Taktline AI
The material-readiness brain for construction
Track: Track 2 — Supply Chain
One-line pitch: Taktline predicts which construction material package will miss its requested-on-job date, proves why, and launches the best recovery plan before the delay reaches site.
The problem
Construction teams do not lack data. They lack a shared, continuously updated answer to one question: will every material package be ready where and when the schedule needs it?
The evidence is scattered across baseline schedules, submittal logs, purchase orders, vendor emails, fabrication updates, inspection records and transport events. Each system reports what has already happened. None of them connects a late approval to a fabrication slip, a missed delivery, a blocked installation crew and the resulting project cost.
By the time a delay is visible on site, the least expensive recovery options are usually gone. Teams then rely on expediting, manual follow-ups and emergency resequencing. The result is avoidable schedule loss, cost escalation and low trust between contractors, vendors and project owners.
Our solution
Taktline is an AI-powered material readiness control tower. It creates a live dependency graph between materials, approvals, vendors, logistics and site activities. It continuously forecasts the arrival date of every package against its requested-on-job date and ranks risks by their probable impact on the project—not simply by how late a status field appears.
For every alert, Taktline provides:
- A probabilistic forecast — predicted on-site date, days early or late, and calibrated confidence.
- An evidence trace — the exact schedule activity, submittal, vendor message, performance pattern or logistics event that changed the forecast.
- A causal explanation — how the signals combine and which dependency is driving the risk.
- Counterfactual recovery plans — split shipments, alternative routes, crew resequencing, substitutions or expediting, ranked by days recovered, cost and probability of success.
- A human-approved action agent — prepared vendor instructions, owner notifications and schedule-change drafts that a project lead can review before anything is sent.
Why this is different
Taktline is not a chatbot placed on top of project files. Its central object is an evidence-conditioned Material Readiness Twin: a construction-specific, probabilistic graph that connects each material package to its approvals, fabrication milestones, logistics events, site location and downstream schedule activities.
The system combines four AI techniques:
- Visual document and message intelligence uses page-image retrieval and structured extraction to recover commitments, dates, quantities, blockers and uncertainty from drawings, schedules, submittals, purchase orders, emails and vendor updates while preserving the exact cited region.
- Entity resolution maps inconsistent package names, vendor SKUs and schedule codes to the same real-world material package.
- Probabilistic forecasting combines survival models, a time-series foundation-model prior and Monte Carlo graph propagation to estimate time-to-event for approval, fabrication, inspection, transit and delivery. Conformal calibration turns those distributions into defensible P10/P50/P90 dates.
- Counterfactual optimization uses constraint-aware scenario simulation—and CP-SAT at portfolio scale—to test recovery actions against the dependency graph and rank them by expected project value: schedule days protected, cost, downstream risk and confidence.
An LLM explains and operates on these outputs, but the dates, costs and confidence scores come from traceable project data and forecasting models. This makes the system both useful and defensible.
Example demo scenario
On the Aurum Tech Park project, the Level 18–26 curtain-wall package is required on site by 19 August. The vendor still reports “fabrication in progress,” which looks normal in a conventional dashboard.
Taktline connects four hidden signals: final glass approval was six days late, the vendor's glass line is currently four days behind, similar packages take 1.18 times the promised duration, and the installation activity has zero float. It predicts delivery on 28 August with 87% confidence—a nine-day miss.
The system evaluates three interventions. Its recommended plan is to split delivery by installation sequence, sending Zones 18–20 first. This protects the critical path, recovers nine days and costs ₹2.1 lakh, with a 91% confidence of success. The project lead reviews the evidence, accepts the plan and launches a workflow that drafts the vendor release instruction, prepares the schedule resequencing proposal and notifies the three accountable owners.
Technical architecture
- Connect: Primavera P6/MS Project, ERP/PO systems, document management, email, vendor portals, GPS/IoT and manual field updates.
- Understand: Docling converts complex files; ColPali retrieves visually relevant pages; Qwen2.5-VL extracts cited construction events; entity matching writes them into the time-aware Material Readiness Graph.
- Predict: Survival models and Chronos/TimesFM priors estimate each stage, Monte Carlo propagates uncertainty through schedule dependencies, and online conformal inference calibrates P10/P50/P90 intervals.
- Decide: Scenario simulation and OR-Tools CP-SAT rank feasible interventions by action cost, expected delay cost and downstream risk.
- Act: A durable, permissioned agent prepares messages and system updates; a deterministic policy gate requires human approval for every external or commercial action.
- Learn: Actual milestone outcomes recalibrate vendor profiles and project-specific forecasts.
Feasibility and build plan
Stage 1 — completed concept and interactive prototype: A realistic command centre demonstrates the complete risk-to-recovery workflow using representative construction evidence. Its server-side engine runs 20,000 Monte Carlo scenarios, produces P10/P50/P90 arrival dates and ranks three recovery options by expected project cost.
Stage 2 — working intelligence: Connect sample P6, submittal, PO and email exports; build entity resolution; train an initial time-to-event model; expose cited evidence and one executable notification workflow.
Stage 3 — pilot-ready system: Add live integrations, uncertainty calibration, permissions, audit logs and multi-project benchmarking. Validate forecasts using back-testing on historical packages.
The first version needs no new site hardware. It works with exports and APIs from tools construction teams already use, which reduces adoption friction and makes a four-week pilot practical.
Expected impact
Taktline is designed to improve four measurable outcomes:
- Earlier warning: move detection from arrival-day exceptions to multi-week forecasts.
- Better decisions: compare recovery choices using cost, days recovered and confidence.
- Less coordination waste: replace repeated status-chasing with a shared, evidence-backed view.
- Organisational learning: turn vendor and package history into a reusable forecasting advantage across projects.
Our pilot targets are a 25% reduction in material-caused schedule variance, 50% faster risk triage, more than 80% evidence coverage for alerts and measurable cost avoided for every accepted recovery action.
Scalability
The Material Readiness Graph is project- and vendor-agnostic. New projects inherit learned priors while retaining local calibration. The same engine can monitor hundreds of projects, compare vendor performance across regions and learn category-specific patterns for façade, MEP, structural steel, elevators and interior finishes.
The commercial path is a per-project or portfolio subscription, with enterprise connectors and private-model deployment for large contractors and owners. Taktline also complements Kaya's Amber vision by providing a specialised prediction and recovery layer for material readiness.
Team
- Parth Porwal IIT Roorkee
- Tanishq Gupta IIT Roorkee
Prototype note
The prototype uses representative project evidence to demonstrate the intended workflow. The forecast and optimization are genuinely computed, but their probability distributions are expert-defined until a historical project dataset is connected. All displayed performance metrics remain illustrative pilot targets, not claims from a completed production deployment.
Built With
- css3
- git
- javascript
- lucide-react
- next.js
- playwright
- react
- tailwind
- typescript
