• Background Infrastructure project schedules cascade from macro milestones (L1) down to micro, executable activities (L5/L6),spanning multiple engineering disciplines - civil, piping, static/rotating equipment, electrical, instrumentation, HSE- each executing and reporting in parallel. While the baseline plan is well-structured (Primavera/MS Project), actual execution data flows back through daily progress reports, site diaries, discipline-wise spreadsheets, and verbal supervisor updates, each in its own format and cadence, largely disconnected from the L5/L6 activity IDs in the plan. • Problem Description There is no reliable, low-friction mechanism to capture actual start/end times of L5/L6 activities across disciplines and auto-link them back to the plan. Input quality varies with manpower skill, reporting discipline, and format.Field execution is often more granular than the planned WBS, and different disciplines describe the same physical progress differently (e.g., 'spool erected' vs. the plan's 'Erect Line 24?-XX').Consequently:
? Actual progress data is fragmented, delayed, and inconsistently structured across disciplines and contractors.
? Manual reconciliation with the baseline schedule is slow, error-prone, and often lags the schedule update cycle by days or weeks.
? Downstream performance analytics, delay/ risk analysis, and forecasting inherit this poor-quality, late data- undermining the AI performance-monitoring stack that depends on it.
? Once a project closes, the hard-won knowledge of what actually happened - real durations, real bottlenecks,real deviations from plan - is rarely captured in a structured, queryable form, so it is lost rather than feeding future project planning.
• Expected Outcome/Solution ? Ingest heterogeneous discipline-wise inputs - free-text daily reports, spreadsheets, scanned diaries,Primavera/MS Project exports - and extract activity-level actual start/end events.
? Offer an LLM-based conversational or voice interface ('time agent') for site supervisors across disciplines to log activity start/end with minimal friction, replacing rigid manual forms while still producing structured output.
? Fuzzy-match and link extracted discipline-specific activity descriptions to the correct L5/L6 plan node,handling terminology differences and granularity mismatches, and flag unmatched/new activities for planner review rather than silently dropping them.
? Auto-update actual start/end dates in the schedule/PMIS in near real time, with a confidence score and audit trail per entry.
? Produce a clean, structured, discipline-tagged actual-progress dataset that serves two purposes: (a) live input for performance analytics, delay/risk pattern discovery, and forecasting, and (b) a foundation for institutional memory building - a growing, queryable repository of real project execution patterns (actual durations, recurring delay causes, discipline-wise productivity) that future projects can learn from, instead of that knowledge staying locked in individual supervisors' experience or scattered paper records.
A working prototype demonstrating ingestion of 2–3 varied input formats (e.g., a free-text daily report and a discipline spreadsheet), extraction, and schedule-linking logic would be ideal; full production-grade OCR/ASR is not required.
• Relevant Data Availability Anonymized/ sample daily progress report formats, sample L5/L6 schedule extracts, and illustrative discipline-wise (civil/ piping/ electrical) site-diary or spreadsheet templates can be shared under NDA with Institute/ Authorised person. Live project data will not be shared; teams should work with synthetic/sample data of similar structure.