Skip to content
Nine Agents
India · Europe · Canada

We buildthe machinethat buildsthe thing.

Not a service desk. A studio that designs the system, ships it, and leaves you owning something that keeps working after we go.

See what we shipped→How we build
Systems shipped
25+
Agents in production
100+
Regions
IN / EU / CA
  • Agentic marketplaces
  • Voice agents
  • Kafka event backbones
  • Extraction pipelines
  • ClickHouse analytics
  • Campaign systems
  • Prometheus + Grafana
  • Job-matching graphs
  • n8n orchestration
  • Rubric feedback loops
  • Multi-source collection
  • Model routing
  • Redis queues
  • Motion systems

Positioning

An agency sells capacity.

We sell the thing that replaces it.

Built to be handed over — not to be depended on.

01scopes the deliverableScopes the problem, argues with the brief
02bills the hoursBuilds what removes the hours
03adds AI to the pitchSpends latency only where the model earns it
04hands over a repoHands over a system your team can extend

Fig. 01 — capability graph

Six systems. One set of parts.

Every project draws on the same substrate. That is the difference between a studio and an agency — the tenth build is cheaper than the first because the parts already exist.

    Hover or focus a node to trace its connections_

    Selected work — 06 systems

    What we shipped, and what it changed.

    Client organisations are described by sector rather than named. Every figure on this page is either something we built or a public statistic with its source attached.

    1. 01Motion-led site for a Belgian IT consultancyIT consulting
    2. 02Campaign operating system for a national NGOSocial impact
    3. 03BolayEdTech
    4. 04ClimbItUpCareers
    5. 05AskJohnnyVoice AI
    6. 06MarketMuseAgentic marketplace
    01
    Client
    A Brussels-based IT consultancy
    Sector
    IT consulting
    Region
    Belgium
    Built with
    Next.jsGSAPLenisMotion design

    Motion-led site for a Belgian IT consultancy

    A consultancy that sells judgement had a website that sold hours.

    The tension

    Technical consultancies all describe themselves with the same nouns — transformation, performance, analytics. When the words are interchangeable, the only differentiator left is how the thing feels to move through. Their existing site read as a brochure; the work behind it did not.

    What we built

    A fully animated marketing site where motion carries meaning rather than decorating it. Sections resolve as you arrive at them, the navigation reacts to the ground it sits on, and every transition is transform-and-opacity only so the whole thing holds a steady frame budget on a mid-range laptop.

    Where the leverage is

    We used AI in the production pipeline rather than in the product: model-assisted copy exploration to pressure-test positioning against a dozen competitor sites, and generative passes for art direction that a designer then cut down. The shipped site runs no inference — it is static, and that is the point.

    The path through

    01Competitor corpusa dozen peer sites, scraped
    02Positioning passesclaims pressure-tested
    03Art directiongenerated, then cut by hand
    04Motion systemfixed frame budget
    05Static exportno runtime inference
    Inference sits in production, not in the product. The shipped site is static HTML.

    What they own now

    • Static export, no server needed — deploys to any CDN edge
    • Motion system with a fixed frame budget rather than ad-hoc animation
    • Positioning rebuilt around what the firm decides, not what it delivers
    02
    Client
    An Indian NGO running health, education and welfare campaigns
    Sector
    Social impact
    Region
    Karnataka · Maharashtra · Gujarat
    Built with
    Next.jsPostgresLLM extractionPayments

    Campaign operating system for a national NGO

    The campaigns were real. The record of them lived in six spreadsheets.

    The tension

    A small team was running menstrual-health awareness drives, eye-donation advocacy and in-kind donation collections across three states — and coordinating all of it over WhatsApp threads and shared spreadsheets. Reach was being achieved and then lost, because nothing that happened in a village on a Tuesday made it into a form a corporate CSR partner could read.

    What we built

    We digitalised the organisation end to end: a campaign system of record covering the menstrual-health, eye-donation and in-kind donation programmes, with field intake, beneficiary tracking, volunteer coordination and donation flows in one place — plus the public-facing site that campaigns are actually launched from.

    Where the leverage is

    Field reports arrive as free text and photos in three languages. We put a language model in the intake path to normalise them into structured records — location, beneficiaries reached, materials distributed — so a volunteer types the way they speak and the database still gets clean rows. Report generation for CSR partners is templated off the same records.

    The path through

    01Field reportfree text, photos, 3 languages
    02Normalisationlocation, reach, materials
    03Campaign recordone row, one source of truth
    04Operationsvolunteers, donations, drives
    05CSR reportinggenerated, never re-keyed
    A volunteer types the way they speak; the database still receives clean rows.

    What they own now

    • One system of record across three state programmes
    • Multilingual free-text field intake normalised into structured data
    • Campaign creation, volunteer coordination and donations unified
    • CSR-ready reporting generated from operational data, not re-keyed
    03
    Client
    A leading Indian communication and public-speaking coach
    Sector
    EdTech
    Region
    India
    Built with
    Next.jsSpeech-to-textLLM rubricsProgress modelling

    Bolay

    You cannot learn to speak by watching someone else speak.

    The tension

    Communication coaching does not scale the way a course does. The value is in a coach hearing you and telling you what to fix — which means the ceiling on how many people one coach can reach is roughly the number of hours in their week. Video courses remove the constraint and the value at the same time.

    What we built

    A communication and business-skills learning platform built around practice rather than playback. Learners work through structured speaking exercises, record attempts, and progress through a curriculum that adapts to where they are actually weak instead of moving in a straight line.

    Where the leverage is

    Speech-to-text plus a rubric-driven language model gives learners feedback on pace, filler-word density, structure and clarity in the seconds after they finish speaking — the fast loop that used to require the coach in the room. The coach stays in the loop where judgement is genuinely needed, which is where their hours are now spent.

    The path through

    01Recorded attemptthe learner speaks
    02Transcriptionrealtime speech-to-text
    03Rubric scoringpace, fillers, structure
    04Feedbackseconds, not next session
    05Curriculum adaptsto the actual weakness
    The fast loop that used to need the coach in the room. The coach keeps the slow one.

    What they own now

    • Practice-first curriculum with recorded attempts as the primary unit
    • Automated rubric feedback on pace, fillers, structure and clarity
    • Coach time redirected from correction to judgement
    • Reach decoupled from the coach’s calendar
    04
    Client
    Commerce and CA students across India
    Sector
    Careers
    Region
    India
    Built with
    Scraping pipelineLLM normalisationLinkedIn graphNext.js

    ClimbItUp

    The jobs exist. The graduates exist. The routing is broken.

    The tension

    Commerce and CA aspirants are served last by every general job board. Listings relevant to articleship, audit, taxation and finance are scattered across a dozen portals and firm career pages, none of which talk to each other — so the search itself becomes a part-time job, and the students without a network lose by default.

    What we built

    A vertical job portal for commerce graduates that scrapes CA- and commerce-relevant roles from multiple platforms into one deduplicated, normalised feed — and then does the part students find hardest, which is turning a listing into a conversation with a person who can actually move it.

    Where the leverage is

    Scrapers pull from sources that all describe the same role differently; a language model normalises titles, seniority and specialisation into a single taxonomy so "Article Assistant" and "CA Industrial Trainee" land in the same bucket. On top of that, the platform surfaces relevant people at the hiring firm on LinkedIn and drafts the opening message — the introduction a well-networked candidate gets for free.

    The path through

    01Multi-source scrapeportals + firm career pages
    02Dedupesame role, four listings
    03Taxonomytitle, seniority, specialisation
    04Matched feedcommerce vertical only
    05Warm introthe right person, drafted
    The listing is the easy half. Turning it into a conversation is the half that decides outcomes.

    What they own now

    • Multi-source scraping into one deduplicated commerce-vertical feed
    • Role taxonomy normalised across inconsistent source listings
    • Warm-introduction routing to relevant people at the hiring firm
    • Network access as a product feature rather than an accident of birth
    05
    Client
    Operators running phone-heavy front desks
    Sector
    Voice AI
    Region
    India
    Built with
    Vapin8nRealtime STT/TTSWebhook orchestration

    AskJohnny

    Most of India still does business on a phone call.

    The tension

    For an enormous share of Indian businesses the first point of contact is not a form or a chat widget — it is somebody picking up a phone. That conversation is where bookings are made, questions are answered and customers are lost, and almost none of it is instrumented, staffed consistently, or available after 7pm.

    What we built

    A voice-agent platform that stands up production phone agents: they answer, understand what the caller wants, take the action — book, reschedule, look up, escalate — and hand off to a human cleanly when the conversation leaves their competence.

    Where the leverage is

    Vapi handles the realtime voice loop — turn detection, interruption handling, sub-second response — while n8n carries every action the agent takes out into the systems that matter: calendars, CRMs, notification channels, internal APIs. Splitting it this way means the conversational layer and the business logic can be changed independently, which is what makes the agents maintainable past week one.

    The path through

    01Inbound callthe way most of India starts
    02Realtime loopVapi — turns, barge-in
    03Intent + toolswhat does the caller want
    04n8n actionscalendar, CRM, internal APIs
    05Resolve or hand offcleanly, with context
    Conversation layer and business logic stay separable — which is what makes it maintainable past week one.

    What they own now

    • Production voice agents with barge-in and clean human handoff
    • Actions executed against real systems, not simulated
    • Conversation logic and business workflow separately maintainable
    • A first point of contact that answers at 3am
    06
    Client
    Marketing teams without a marketing-ops function
    Sector
    Agentic marketplace
    Region
    India
    Built with
    n8nLangChainMulti-model routingConnector layer

    MarketMuse

    Ten agents that do the work. Fifty utilities that do the errands.

    The tension

    The gap between "AI could do this" and "AI does this on Tuesday morning" is almost entirely plumbing. A marketing team can describe ten things they would automate and implement none of them, because each one needs credentials, a trigger, error handling and somewhere for the output to land.

    What we built

    An agentic marketplace for digital marketing: ten-plus AI agents covering lead generation, influencer tracking and analysis, ad creative generation, campaign reporting and outreach — alongside fifty-plus smaller utility apps for the single-purpose jobs that do not deserve an agent. Pick one, connect it, it runs.

    Where the leverage is

    Every agent is an orchestrated graph rather than a prompt: n8n carries state and retries, model calls are routed by task — cheap models for extraction and classification, frontier models where judgement is required — and each agent shares a common connector layer so adding the fifty-first utility is a configuration, not a build.

    The path through

    01Triggerschedule, webhook or click
    02Agent graphn8n — state, retries
    03Model routercost tracks difficulty
    04Connector layershared across 50+ apps
    05Output landswhere the team already works
    Adding the fifty-first utility is a configuration change, not a build.

    What they own now

    • 10+ marketing agents: lead gen, influencer analysis, ad creative, reporting
    • 50+ single-purpose utility apps alongside them
    • Shared connector layer — new utilities are configured, not built
    • Model routing by task, so cost tracks difficulty

    Fig. 02 — where the work sits

    Software written in India,
    answering to India.

    It is easy to build for a market you have read about. Most of what we ship answers to conditions you have to live inside to design for — a volunteer filing a report in a language the form does not offer, a graduate with the qualification and none of the network, a business whose front door is a ringing phone.

    Programme footprint
    Maharashtra · Karnataka · Gujarat
    Offices
    Dehradun ·Ahmedabad ·Indore ·Surat ·Mumbai ·Pune ·Hyderabad ·Bengaluru
    Beyond this frame
    Brussels · Canada
    MAHARASHTRAKARNATAKAGUJARATDehradunAhmedabadIndoreSuratMumbaiPuneHyderabadBengaluruBRUSSELSCANADA

    Locations are office addresses — ours and those of the organisations we build for. Shaded states mark where a multi-state programme operates, not a single site.

    Fig. 03 — method

    The tooling is not the interesting part.

    Everyone has access to the same models. The difference is knowing which eleven steps of a workflow to leave alone, and which three are worth the latency, the cost and the failure modes a model brings with it.

    build session
    _

    • Orchestration

      • n8n
      • Webhook graphs
      • Retry + state

      Where the work actually gets carried out into other systems.

    • Models

      • Claude
      • GPT
      • Gemini
      • Task-based routing

      Cheap models for extraction and classification. Frontier models only where judgement is required.

    • Voice

      • Vapi
      • Realtime STT
      • TTS
      • Turn detection

      Sub-second loops with barge-in, because people interrupt.

    • Collection

      • Scraping pipelines
      • Dedupe
      • Taxonomy normalisation

      Getting messy public data into a shape a schema will accept.

    • Agents

      • LangChain
      • Tool use
      • Competence boundaries

      An agent that knows what it cannot do is worth three that do not.

    • Streaming & storage

      • Kafka
      • ClickHouse
      • Postgres
      • Redis

      Kafka when events must survive the consumer being down. ClickHouse when someone wants to ask a question of a billion rows without waiting.

    • Observability

      • Prometheus
      • Grafana
      • OpenTelemetry
      • Structured logs

      An agent you cannot see is an agent you cannot operate. Every system ships with dashboards and alerts on day one, not after the first incident.

    • Surface

      • Next.js
      • TypeScript
      • Tailwind
      • Motion systems

      Built to a frame budget — including this page, which does no per-frame layout work while you scroll it.

    Start a build

    Bring the problem, not the spec.

    The most useful first conversation is the one where you describe what keeps going wrong, and we argue about what is actually causing it. Everything after that is easier.

    hr@nineagents.in
    Good fit
    Agentic systems, voice, extraction pipelines, the interfaces they arrive through
    Also good
    Rescuing a build that shipped and then stopped being maintainable
    Poor fit
    Staff augmentation, or AI added to a deck rather than to a workflow
    Where
    India — working across IST and CET