistants and agents run on time and materials or a dedicated team.
- Fixed priceA defined scope with a clear end state: an MVP, a website, a well-specified module or an integration.
- Time and materialsEvolving products, research-heavy work such as AI features, and engagements where the backlog is set sprint by sprint.
- Dedicated teamOngoing product development where you want engineers who know the codebase and stay on it, working as an extension of your own team.
Questions buyers ask
What buyers ask most about this service, answered before the first call.
Tell us what you are building.
NDA on request. Written estimate within 48 hours of a scoped call. Reply within one business day.