Oriental Nippon Development
Back to portfolio
Case studySaaS platform

A Platform of 20 Task-Specific AI Agents for Long-Tail Knowledge Work

Email triage, CV screening, meeting notes, security scans, each too small for a product, so all stayed manual. We built a shared agent platform and twenty specialised agents, each with its own quality bar.

Industry
Software & SaaS
Region
Global
Engagement
R&D Lab (joint innovation)

Outcomes

  • 20 specialised agents
  • Hundreds of emails triaged
  • 100+ candidates screened

Stack

Python · LLM agents · Orchestration · Web dashboard · API integrations

Written byOriental Nippon Team

Almost every organisation has a long tail of repetitive knowledge work: triaging email, scoring job applications, writing up meetings, running routine security scans, sending outreach. Each task is too small to justify buying a product. Together they eat a lot of time.

A SaaS platform asked our R&D lab to build a different answer: one agent per job, each built properly, all on one dashboard.

The problem

  • Done manually, in volume. The tasks were repetitive but still handled by people.
  • Too small to buy for. No single task justified its own tool, so they stayed manual.
  • General assistants are approximate. They handle each task adequately and none of them well, because good results depend on task-specific structure.
  • Point tools multiply. Several tools mean several interfaces and no single view.
Agent examplesWhat it does
Email triageSorts and labels incoming mail
CV scoringScreens and ranks applicants
Meeting notesTranscribes and summarises
Security scanningRuns routine scans on a schedule
Network outreachHandles professional-network outreach

What we built

  • Twenty task-specific agents

    Each designed around one job instead of one agent asked to do everything.

  • Scheduled autonomous runs

    Agents work on their own cadence instead of waiting to be asked.

  • A unified dashboard

    Configuration, output and status for every agent in one place.

  • Per-agent pricing

    A customer turns on the two agents that matter to them, not a whole suite.

Technology and the role it played

TechnologyRole in this project
PythonShared platform layer and agents
LLM agentsEach built for one job
Scheduled orchestrationAgents run on their own cadence
Web dashboardConfiguration, output and status for all agents
API integrationsConnections to email, recruiting, security and outreach tools

How we delivered it

We chose tasks by two properties: volume and rule clarity. Those two decide whether automating a task is genuinely worth it.

  1. Choose tasks. Ranked by volume and how clear the rules are.
  2. Build the shared layer. Scheduling, credentials, logging and dashboard, so each agent is a specialisation rather than a rebuild.
  3. Build agents one at a time. Each with its own evaluation set and quality bar.
  4. Instrument everything. What an agent did, and why, is always inspectable.
  5. Release incrementally. Agent by agent.
Delivery timelineIndicative total: 9–12 weeks
  1. 01Task selection

    2 wks

    Candidate tasks ranked by volume and rule clarity

  2. 02Platform layer

    5–7 wks

    Scheduling, credentials, logging, dashboard

  3. 03Instrumentation

    2–3 wks

    Action logging, output inspection, error handling

  4. 04Agent waves

    ongoing

    Agents built and released individually, each with evaluations

Risks and how we handled them

RiskHow we handled it
One agent doing everything badlyOne agent per task, each with its own evaluation set
Each agent rebuilt from scratchA shared layer for scheduling, credentials and logging
Agents acting without accountabilityEvery action logged and inspectable

Outcome

In its operating history the platform has handled real volume, including several hundred emails triaged, more than a hundred candidates screened and scored, and dozens of security scans, without those tasks occupying anyone's day.

Those are throughput figures for the platform, and we quote them as such. What a given organisation gets back depends on which agents it turns on and how much of that work it was doing by hand.

What the client keeps

  • Platform and agent source
  • Per-agent evaluation sets
  • Action logs
  • Scheduling and credential configuration
  • The dashboard

Sources and further reading

  1. Top 10 for Large Language Model Applications, OWASP
  2. AI Risk Management Framework, NIST

Facing something similar? This is the kind of work our AI, Data Science & Automation team delivers every day. Talk to us about your stack.