Hakuhodo Vietnam — HR Department·Project workspace

Hakuhodo Vietnam
HR Recruitment System

An Applicant Tracking System (ATS) built for sei-katsu-sha.

A high-level design for a recruitment system that meets every 2026 ATS expectation — and wins on three things off-the-shelf software cannot deliver in Vietnam.

Current phase
Phase 0 · Discovery
Prepared by TrendAI · Updated 2026-06-03
01 · Why we are here

A working brief for an ATS that fits Vietnam.

This site is the live working brief — requirement verification, high-level design, and a running view of how the project is being delivered.

Hakuhodo Group Vietnam runs three entities — Hakuhodo & Saigon Advertising, Hakuhodo Vietnam, and Hakuhodo Digital Vietnam — with ~800 people across HCMC and Hanoi.

The HR department needs a single system to manage recruitment end-to-end: requisitions, résumé intake, screening, interview scheduling, pipeline management, and a searchable archive of past candidates.

This site is the live working brief: requirement verification, high-level design, and a running view of how the project is being delivered.

Project objectives
  1. 01Manage recruitment workflows and candidate résumés.
  2. 02Score and match résumés against job descriptions.
  3. 03Schedule interviews across hiring panels.
  4. 04Manage and order candidates through the pipeline.
  5. 05Archive past applicants and re-search the talent pool against new roles.
  6. 06Make scoring granular — based on multiple weighted criteria.
02 · What HR asked for

Six requirements. Two of them are differentiators.

Parsing, scheduling and pipeline are universal across commercial products. Granular scoring (R3) and talent rediscovery (R6) are where this project earns its keep.

R1
Primary

Résumé management

Upload, store, search and organise candidate files across PDF, DOCX and scanned documents.

R2
Primary

Résumé parsing

Extract structured fields — contact, work history, education, skills, languages — into a queryable record.

R3
Primary

Granular weighted scoring

Score a candidate against a JD across multiple criteria, with weights tunable per role family and per-criterion justification.

R4
Primary

Interview scheduling

Calendar-synced, panel-aware booking with self-service slot selection for candidates.

R5
Primary

Pipeline ordering

Stage progression with audit trail; bulk actions; structured rejection reasons.

R6
Secondary

Talent rediscovery

Re-search the archive against a new JD; surface silver medalists and prior strong applicants before sourcing externally.

03 · The system in two halves

Table stakes — built to match the market. Differentiators — built to beat it.

Every modern ATS has the first list. The second list is what wins this against Manatal, Greenhouse and Base.vn.

Non-differentiator modules
  • Job & requisition management
    Versioned JDs, approval flow, multi-entity routing across the three Hakuhodo VN entities.
  • Résumé intake
    Public job page, email-to-apply, LinkedIn import, dedupe.
  • Résumé parsing
    Licensed parser wrapped behind a swappable interface so vendor lock-in is avoided.
  • Candidate database
    Master profile, application history, attachments, tags, notes.
  • Pipeline management
    Configurable stages per requisition type; event-sourced state machine.
  • Interview scheduling
    Microsoft 365 + Google Calendar sync; slot-finder; panel coordination.
  • Comms
    Templated email + inbound parsing — every reply threaded back onto the candidate record.
  • Hiring manager collaboration
    Shareable shortlists, structured feedback forms, scorecards.
  • Analytics & reporting
    Time-to-hire, source-of-hire, funnel conversion, recruiter throughput.
  • Auth & RBAC
    Azure AD SSO; recruiter / hiring manager / admin / read-only roles; per-entity scoping.
  • Integrations
    Job boards, LinkedIn, calendar, HRIS handoff on offer accept.
Differentiator modules — our selling points
D1

Granular, explainable scoring

Per-JD weighted rubric across skills, experience, education, portfolio, language and culture-fit signals. Every score shows criterion-by-criterion breakdown plus the résumé evidence that produced it.

Why incumbents lose here — Manatal offers weighting but shallow explanation; foreign SaaS rubrics are opaque or non-portable.

D2

Talent rediscovery as a first-class loop

Every new JD automatically re-searches the archive with semantic match and lifecycle filters before external sourcing spend. Silver medalists surface with their prior reject reason attached.

Why incumbents lose here — Greenhouse archive is keyword-only; SeekOut and Eightfold do this well but cost enterprise figures and don't host in-country.

D3

PDPL-native data residency

Built natively against the PDPL — Vietnam's Personal Data Protection Law, effective 2026-01-01 (which extends the 2023 PDPD). Consent ledger with granular purposes, per-candidate retention timer, Cross-Border Transfer Impact Assessment workflow, DPO audit dashboard, breach notification timer.

Why incumbents lose here — Foreign SaaS treats Vietnamese data protection as a checkbox. With fines up to 5% of revenue for cross-border breaches under the PDPL, the legal posture deserves to be the platform.

D4

Creative portfolio vault

First-class portfolio links (Behance, Dribbble, reel URLs), uploaded work samples with versioning, structured review forms per creative discipline.

Why incumbents lose here — Generic ATS jams portfolio links into a notes field. Agencies hire on craft; the tool should respect that.

D5

Hakuhodo skills ontology

Curated taxonomy across creative, strategy, media, digital, PR/social, CRM-CDP, e-commerce, events and retail design. Synonyms collapse — “ATL planner” = “above-the-line strategist”; “ECD” = “Executive Creative Director”.

Why incumbents lose here — Generic ontologies (ESCO, O*NET) miss agency vocabulary and force recruiters to re-tag everything by hand.

D6

Bias audit + explainability layer

Every AI decision logged with input, model version, weights, output and explanation. Quarterly impact-ratio audit job and a recruiter-facing fairness dashboard.

Why incumbents lose here — Aligns with EU AI Act high-risk hiring posture and gets Hakuhodo ahead of inevitable Vietnamese AI regulation.

D7

Conversational scheduling

Candidate-facing scheduling via Zalo, WhatsApp and email — panel-aware slot picking and reschedule-without-recruiter flow. Zalo coverage matters because it is the messaging app most Vietnamese candidates already use.

Why incumbents lose here — Paradox dominates email + SMS but Zalo scheduling is wide open in the Vietnamese market.

Design principles
P1Residency-first, then features

Every module assumes VN-hosted by default; cross-border calls require a CTIA-logged consent path.

P2Explainability over raw accuracy

Every score, ranking and recommendation ships with a human-readable justification stored alongside the decision.

P3Pluggable parsing & matching

Wrap parsing and matching behind interfaces so vendors are swappable without touching workflows.

P4Archive is an asset, not a dump

Every closed application enriches a searchable talent pool with consent state and re-contact timers.

P5Audit trail = source of truth

Every state change, score change, weight change and AI invocation is event-sourced.

P6Human-in-the-loop on every decision

AI ranks and recommends; recruiters approve. No auto-reject, no auto-advance.

04 · How it is put together

Architecture — VN-hosted core, gated external calls.

The shape in one frame: a VN-hosted core that splits cleanly into table-stakes services and the differentiator services that earn this project its keep. The dashed edge to external is CTIA-gated — every cross-border call passes through consent + impact assessment first.

System architecture · component view

Scoring engine — internals

The rubric profile is first-class data. Recruiters can clone, edit and version it without engineering involvement. Every score carries evidence anchors — “scored 9/10 on Skills because the résumé shows led Honda regional brand campaign 2023, which maps to Automotive FMCG planning via ontology node H4-AUTO-PLAN.”

D1 · scoring engine

PDPL layer — how compliance becomes infrastructure

The PDPL — Vietnam's Personal Data Protection Law, effective 2026-01-01 — extends the 2023 PDPD with stricter cross-border rules and fines up to 5% of revenue for breaches. Consent, retention and CTIA workflows are treated here as platform primitives, not bolt-ons.

D3 · PDPL compliance layer
05 · The work in motion

End-to-end recruitment, with rediscovery in the loop.

Every new requisition sweeps the archive first. External sourcing only spends once the silver-medalist shelf has been checked.

End-to-end recruitment workflow

Talent rediscovery — the deep loop

Rediscovery is opt-in per candidate via the consent ledger — no consent, no surfacing, no exception. Silver medalists (lost final round) get an explicit flag; industry data suggests they convert at roughly three times the rate of fresh applicants. Cooldown defaults to 90 days, tunable per role family.

D2 · talent rediscovery loop
06 · How we deliver

Lean agile, CI on every push, working software over status reports.

The project does not run on slide decks. It runs on a backlog, one-week sprints, automated checks, and this site.

Sprint loop · backlog → planning → build → CI → CD → preview → review → retro → backlog

One-week sprints

Short cadence keeps feedback tight. Each sprint opens with a planning ceremony, ships a demoable increment, and closes with a retro that updates the backlog before the next planning.

Master backlog as one file

Stories are markdown blocks with YAML frontmatter — estimate, owner, status, acceptance criteria. The backlog is git-tracked so every scope change has an author, a timestamp and a diff.

Definition of Done is explicit

A story closes only with: code merged, type-checks + lint passing, deployed to preview, demoable on the project site, and an entry in the time log so velocity is real.

CI on every push

TypeScript, ESLint and unit tests run before any merge to master. Failures block the merge — no override path. The build that ships to production is byte-identical to the build that passed CI.

CD via static export + Cloudflare Pages

Master pushes trigger a Pages build. Production is one git push; preview environments come for free on every branch. Rollback is one redeploy of a prior commit.

Working software over status reports

This site is the status report. Every section reflects what is true in the repo right now, not a slide deck assembled the week before.

Daily async stand-up

Written in a shared thread, not a meeting. The lane state — what is in flight, what is blocked, what is parking — is visible at all times.

Decisions captured in a register

Architectural calls are written down with date, options considered, and the why. The register is the audit trail for every reversible choice.

1 wk
Sprint cadence
100%
Pushes that run CI
Pages preview
Deploy target per branch
07 · Where we are

Phased delivery — discovery now, MVP next.

Five phases across fourteen to sixteen weeks for the differentiated MVP. One-week sprints, parallel work where dependencies allow, single-entity go-live in week sixteen. Multi-entity rollout and differentiator hardening sit deliberately outside this envelope — listed below — so the timeline doesn't quietly absorb them.

  1. 0
    Discovery close-out
    In progress

    Confirm volume + residency hardness, pick parser vendor, lock the role-family rubric list, sign off on the assumption set.

    Greenlit scope + vendor short-list.

    Weeks 1–2
  2. 1
    Foundation
    Next

    Job management, candidate ingest, parsing wrapped behind a swap-friendly interface, pipeline state machine, calendar scheduling, HR console, D3 consent ledger v1 — single entity (Hakuhodo VN) to start.

    Recruiters can run a requisition end-to-end; PDPL-safe from day one.

    Weeks 3–6
  3. 2
    Differentiators core

    D1 scoring engine with role-family rubrics + per-criterion explanation; D2 rediscovery against archive with semantic match; D6 audit log of every AI decision.

    Both differentiating requirements (R3, R6) live and explainable.

    Weeks 7–10
  4. 3
    Agency + compliance depth

    D4 portfolio vault, D5 ontology v1 across creative / strategy / media / digital, CTIA workflow + DPO dashboard.

    Agency-specific value visible; PDPL posture audit-ready.

    Weeks 11–13
  5. 4
    Polish + go-live

    D7 email scheduling + Zalo prototype, D6 fairness dashboard read-only, recruiter training, HRIS handoff stub, cutover from existing process, week-one support.

    Live for the lead entity. Candidates landing through the new system.

    Weeks 14–16
Outside the MVP envelope — sequenced after go-live
Multi-entity rollout
~4 weeks following go-live

Onboard Hakuhodo Digital Vietnam + Hakuhodo & Saigon Advertising onto the same instance with entity-scoped data isolation, separate Azure AD groups, and per-entity rubric profiles.

Differentiator hardening
Sized after week-one feedback

Production-grade D6 bias dashboard (impact-ratio reports), Zalo Official Account scheduling beyond prototype, on-prem LLM inference path if residency tightens, LinkedIn Recruiter deep integration.

Open decisions — to confirm before Phase 1
Hiring volume & recruiter seats
Sizes vector DB and search infra. < 50 reqs/yr keeps pgvector indefinitely; > 500/yr brings Qdrant in early.
PDPL residency — hard or preferred?
Whole hosting + LLM gateway design. Hard residency requires on-prem inference path; preferred allows cloud LLM with CTIA gating.
Parser vendor
Affinda, Textkernel, or in-house. Wrap-and-swap means low cost of being wrong; vendor choice still drives D5 effort.
Multi-entity isolation
SAC, Hakuhodo VN and Digital are legally distinct entities; one tenant with scoping vs separate DBs changes both data layer and PDPL posture.
Skills ontology — build vs license + extend
Licensing ESCO/O*NET and extending is faster than greenfield; ownership of vocabulary is the trade.