AI agents that still work when the demo is over.
Chat over your documents, an agent that does the boring back-office work, a voice line that answers the phone. The demo takes a week. The version that handles a 40-page PDF, a 60-second model call and a user who types in Hindi takes engineering. That is what we build.
What we build
- 1Chat and search over your own data: contracts, tickets, product docs, with citations and a real ingestion pipeline instead of a synchronous upload.
- 2Agents that act: triage inbound email, update the CRM, draft the reply, escalate the edge cases to a human with context.
- 3WhatsApp, Instagram and web chatbots for support and lead capture, with handoff to your team.
- 4Voice agents for inbound calls: qualification, booking, post-call summaries into your CRM.
- 5Internal tools: parsers, scrapers and classifiers that turn messy inputs into structured rows.
How an AI build runs
The order matters: the boring infrastructure decides whether the model looks smart.
1
Define the job and the failure cost
What the agent must do, what it must never do, and what a wrong answer costs. That sets the guardrails, the human-in-the-loop points and the eval set before any prompt is written.
2
Ingestion as a job, not a request
Uploads go to storage; a worker parses, chunks, embeds and indexes. The chat route only retrieves and generates. This is the single biggest reason production AI features time out.
3
Build with evals from day one
A set of real questions with expected answers runs on every change. Prompt edits that regress get caught before users do.
4
Budgets and limits
Caps on file size, tokens, tool calls and per-user spend. Rate limits on the "Generate" button. Cost tracked per feature so AI is part of your COGS, not a surprise.
5
Ship, watch, tune
Traces on every call, a review queue for low-confidence answers, and a weekly look at what users actually asked. Then we tune prompts, retrieval and models against the eval set.
What every AI build includes
- Async ingestion pipeline with a queue (Inngest or Postgres-backed)
- Retrieval with citations; users can see where an answer came from
- An eval set with pass/fail thresholds wired into CI
- Per-user and per-feature cost tracking and hard caps
- Guardrails: allowed actions, refusal rules, human escalation path
- Model choice by task (OpenAI, Anthropic, open-weight) with a swap path
- Tracing on every call, and a queue for reviewing low-confidence outputs
- Handover with the eval set, so you can change prompts safely without us
What it costs
Fixed-price builds. Running costs (model usage) are yours and we design to keep them predictable.
| Scope | Typical price | Time |
|---|---|---|
| Focused assistantChat over your docs, or one agent with a small set of actions, with evals and limits. | $6,000–$15,000 | 3–5 weeks |
| Agent workflowMulti-step agent with tools, CRM/email integration, review queue, voice or WhatsApp channel. | $15,000–$40,000 | 5–10 weeks |
| AI inside your productAI features shipped inside a customer-facing SaaS, with billing, usage metering and multi-tenant data isolation. | From $40,000 | 8–16 weeks |
USD. Typical running costs for a support assistant handling 1,000 conversations a month are in the low hundreds of dollars; we show you the number before launch.
Questions we get about this
Why does our AI feature time out in production?
Usually because one request does ingestion and chat together: download, parse, chunk, embed, then call the model. That is 30 to 90 seconds. Splitting ingestion into a background job and capping inputs brings chat under ten seconds for most apps.
Which models do you use?
Whichever fits the task and the cost. OpenAI and Anthropic for most reasoning and generation; smaller or open-weight models for classification and extraction at volume. Every build has a swap path so a pricing change does not mean a rewrite.
How do you stop the agent doing something harmful?
Explicit allowed actions, refusal rules in the system prompt backed by code checks, human approval on anything irreversible or expensive, and an eval set that includes the adversarial cases. The model never gets more permissions than the job needs.
Can you add AI to our existing app?
Yes, that is most of the work. It starts with a short assessment of your data, auth and infrastructure, because the AI feature inherits whatever is fragile underneath.
Do you build WhatsApp and voice bots?
Yes, on the Meta and telephony APIs, with proper opt-in, handoff to humans, and conversation logs in your CRM. See the WhatsApp and voice pages for specifics.
Read how we fix it
Vercel Function Timeouts for AI in 2026
Shipping AI on Next.js + Vercel? Why routes hit 10s/60s limits in prod, plus fixes that actually work in 2026.
Server Actions: Stop Double Charges (2026)
Fix double-submit Server Actions with idempotency keys, Redis rate limits, and retry-safe patterns (Next.js in 2026).
Fix Inngest 401s on Vercel (2026)
Inngest hitting 401/403 on Vercel? Deployment Protection is usually the cause. Here’s the bypass fix, plus safer options.
R2 Presigned PUT 403: Hidden Header Traps (2026)
Fix Cloudflare R2 presigned PUT 403s caused by header mismatches (Content-Type, checksums, CORS). A dev checklist you can use today.
Tell us the job the agent should do.
One of the founders reads it, not a sales team. Within 48 hours you have a written assessment and a fixed price. If it is a two-hour fix, we tell you that too.
Prefer to talk first? Book a 30-minute call with Mahima. No pitch deck.