AI product development

AI product development for startups.

We build AI features that work past the demo: LLM integration, search over your own data, agents, evals and guardrails. Use an API model, or an open model hosted in India.

IntegrateEvaluate
AI product

What we build

LLM featuresChat, summaries, extraction and drafting inside your product.
Search over your dataRetrieval that respects who is allowed to see what.
AgentsNarrow assistants that call your tools, with limits.
EvalsA test set that decides if a change ships.
GuardrailsInput and output checks, and human approval where it matters.
Cost and logsTraces, spend per feature and a record of what happened.
Six partsTested before launch
Run safely

From pilot to production.

Most AI pilots stall on the same six things. We deal with each one before launch, not after the first bad answer. If you are adding AI to a product that already exists, the same list applies, and our guide to adding AI to an existing SaaS product covers the architecture and rollout.

Evals before launch

Where pilots failA demo that works on five examples and fails on the sixth.

What we doA test set built from real inputs, with expected answers, scored on every change to the prompt, model or search. A change that makes results worse does not ship.

Retrieval you can inspect

Where pilots failAnswers that sound right but come from the wrong document, or from none.

What we doSearch that returns its sources, tested on its own before the model is involved, so you can tell a search failure from a model failure.

Observability

Where pilots failNo way to see why a user got a bad answer last Tuesday.

What we doTraces for each request, the prompt and context used, latency, cost and a way to flag bad outputs. Our guide covers an LLM observability setup for small teams. Read the observability guide.

Cost control

Where pilots failA bill that grows faster than usage.

What we doCost per feature and per customer, caching where it is safe, smaller models for simple steps and limits on runaway loops.

Fallbacks

Where pilots failOne provider has an outage or a rate limit and your feature goes down.

What we doA second model or provider behind the same interface, with rules for when to switch. We lay out multi-model design in a separate guide. Read the multi-model guide.

Personal data in prompts

Where pilots failNames, phone numbers and documents sent to a model provider by default.

What we doMasking before the prompt where you can, a clear list of what each provider sees, and retention settings checked and recorded.

Search over your data, with permissions.

Retrieval is where private data most often leaks into AI answers. Our guide to permission-aware RAG goes through the design in detail. These are the points we check first.

Filter before the model sees anything

Access rules must apply to the search, not to the answer. If a document reaches the model, assume it can leak into the reply. Each chunk carries the permissions of its source document.

Check filters work with your vector index

In pgvector, with approximate indexes, filtering is applied after the index is scanned, so a strict filter can return fewer rows than you asked for. Iterative scans (from version 0.8.0), partial indexes or partitioning address this. Test it with your real filters.

Isolate tenants the efficient way

Qdrant advises against a separate collection for every tenant in most cases. It recommends one collection with a tenant payload index, and separate collections only for a limited number of tenants that need strict isolation.

Use database rules as a second lock

Postgres row level security restricts which rows each user can read. With it on and no policy, access is denied. Superusers and roles with BYPASSRLS skip it, and table owners skip it unless you force it.

Expect synced permissions to lag

Copies of permissions from SharePoint or a drive lag behind the source. Azure AI Search says permission refresh does not provide immediate revocation, and its ADLS ACLs refresh about once a day. If access must end at once, check the source system at query time.

Agents, with evals and guardrails.

An agent is a model that can take actions, so a wrong answer can become a wrong action. We build for the cases where it goes wrong, and we test those cases before launch.

One job, a short tool list

An agent that does one task with three tools is easier to test than a general assistant with thirty.

Permissions of the user, not the agent

The agent acts with the rights of the person who asked. It never holds a master key.

Approval for actions that matter

Sending money, deleting data and emailing customers wait for a person to approve.

Evals on whole runs

We test the full sequence of steps and tool calls on realistic tasks, not only single answers, including tasks designed to trick it.

Limits and a kill switch

Step counts, spend caps and timeouts, plus a way to turn the agent off without a deploy.

A log of every step

What it was asked, which tools it called, with what inputs, and what came back.

Models we work with, and their licences.

We work with API models from the main providers and with open-weight models such as Llama, Gemma, Qwen and Mistral. We pick by task, language, cost, licence and results on your own examples. A licence belongs to a model, not to a vendor, so we check each one. Our guide to open-weight model licences goes further. This is a summary, not legal advice: read the licence on each model card.

Llama 4 (Scout, Maverick)

What the licence saysMeta’s Llama 4 Community License. Companies with more than 700 million monthly active users need a licence from Meta. Distributors must display “Built with Llama”, and derivative model names must start with “Llama”. The Acceptable Use Policy applies, and for multimodal models it withholds the licence grant from EU-based individuals and companies. End users of products are exempt.

What to rememberOpen-weight under Meta’s community licence, not open source. Check the multimodal EU terms before you ship to Europe.

Gemma

What the licence saysGemma 4 is under Apache 2.0. Earlier Gemma models are under Google’s Terms of Use, which pass restrictions downstream and let Google restrict use remotely or otherwise.

What to rememberTreat Gemma 4 and earlier Gemma as two different licences.

Qwen

What the licence saysMixed. Some models are Apache 2.0. Others use the Qwen Community License, which asks larger products to show the model name and needs a separate licence for a model-as-a-service or AI work assistant business. One image model is non-commercial research only.

What to rememberNever assume “Qwen is Apache 2.0”. Read each model’s licence file.

Mistral

What the licence saysMixed. Some models are Apache 2.0. One medium model uses a modified MIT licence that bars companies above a monthly revenue limit. The large model has its own open-weight licence that needs a commercial licence for model-as-a-service businesses. Some older models carry a research or non-production licence.

What to rememberNever write “Mistral is open source” as a blanket statement. Read each model’s licence file.

Running your own model, in practice.

For teams that choose an open model, this is what we plan for. Our private LLM deployment in India page covers the full build for regulated data.

Serving

vLLM runs an OpenAI-compatible server with one command, so your product code can switch between a hosted API and your own model. Memory is set with --gpu-memory-utilization, context with --max-model-len and multi-GPU with --tensor-parallel-size.

Sizing, as a rough estimate

Weights take about parameters times bytes per parameter: 2 for BF16, 1 for FP8 or INT8, about 0.5 for 4-bit. A 31 billion parameter model is about 62 GB in BF16 and about 16 GB at 4-bit, before the KV cache and overhead, which grow with context length and concurrent users. We test on your own prompts before choosing a GPU.

Quantisation

Quantising a model trades precision for a smaller memory footprint. vLLM supports several methods, and FP8 needs Ada, Hopper or AMD GPUs. We check quality on your eval set before and after.

Gateway and routing

A gateway such as LiteLLM, an open-source proxy (MIT licence, except its enterprise folder) with virtual keys, spend tracking and fallbacks, can sit in front of several models. A hosted router such as OpenRouter is another processor that sees your prompts. For sensitive personal data, use a self-hosted gateway in front of India-hosted models.

Hosting in India, the options.

Where the model runs affects latency, data location and cost. DPDP does not require data to stay in India. Hosting here avoids future Section 16 transfer restrictions and makes sector localisation rules easier to meet. Details checked 20 November 2026 and they change, so we recheck before you commit.

AWS

Mumbai (ap-south-1) and Hyderabad (ap-south-2, opt-in required).

Azure

Central India (Pune), India South Central (Hyderabad), South India (Chennai) and West India (Mumbai). Availability zones differ by region.

Google Cloud

Mumbai (asia-south1) and Delhi (asia-south2). GPU types differ by zone, so check before planning.

Indian GPU clouds

E2E Networks lists B200, H200, H100, A100 and RTX PRO 6000 GPUs and shows a MeitY empanelment label. Yotta Shakti Cloud lists H100 and L40S, with B200 announced.

IndiaAI compute

Subsidised rates for approved projects, through a request process that opens monthly on the IndiaAI portal. Eligibility follows the portal’s criteria. It is a monthly, per-request subsidy, so don’t plan production capacity on it.

How it runs.

Phases, not promises of dates. We agree the plan with you once we know your product, your data and your users.

  1. Step 1

    Scope

    We agree the one job the AI must do, who uses it, what data it touches and how you will judge a good answer.

  2. Step 2

    Prototype

    We build the thinnest working version and the first eval set from your real examples.

  3. Step 3

    Harden

    We add retrieval permissions, guardrails, fallbacks, logging and cost limits, and fix what the evals show.

  4. Step 4

    Launch

    We ship to a small group, watch the traces and widen access as the numbers hold.

  5. Step 5

    Hand over

    Documentation, the eval set and the dashboards, so your team can keep changing it safely.

AI product development FAQs.

Should we use an API model or host an open model?

Start with the question your data and buyers ask, not the model. A hosted API is the quickest way to learn whether the feature works. An open model on infrastructure you control gives you more say over where data goes and how it is retained. Many products use both, behind one interface, so you can change later.

Can you add AI to a product we have already built?

Yes, if the product has a codebase we can work in. We look at where AI would help a real user, what data it needs and how you will know it works, then add it behind a feature flag so you can switch it off.

How do you know an AI feature is ready to launch?

When it passes an eval set made from realistic inputs, including hard and adversarial ones, and a small group of real users has used it with traces switched on. We agree the pass mark with you before building, so nobody argues about it afterwards.

Do we have to host the model in India?

No. The DPDP Act does not require data to stay in India. Hosting in India avoids future Section 16 transfer restrictions and makes sector rules on localisation easier to meet. Your regulator and your buyers may ask for it, so we design so that the model location can change.

Does Mono have AI products in its portfolio?

Our work shows full product builds, such as Prosprr and Maple Drive, and design and frontend staffing for the MVP of Otto’s AI platform. It does not show AI feature builds, so we do not cite it as proof of them. Ask us what we would do for your use case and how we would test it.

Is an AI agent safe to let act on its own?

Only within limits. We give an agent a narrow job, the permissions of the user it acts for, spend and step limits, and a person’s approval for actions that move money, delete data or contact customers. We test full runs before launch and keep a log of every step.

Build an AI feature that holds up in production.