Oriental Nippon Development
Back to portfolio
Case studyOrganisation running concurrent projects

A Virtual Project Manager: Grounded Answers and Meeting Intelligence

Project status lived across a knowledge base, several tools and the last meeting. Our assistant attends meetings, cites its sources, and answers 'where is this?' on demand, with consent built in from day one.

Industry
Professional Services & Staffing
Region
United States
Engagement
R&D Lab (joint innovation)

Outcomes

  • Cited answers on demand
  • Meetings become searchable record
  • Fewer status meetings

Stack

Python · LLM + retrieval · Deepgram · Meeting APIs · Knowledge base

Written byOriental Nippon Team

An organisation running many projects at once had a simple question it could never answer quickly: where is this project? The answer was spread across a knowledge base, several tools, and whatever was said in the last meeting.

Getting it meant asking several people and reconciling their answers. Many meetings existed mostly to pass status around.

The problem

  • Status by interrogation. Every answer meant chasing several people.
  • Meetings as data transfer. An expensive way to move information.
  • Lost decisions. What was agreed in a call often wasn't written down.
  • Patchy updates. Writing updates competed with doing the work.
BeforeAfter
"Where is this project?"Ask aroundAsk the assistant, with sources
Meeting notesWhoever rememberedCaptured and transcribed
Status updatesWritten by people, irregularlyDaily and weekly, generated
Assistant presencen/aVisible to every participant

What we built

  • A grounded assistant

    Answers come from the organisation's knowledge base and cite the source.

  • Meeting capture

    The assistant joins meetings, visibly, and transcribes them with Deepgram, so what is said becomes part of the record.

  • Automatic summaries

    Daily and weekly progress updates produced from that record.

  • Direct Q&A

    Progress, blockers and next steps on demand.

Technology and the role it played

TechnologyRole in this project
PythonAssistant services and integrations
LLM with retrievalGrounded answers that cite the knowledge base
DeepgramMeeting transcription
Meeting-capture APIsJoining meetings visibly
Knowledge-base integrationIngesting the authoritative project sources

How we delivered it

The first step was deciding what counts as the record: which sources are authoritative, and what every answer must be grounded in.

  1. Define the sources. Authoritative sources, grounding rules and consent policy.
  2. Retrieval. Knowledge-base ingestion with citations back to source.
  3. Meeting capture. With explicit consent handling and a clear signal that the assistant is present.
  4. Summaries and Q&A. Checked against what a good project manager would have written.
  5. Pilot. A few projects, with output compared to human updates.
Delivery timelineIndicative total: 11–16 weeks
  1. 01Source definition

    1–2 wks

    Authoritative sources, grounding rules, consent policy

  2. 02Retrieval layer

    3–4 wks

    Knowledge-base ingestion, retrieval, citation

  3. 03Meeting capture

    2–3 wks

    Attendance, transcription, participant notification

  4. 04Summarisation and Q&A

    3–4 wks

    Daily and weekly updates, question answering

  5. 05Pilot

    2–3 wks

    A few projects, output compared against human updates

Risks and how we handled them

RiskHow we handled it
Answers that aren't groundedAuthoritative sources defined and every answer cited
Recording people without consentConsent handling and visible presence built in from the start
Summaries worse than a human'sPilot output compared against human-written updates

Outcome

Stakeholders get an accurate project answer on demand. A whole category of status meeting, and a lot of chasing, went away.

An assistant that joins meetings has to be visible to everyone in them. Building consent in from the start avoided a conversation nobody wants to have afterwards.

What the client keeps

  • The assistant and its source
  • Retrieval and grounding configuration
  • The consent and notification policy as implemented
  • Summary history

Sources and further reading

  1. Deepgram documentation, Deepgram
  2. Guidelines for human-AI interaction, Microsoft HAX Toolkit
  3. Top 10 for Large Language Model Applications, OWASP

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