AI agents
Multi-step workflows that take real actions — raising tickets, updating records, escalating when confidence drops below a threshold you set.
AI & Intelligent Systems
Most AI development stalls at the demo. Ours reaches production.
A convincing prototype takes a weekend. A system your team can trust with real customers takes grounding in your own data, permission-aware retrieval, evaluation before launch and monitoring after it. That is the part we build — and the part most proposals leave out.
RAG & knowledge assistants
Answers grounded in your documents, with citations back to source and retrieval that respects each user's existing permissions.
Copilots inside your product
Assistance where users already work — the CRM, the dashboard, the admin panel — not bolted on as a separate chat box nobody opens twice.
Document, vision & voice
Invoices, scans, forms, calls and images turned into structured records your existing systems can process without a human retyping them.
What AI development means here
AI development is engineering software where a language or vision model is one component inside a system — not the whole product. Canopus builds retrieval pipelines, tool-using agents, document processing and in-product copilots, grounded in your own data and wrapped in the evaluation, access control and monitoring that production actually requires.
What we build
Four shapes, and what each one actually takes to build
Named honestly, including the parts that are harder than they look.
Knowledge assistants that cite their sources
Your policies, contracts, manuals, tickets and wiki, made answerable in plain language — with a citation the user can click to check. The retrieval layer respects the permissions a user already has, so a support agent and a director asking the same question get different answers from the same index.
The hard part is not the model. It's chunking documents so a clause doesn't get split from the condition that qualifies it, and keeping the index fresh when someone edits a policy on a Friday afternoon.
Typical: 6–9 weeks · hybrid vector + keyword search over 10k–500k documents · a citation on every answer
Agents that do the work, not just describe it
A tool-using agent that reads a request, checks the systems of record, takes the action and logs what it did — refunds inside a policy limit, ticket triage and routing, order status chased across three carriers, onboarding checks against a document pack.
The engineering that matters is the boundary: what it's allowed to do without a human, what requires approval, what happens when a downstream API times out halfway through a multi-step action, and how you undo a wrong decision. Autonomy without a rollback path is not automation, it's exposure.
Typical: 10–14 weeks · human-in-the-loop review for the first release · every action logged and reversible
Document intelligence at volume
Invoices, delivery notes, purchase orders, ID documents, claim forms, lab reports — read, validated against your rules, and written into the system as structured records. Confidence below the threshold routes to a human queue instead of writing a wrong number into your ledger.
We measure against a labelled sample of your real documents before quoting the accuracy, because a vendor's benchmark on clean PDFs tells you nothing about a phone photo of a crumpled delivery note.
Typical: 8–12 weeks · accuracy measured on 200–500 of your own labelled files · human review queue for exceptions
Copilots inside software you already run
The highest-adoption AI work we do, and the least glamorous: search that understands intent, a draft reply pre-filled in the reply box, a summary at the top of a long case, a field that suggests the right code. No new interface for anyone to learn, no separate tab, no change-management programme.
If you already have a product, this is usually the right first project — it lands inside an existing habit rather than asking users to build a new one.
Typical: 6–10 weeks · integrated into your existing application · no new interface for users to learn
Where we tell clients not to use AI
If the rule is deterministic, write the rule. A model that gets a tax calculation right 97% of the time is worse than a function that gets it right every time, and it costs more to run. We have replaced proposed AI features with a lookup table more than once, and the client kept the budget.
If nobody can say what a correct answer looks like, the project isn't ready. Evaluation needs a set of questions with agreed answers. If your own experts disagree on those, the disagreement is the project — not the model.
And we don't promise accuracy percentages before measuring on your data. Any number quoted in a first meeting is a guess dressed as a commitment.
How we de-risk it
Two weeks to find out if it works at all
AI is the one service line where we insist on a small paid evaluation before a build quote. Roughly one project in five stops at this stage — which is the point.
WEEK 1
Build the test set
50–200 real questions or documents with answers your experts agree on. This is the deliverable that outlives the project.
WEEK 2
Thin prototype
Retrieval and prompting against real data — ugly on purpose. We're measuring accuracy, cost per call and latency, not design.
DECISION
Numbers, then a choice
Measured accuracy, monthly inference cost, response time and the gap to what you need. Go, adjust scope, or stop.
THEN
Build with the harness
The test set runs in CI from then on, so a prompt change that breaks accuracy fails the pipeline instead of a customer.
What the evaluation covers, and what you keep
Two weeks, quoted as a fixed engagement once we've seen the data and the question you want answered. You keep the test set, the measurements and a written recommendation — including the case for not proceeding, if that's what the numbers say. It's the cheapest possible way to find out that a quarter of engineering time won't reach the accuracy your process requires.
Stack
What production AI is actually made of
The model is one line in a system that has a dozen others. These are the others.
Retrieval & grounding
pgvectorPineconeQdrant
Hybrid vector + BM25Re-ranking
Semantic chunkingPermission-filtered retrieval
Models & orchestration
ClaudeGPTGemini
Llama & Mistral (self-hosted)Azure OpenAI
Amazon BedrockTool calling & MCPStructured output schemas
Evaluation & safety
Golden-set regression tests in CILLM-as-judge scoring
Guardrails & input filteringPrompt-injection defences
PII redaction before inferenceHuman-in-the-loop review
Operations
Prompt & model versioningPer-feature cost tracking
Latency budgets & cachingFallback model routing
Feedback capture with contextAudit logging
Questions
Cost, accuracy, data and models
How much does an AI project cost?
Nothing here is priced off a feature list. What moves it is how much content or how many systems retrieval has to reach, whether the system only answers or also takes actions you must be able to review and undo, and the accuracy bar your process needs before anyone trusts it. We settle all three in the two-week evaluation, and you leave with a fixed written estimate and a test set you own even if you build elsewhere. Budget separately for inference, which is not a Canopus fee at all — it's your own account with Anthropic, OpenAI or Azure, and most internal assistants land between $200 and $1,500 a month in model calls. We measure that during the evaluation rather than guessing.
Which models do you use?
Whichever passes your evaluation set at acceptable cost and latency — and we design so the model is swappable. In practice: a frontier model such as Claude or GPT for reasoning-heavy steps, a smaller or open-weight model for classification and extraction, and a local embedding model where documents can't leave your network. Model choice is a runtime configuration, not an architectural commitment.
How do you stop it making things up?
Three layers. Retrieval grounds every answer in your own content and returns citations the user can open. The prompt and output schema constrain the response shape and force an explicit "not found" path. An evaluation set of real questions with known answers runs on every change, so a prompt edit that quietly breaks accuracy fails in CI rather than in front of a customer.
Will our data be used to train someone else's model?
Not under the enterprise API terms we deploy on — Anthropic, OpenAI and Azure OpenAI all contractually exclude API traffic from training by default. Where data can't leave your infrastructure at all, we deploy open-weight models such as Llama or Mistral on your own GPU instances. The chosen route goes in writing before the build starts.
How long does an AI build take?
Six to fourteen weeks, after a two-week evaluation phase that establishes whether the accuracy you need is reachable at all. Roughly one in five projects stops there — a good outcome, because it costs a fortnight instead of a quarter.
What does it take to run in production?
Cost and latency monitoring per feature, prompt and model versioning so you can roll back a regression, a feedback channel that captures bad answers along with their retrieved context, and a human escalation path for low-confidence cases. Without those, quality drifts silently as your content and the underlying model both change.
Can you add AI to software we already have?
Yes, and it's the most common request. We integrate into your existing application rather than shipping a separate chat window — search that understands intent, drafting inside the form the user is already filling, classification on a queue an agent already works. The integration surface matters more than the model. See API and system integration for how that connects.
Do you handle regulated or sensitive data?
We build with PII redaction before inference, per-user permission filtering on retrieval, regional hosting and full audit logging. For healthcare specifically, we work to HIPAA-aware delivery practices — but we're engineers, not your compliance officer, and we're explicit about that line.
Related services
Get a Free Quote
Bring us a question your team answers a hundred times a week.
That's usually the best first AI project. Tell us what it is and where the answer lives, and an engineer replies within one business day — including if we think a model is the wrong tool.
- Email — [email protected]
- Phone / WhatsApp — +91 817 979 7732
- Two-week evaluation — fixed scope, fixed fee, and you keep everything it produces