ACIP is a full-stack intelligence platform for construction safety, powered by the Live Risk Intelligence Engine (LRIE). It fuses computer vision, machine learning, document intelligence, and multi-agent orchestration into one deterministic decision engine that observes, scores, decides, and escalates — closing the loop from detection to action without a human bottleneck.
Construction is the most dangerous major industry — one in five workplace fatalities happens on a site. Today's tools either report incidents after they happen or display yesterday's data in dashboards. ACIP is built to be the first platform that spans every phase of construction, fuses every signal, and acts in seconds — preventing the incident instead of documenting it.
Detection to action with no human in the bottleneck. LRIE detects, scores compounding risk, retrieves the right protocol, decides, and escalates to the right person in seconds — every step logged.
Not "missing hard hat = HIGH." Instead: missing hard hat + 12th floor + 20 mph wind + this trade's three violations this week + a crane pick in 30 minutes = CRITICAL, with specific mitigation steps.
Projects are temporary; their lessons usually leave with the people. ACIP's vector store and trained models turn every project's learnings into persistent organisational memory applied to the next site automatically.
ACIP's deepest capability is live execution monitoring — but its document and predictive modules extend risk intelligence upstream to planning and downstream to closeout.
Every AI output — vision, ML, or document — passes through LRIE's state machine before any action is taken. Each transition is typed, logged, and auditable for OSHA compliance and insurance.
Ingests site camera images, inspection photos, PDF reports, CSV logs, IoT streams, and manual entries. Each input is typed and routed to the right AI module.
Pydantic v2 schema validation, image resolution checks, document completeness verification, and duplicate detection before anything enters the pipeline.
YOLOv8 for vision, XGBoost for risk, LLM for documents — the right model runs for the input type.
A calibrated enterprise confidence score (0.0–1.0) — not a raw model output — weighing model confidence, historical accuracy, and context (time, phase, weather).
Maps score to risk class: LOW (log only), MEDIUM (dashboard flag), HIGH (notify safety officer + retrieve protocol), CRITICAL (executive escalation + possible work-stoppage).
Role-based escalation routes to the right person; RAG retrieves the specific applicable protocol via Groq LLM for an actionable mitigation — not a generic rule.
The full reasoning chain — detection through resolution — is preserved in the PostgreSQL audit log with timestamps, responsible parties, and outcome.
ACIP doesn't just detect current hazards — it learns from every project. Patterns invisible in spreadsheet reviews are surfaced by the model and applied predictively to new sites.
XGBoost finds multi-factor correlations humans miss — e.g. "new subcontractor mobilisation + concrete pour + temperature above 90°F" correlating with a 3.2× rise in heat incidents.
A new 40-story tower in Miami automatically inherits risk patterns from prior high-rise projects in hot climates — seasonal curves, trade violation profiles, critical-activity windows.
A pattern detected on Project A — scaffold violations spiking in week 3 of envelope work — is encoded into the model and applied predictively on Project B at the same phase.
Near-misses reveal failure modes without injury. ACIP treats them with the same rigour as incidents, vectorising the narratives for semantic retrieval when similar conditions recur.
Weather, temperature, and wind are correlated with incident timing so risk scores rise before conditions turn dangerous, not after.
Trained models and the vector store keep safety knowledge after a project ends — institutional memory that no longer walks out the door with the team.
Cameras catch what's visible. ACIP's Document Intelligence infers what's hidden below ground — utilities, soil conditions, abandoned infrastructure — by correlating decades of plans, surveys, and reports before a single shovel breaks ground.
A renovation in a 1960s office building needs new basement foundation work. The contractor uploads the current demolition plan. ACIP processes it and also retrieves the original 1962 mechanical drawings stored in ChromaDB at project setup. Those drawings show a 6-inch cast-iron domestic water line running diagonally through the zone now marked for excavation — capped during a 1990s renovation but never removed, so it appears on no current survey.
ACIP unifies modern computer vision, gradient-boosted ML, retrieval-augmented generation, and multi-agent orchestration — every layer chosen for production reliability and auditability.
Construction risk decisions branch. Should we retrieve safety protocols? Should we escalate? That requires a graph, not a chain. LRIE is built as a LangGraph StateGraph — each node a pure, typed, logged function.
A StateGraph lets LRIE branch, merge, and retry without losing state. Conditional edges decide whether to retrieve protocols (risk > 0.6) and whether to escalate (risk > 0.8 or critical violation). Every node's I/O is typed with Pydantic, and the full execution trace is logged.
LangChain loads documents, splits them with RecursiveCharacterTextSplitter, embeds chunks via all-MiniLM-L6-v2 into ChromaDB, and runs the retrieval chain that feeds the Groq LLM. Prompt templates keep every interaction structured and consistent.
Every decision logged with input provenance, processing chain, model metadata, confidence rationale, action taken, and resolution — designed for OSHA, insurance, and litigation defence.
Configurable hierarchies from site worker to CXO. Unacknowledged HIGH alerts auto-escalate after a configurable window; CRITICAL findings bypass the queue.
Box, SharePoint, Procore, Autodesk Construction Cloud, Revit/Navisworks, Teams, Slack, and IoT sensor feeds — ACIP fits existing construction stacks.
JWT auth via Supabase, role-based access at site/project/portfolio level, TLS in transit, encryption at rest, configurable data residency.
Designed to support OSHA 1926 documentation, ISO 45001 safety-management requirements, and client-specific EHS reporting.
CXO-ready dashboards with trend analysis, site comparison, and audit-ready decision trails — built on React, MUI, and Recharts.
ACIP is being engineered progressively through a structured curriculum, from foundational architecture to a deployed, industry-ready platform. We're transparent about where it stands.
| Architecture | Complete — system design, LRIE state machine, LangGraph orchestration, API contracts, folder structure |
| Frontend | Initialised — React + TypeScript + MUI; nine screens designed (Executive Dashboard, Live Risk Monitor, Vision AI, Document Intelligence, and more) |
| In development | LRIE state modelling, decision & confidence engines |
| Roadmap | ML risk scoring → computer vision → document intelligence → agent orchestration → full LRIE assembly → production deployment |
| Deployment target | Containerised (Docker), CI/CD via GitHub Actions, cloud-native |