Azure Functions poll APIs, parse RSS feeds, and trigger on inbound emails. Adapters for GeBIZ, ePerolehan, Vietnam eProcurement, World Bank, ADB, UNGM, TED EU, SAM.gov, and all major procurement portals. Apify actors for portals with no official API.

Tender descriptions embedded with text-embedding-3-large, matched against OneLake vector index of historical wins/losses. GPT-4o classifies sector, contract type, capabilities, risk flags, and relevance score 0–100 with RAG context from similar past bids.

Classified tender matched against a Delta table of PCU business unit capabilities per country. Resolves primary assignee and CC assignees for multi-country tenders. Rules engine applies hard filters (blacklisted markets, value floors).

WF-01: structured email + Teams card + SharePoint folder creation. WF-02: bid/no-bid capture from any surface. WF-03: 48h escalation chain (officer → manager → country director). WF-04: daily deadline reminders with urgency banding.

All deadlines stored UTC. Client-side conversion to all 6 configured PCU office timezones (MY, VN, SG, ID, TH, UTC). Issuer timezone highlighted. On-click popover in feed and dashboard; always-expanded table in detail panel; pre-rendered in email.

Officers assigned to multiple Entra groups see a unified dashboard across all their assigned country/BU combinations. Data scoping enforced per-request from JWT group claims — no separate logins. Multi-country tenders flagged with joint-bid routing.

Token-based design system across Power Apps portal, Teams tab, and Outlook add-in. System preference respected on first load; officer override persisted to Entra user profile extension attribute. Email always renders light-mode regardless of preference.

Six roles (Reader, Officer, Manager, Admin, FoundryPrincipal, ServicePrincipal). Four Conditional Access policies (MFA, device compliance, location, risk). Foundry workloads use managed identity — zero secrets stored anywhere.

Azure API Management exposes all TIS endpoints: tenders CRUD, status PATCH, reassign, capability matrix, timezone registry, user preferences, sources health. JWT validation, rate limiting, App Insights logging on every call.

OneLake Gold layer extended with bid outcome fields: submitted, won, lost, cancelled, contract value, delivery performance, relationship factor, competitor names where known. This becomes the training corpus for all Phase 2 and 3 AI workloads.

Fabric AutoML experiment using tender features (sector, value, issuer country, contract type, days to deadline, risk flags) + PCU capability profile + historical outcome to train a gradient-boosted win-probability classifier. Outputs probability score separate from relevance score.

Nightly Fabric notebook analyses win/loss outcomes to update capability matrix weights automatically. Markets where PCU consistently loses trigger a capability gap flag. Sectors with high win rates get score boosted. Admin approval before each update cycle publishes.

Extract competitor names from award notices (World Bank, ADB, TED, GeBIZ all publish awarded suppliers). Build competitor presence map by sector and country. Alert officers when a known strong competitor has already won in the same sub-sector recently — inform bid/no-bid decision.

Training data for each country office stored in country-specific OneLake workspaces with Entra RBAC country-scoping. Malaysian win data never used to train Vietnam model and vice versa. Foundry workspace boundary enforces this. Microsoft Purview labels training datasets as Country-Restricted.

Power Apps officer portal adds a structured feedback form on every bid/no-bid decision: reason code, confidence, notes. For submitted bids: outcome field (won/lost), loss reason, contract value if won. This structured feedback is the highest-quality training signal — richer than outcome data alone.

SharePoint library of all past submitted proposals ingested into OneLake. AI extracts: executive summary, methodology sections, team CVs, company credentials, pricing structure, differentiators used. These become retrievable chunks in the Phase 3 proposal generation RAG index.

Power BI semantic model extended with win-rate by sector, country, issuer type, contract value band, and contract type. Officers see their team's historical performance before making a bid decision. Foundry evaluation harness publishes weekly model accuracy metrics to the same Power BI report.

Foundry weekly evaluation run against 200-tender labelled holdout set. Tracks sector accuracy, relevance score calibration, routing accuracy, RAG precision, latency p95. Power Automate sends alert to TIS.Admin if any metric drops below threshold. Prevents silent model degradation.

Foundry prompt flow reads the RFP document (uploaded PDF) and identifies required sections, mandatory attachments, evaluation criteria, and word limits. Generates a proposal outline mapped to PCU's standard section structure. Officer confirms structure before drafting begins.

Each proposal section drafted by GPT-4o using: (1) the RFP requirements for that section, (2) the most similar past-win paragraphs from PCU's proposal corpus, (3) the team's current capability profile, (4) relevant compliance clauses. Officer edits the draft — AI never auto-submits.

Foundry agent reads the RFP and cross-references PCU's compliance library (ISO certifications, Bumiputera equity statements, financial statements, professional indemnity certificates). Auto-attaches applicable documents and flags missing certifications that need to be obtained before submission.

Historical pricing data (from past proposals in OneLake) surfaces typical day-rate ranges, overhead multipliers, and margin bands for similar past contracts in the same sector and country. AI pre-fills a pricing template. Finance team reviews and approves before inclusion. PCU pricing data is country-locked and never shared externally.

For tenders flagged as multi-country mandates, Foundry prompt flow assembles capability sections from multiple PCU country offices into a single coherent proposal — matching sections to each country's demonstrated track record. Resolves lead entity, consortium structure, and local content requirements.

AI-drafted proposal exported to SharePoint as a Word document (.docx). Officer reviews in Word with tracked changes. All edits captured as a feedback signal back to the training loop (Phase 2). Approved version triggers the submission checklist workflow.

Power Automate WF-05 assembles the final submission package: cover letter, main proposal (PDF), appendices, signed declarations, financial statements. Generates a submission checklist confirming all mandatory items are present and within page/file limits. Creates final package in SharePoint for officer to download and submit.

SharePoint version history tracks every proposal draft, review, and approval step with timestamps and actor attribution. Approval chain: Lead Officer → BD Manager → Country Director (for high-value tenders above threshold). Approved version locked against further edits. Full audit trail in OneLake.

After submission, TIS monitors for award notices via the same signal sources. When an award is published that matches a submitted tender, it auto-updates the outcome field (won/lost), notifies the officer, and triggers the Phase 2 outcome data pipeline. Closes the training loop.

Foundry agent with Bing Search grounding monitors pre-procurement signals — government budget announcements, project pipeline publications, parliamentary debates about procurement — before tenders are formally published. Alerts BD team 3–6 months ahead of expected tender release.

For each new tender, AI checks PCU's current eligibility: active certifications, financial capacity thresholds, minimum experience requirements, local content rules. Flags tenders where PCU is not currently eligible and what would need to change — before officer time is spent reviewing.

For tenders where PCU has a partial capability gap, AI searches PCU's registered subcontractor and partner database to find entities that could fill the gap — enabling a joint bid rather than a decline. Ranks partners by past performance, geographic presence, and sector certification.

Foundry agent monitors government gazette publications, procurement law amendments, and Bumiputera policy changes across all PCU operating markets. Flags changes that affect TIS routing rules, capability matrix, or compliance requirements. Triggers a TIS.Admin review workflow for rule updates.

AI analyses contract terms in published RFPs for: liquidated damages clauses above PCU tolerance, unusual payment terms, performance bond requirements, governing law jurisdiction risks, and forex exposure. Generates a contract risk brief alongside the bid brief for legal and finance team review.

Won tenders trigger project creation in PCU's project management system. TIS passes structured contract data (scope, milestones, payment schedule, performance KPIs) directly to delivery team. Creates a feedback loop between delivery performance and future bid scoring — bids for contracts PCU consistently delivers well on get boosted.