◈ PRISM

AI Construction & Property Management Intelligence Platform

Two domains. One platform. Document-aware AI that works like an expert who never sleeps.

Upload contracts, drawings, RFIs, daily reports, leases, and maintenance requests — get instant insights, risk flags, delay claim detection, approval workflows, and ready-to-send documents. Built for how construction and property management actually works.

What PRISM Does

Most construction and property management software is just digital paperwork. PRISM is different — it reads your documents, understands your project, and acts like an expert assistant with full context of your entire document stack.

Construction side: Upload a contract and PRISM extracts every obligation, deadline, and risk clause. Upload a daily report and it writes the formal narrative, flags potential delay claims, and detects weather impact events. Upload an RFI and it drafts a professional response referencing your indexed documents.

Property management side: Upload a lease PDF and PRISM extracts every key term — rent escalations, CAM caps, renewal options, TI allowances. Score tenant creditworthiness with AI. Forecast NOI and cap rates. Track maintenance orders with AI diagnosis. Everything that gets submitted — leases, high-priority maintenance, CAM bills, NOI reports, high-risk tenants — flows through a human approval queue before activation.

Platform Overview

Domain	Pages	DB Tables
Construction	Dashboard · Documents · Projects · RFI Register · Change Orders · Obligations · Punch List · Submittals · Daily Log · Scheduling · Analytics · Intelligence · Safety · Generate Docs · Knowledge Graph	rfis · change_orders · obligations · punch_items · submittals · daily_logs · workers · projects · schedule_tasks
Property Management	Lease Abstraction · Tenant Risk · NOI Forecaster · Maintenance Orders · CAM Reconciliation · Daily Log · Unit Turnover · Approval Queue	leases · tenants · maintenance_requests · noi_reports · cam_reconciliations · pm_daily_logs · unit_turnovers · approvals
Shared	Settings	user_settings

Screenshots

Features

Construction — Document Intelligence

Upload any construction document — PRISM classifies it, indexes it, and makes it queryable across your whole project.

Capability	What it does
Auto Classification	AI identifies document type: Contract, Drawing, RFI, Spec, Daily Report, Change Order, or Meeting Minutes
Cross-Document Q&A	Ask questions in plain English — get answers with source citations across all uploaded documents
Project Dashboard	Automatically extracts project name, contract value, parties, key dates, and liquidated damages
Risk Intelligence	Continuously scans all documents for risks and surfaces them without being asked
Knowledge Graph	Entity-relationship visualization — people, companies, obligations, clauses, and how they connect

Construction — Contract & Compliance

Capability	What it does
Contract Risk Analysis	Clause-level risk scoring with severity ratings (Critical / High / Medium / Low) and recommended actions
Obligations Tracker	Extracts every deadline, notice requirement, and milestone — color-coded by urgency, never miss a date
Spec Compliance	Checks submittals against specification requirements with a compliance score and line-by-line gap analysis

Construction — Project Administration (AI-augmented)

Tool	What makes it different
RFI Register	Full log with status tracking, days-open counter, AI-generated responses, and refresh button to surface agent-created RFIs instantly
Change Order Register	Contract value waterfall (Original → Approved → Pending → Revised), AI clause assessment, one-click approve/reject
Punch List	Priority-scored deficiency tracking with ball-in-court workflow and AI auto-categorization
Unit Turnover Checklist	25-point inspection checklist auto-saved to DB on every toggle — tracks all saved units with In Progress / Complete status
Submittals	Ball-in-court tracking, days-in-review counter, AI pre-review with compliance scoring and red-flag detection
Daily Log	Structured entry → AI writes the formal narrative, detects delay claims, flags weather impact events
Projects	Create and manage projects with live status, completion %, phase, RFI count, and workers — all DB-persisted
Scheduling	Gantt-style task management with phase grouping, Kanban view, natural language task entry — DB-persisted
Workforce	Crew management, trade tracking, daily rate, and attendance — persisted per session

Construction — Cost & Schedule Intelligence

Capability	What it does
Cost Forecasting	EAC, CPI, SPI, cost variance analysis, and scenario range from your project data
Predictive Delay Risk	Phase-by-phase delay probability with recovery recommendations
Schedule Optimization	AI reorders and compresses your task list to find time savings
Natural Language Task Entry	Add schedule items in plain English — AI parses dates, durations, and dependencies
Subcontractor Scorecard	Rates subcontractors on quality, schedule, safety, communication, and cost
Meeting Intelligence	Upload a meeting PDF → extracts action items, decisions, risks, and follow-ups

Construction — Safety & Field

Capability	What it does
Safety AI	Upload a site photo — detects PPE compliance, identifies hazards, generates a safety score
Blueprint Analysis	Gemini Vision reads drawing sheets — counts components, flags coordination issues, extracts dimensions
Document Generation	One-click generation of RFI responses, delay notices, change order assessments, and weekly summaries

Property Management — Lease Abstraction

Capability	What it does
AI Lease Parsing	Upload any lease PDF — extracts rent, escalations, CAM cap, TI allowance, renewal options, termination clauses
Key Term Summary	Plain-English summary of every lease alongside the raw extracted data
Approval Gate	Every abstracted lease is submitted to the Approval Queue before status becomes Active
Status Tracking	Pending Approval → Active / Rejected — shown inline on each lease row

Property Management — Tenant Risk Scoring

Capability	What it does
AI Credit Assessment	Scores each tenant across 5 risk dimensions — payment history, financial stability, business type, lease terms, references
Risk Level Classification	Low / Medium / High with detailed narrative
Auto-Escalation	High-risk tenants are automatically submitted to the Approval Queue for supervisor review
Manual Escalation	One-click escalate button for any tenant flagged as high risk without a pending approval

Property Management — NOI Forecaster

Capability	What it does
NOI Calculation	Gross potential rent → vacancy loss → other income → effective gross → operating expenses → NOI
Cap Rate Analysis	Automatic cap rate with AI performance benchmarking
AI Recommendations	Actionable suggestions to improve NOI based on the property's numbers
Approval Gate	Every NOI report requires supervisor approval before it can be Published

Property Management — Maintenance Orders

Capability	What it does
Work Order Tracking	Full maintenance lifecycle — Open → In Progress → Resolved with vendor assignment and cost tracking
AI Diagnosis	AI assesses the issue, estimates root cause, and recommends resolution approach
Dispatch Approval	High-priority maintenance orders require dispatch approval before work begins
Approval Status	Awaiting Dispatch / Dispatch Approved / Dispatch Rejected — shown inline on each order

Property Management — CAM Reconciliation

Capability	What it does
Pro-Rata Allocation	Calculates each tenant's CAM share based on leasable area
Cap Enforcement	Applies CAM cap percentages and flags tenants over cap
AI Summary	Narrative explaining the reconciliation, variances, and billing recommendations
Approval Gate	Every CAM reconciliation requires supervisor sign-off before billing goes out

Property Management — Unit Turnover

Capability	What it does
25-Point Checklist	Covers all rooms, systems, appliances, and exterior items
Auto-Save on Toggle	Every checkbox change is debounced 800ms and persisted to Supabase — no Save button needed
Multi-Unit Management	Switch between units; each unit's checklist is independently saved and loaded from DB
Completion Tracking	Saved units strip shows Complete (green) or In Progress (amber) status across all units
Notes	Free-text notes field per unit — also auto-saved

Human Guardrails — Approval Queue

Every significant action flows through a supervisor approval layer. No AI decision takes effect without a human sign-off.

Trigger	Type	Approval Required	Status Transition
Lease abstracted	`lease`	Supervisor	Pending → Active / Rejected
High-priority maintenance	`maintenance`	Dispatch manager	Pending → Dispatch Approved / Open
CAM reconciliation run	`cam`	Supervisor	Pending → Finalized / Draft
NOI report generated	`noi`	Supervisor	Pending → Published / Draft
High-risk tenant added	`tenant`	Supervisor	Pending → Reviewed / Action Required

The Approval Queue page shows:

Stats cards: Total · Pending · Approved · Rejected
Pending tab with full context cards (type, title, description, submitted by, date, reference ID)
Approve / Reject with a notes modal — review notes are stored and shown on the record
History tab with all resolved approvals and reviewer identity + timestamp

AI Copilot — Streaming Multi-Agent Chat

Every page has a floating AI Copilot panel powered by a LangGraph multi-agent system with real-time token streaming.

Feature	Detail
Multi-Agent Routing	Keyword supervisor routes to the right specialist — SiteOps, RFI & Contracts, Schedule & Punch, Risk & Analysis, or Property Management
SSE Token Streaming	Token-by-token output via Server-Sent Events — text appears as it generates, no waiting
Live Tool Chips	Yellow pulsing chips show tools being called in real time; turn green when complete
Smart Model Routing	Simple questions → LLaMA 3.1 8B (fast). Complex analysis → LLaMA 3.3 70B. Falls back automatically on rate limit
Cross-Session Memory	Key facts extracted from conversations and stored in Supabase — surfaced as context in future sessions
Semantic Cache	Repeat questions answered instantly from cache (0.92 cosine similarity threshold) — no LLM call needed
Feedback Buttons	👍 / 👎 on every response — stored in Supabase for continuous evaluation
RAG Integration	All specialists can call `search_documents` to answer questions grounded in your uploaded documents
Voice Agent	🎙 mic button → speak your question → agent answers → 🔊 browser reads response aloud (Web Speech API, Chrome/Edge)
Write Tools	Agents can create records directly — `create_maintenance_request` (unit, category, priority), `create_rfi` (subject, assigned_to, priority, due_date) — Pydantic-validated schemas, results visible in UI immediately

Specialists:

Agent	Handles
SiteOps	Daily logs, weather risk, crew, project status
RFI & Contracts	RFIs, change orders, obligations, submittals
Schedule & Punch	Tasks, overdue items, punch list, progress
Risk & Analysis	Weather impact, delay claims, compliance, general questions
Property	Leases, tenants, maintenance, NOI, CAM

Guardrails — Safe AI at Every Layer

Four layers of protection run on every input and output before the agent ever sees a message.

Layer	What it catches
Prompt Injection Detection	16 regex patterns — jailbreaks, role-play overrides, "ignore previous instructions", DAN, system prompt leaks
Topic Guardrail	Off-domain questions (recipes, politics, harmful content, weapons) are rejected with a clear message
Output Moderation	Harmful content patterns blocked on the way out — with construction-safe exceptions (e.g. "explosive" in a demolition context is fine)
JSON Schema Validation	Every agent response coerced to `AgentResponse` Pydantic schema — malformed outputs never reach the UI

RAG Pipeline — Document-Grounded Answers

Component	Detail
Embeddings	`fastembed` with `BAAI/bge-small-en-v1.5` (384-dim, ONNX, CPU-only — no GPU needed)
Vector Store	PGVector in Supabase — `document_chunks` table with `pgvector` extension
Chunking	Smart chunking: section-aware for contracts, paragraph-aware for reports, 512-token overlap fallback
Hybrid Retrieval	70% cosine similarity + 30% BM25 keyword score — best of both worlds
Reranking	Cohere Rerank API (optional) for precision boost; falls back to hybrid score if key absent
Semantic Cache	Cache hits at ≥0.92 cosine similarity skip the LLM entirely — 20-minute TTL, auto-invalidated on any write operation
Cross-Session Memory	LLM extracts key facts after each answer and stores them in `agent_memory` — recalled by future sessions

Evaluation & Observability

Feature	Detail
RAGAS Metrics	`faithfulness`, `answer_relevancy`, `context_precision` — run with `python -m eval.ragas_eval --mlflow`
Benchmark Runner	20 Q&A pairs across all 5 agents — scores routing accuracy, tool accuracy, keyword hit rate, latency
MLflow Tracking	`--mlflow` flag on both eval scripts logs all metrics to `prism-benchmark` / `prism-ragas` experiments — view at `mlflow ui`
DVC Pipelines	`dvc repro` runs the full eval pipeline (benchmark → ragas) and tracks data + output versions
Local Trace Logging	Every agent run logged to `agent_traces` in Supabase — latency, model, tools called, estimated cost
Human Feedback	`POST /eval/feedback` stores 👍/👎 with session and message ID — queryable via `GET /eval/feedback/{session_id}`
LangSmith	Set `LANGCHAIN_TRACING_V2=true` + `LANGCHAIN_API_KEY` in `.env` for full LangSmith trace dashboard
E2E Tests	25 Playwright tests covering auth, navigation, copilot, and RFI — run with `npm run test:e2e` from `frontend/`

MCP Integration — Talk to Your Project in Claude Desktop

This branch adds an MCP (Model Context Protocol) server so Claude Desktop can call all of PRISM's construction AI tools directly in conversation — no UI required.

"Create an RFI asking about the rebar spec at grid line C4" "Assess this change order — $45,000 for unforeseen rock excavation" "What does the contract say about liquidated damages?" "Predict delay risk for Foundation, Structure, and MEP phases"

MCP Tool	What it does
`ask_project`	Cross-document Q&A with citations
`get_dashboard`	Project summary, contract value, key dates
`contract_risk_analysis`	Clause-level risk scoring
`get_obligations`	All contract deadlines and milestones
`create_rfi` / `respond_rfi` / `list_rfis`	Full RFI lifecycle
`create_change_order` / `assess_change_order` / `list_change_orders`	Full CO lifecycle
`predict_delays`	Phase-level delay probability
`cost_forecast`	EAC, CPI, SPI, variance analysis
`create_punch_item`	New deficiency item
`generate_daily_narrative`	Formal narrative + delay claim detection
`optimize_schedule`	Critical path compression
`safety_analyze`	PPE + hazard detection from site photo
`subcontractor_scorecard`	Performance ratings across 5 dimensions
`generate_document`	Ready-to-send RFI / delay notice / CO / weekly summary

Tech Stack

Layer	Technology
Frontend	React 19 · TypeScript · Vite
Styling	Inline styles — zero CSS dependencies · Outfit + JetBrains Mono
Routing	React Router v7
Backend	FastAPI · Uvicorn · Python 3.11
Database	Supabase (PostgreSQL) — 15 tables, all RLS disabled (service key)
LLM	LLaMA 3.1 8B Instant via Groq API
LLM Router	LLaMA 3.1 8B (fast) · LLaMA 4 Scout 17B (balanced) · LLaMA 3.3 70B (complex) via Groq
Multi-Agent	LangGraph 1.x `create_react_agent` — 5 specialist agents with keyword supervisor routing, recursion limit 50
HuggingFace Models	BART-large-cnn (summarization) · BART-MNLI (zero-shot) · BERT-NER · RoBERTa-SQuAD2 (extractive QA) · CLIP · BLIP — selectable per task in Settings
Voice Agent	Browser Web Speech API — speak to AI Copilot, responses read back via SpeechSynthesis (Chrome/Edge, free)
Vision AI	Gemini 2.0 Flash (primary) · Groq LLaMA 4 Scout vision via frame extraction (fallback)
Embeddings	fastembed `BAAI/bge-small-en-v1.5` (384-dim, ONNX, CPU-only)
Vector Store	PGVector in Supabase + FAISS (legacy document Q&A)
RAG	Hybrid BM25 + cosine retrieval · Cohere Rerank (optional) · semantic cache
Evaluation	RAGAS · LangSmith · local Supabase trace logging
MLOps	MLflow experiment tracking · DVC data + pipeline versioning
E2E Testing	Playwright (25 tests — auth, navigation, copilot, RFI)
PDF Parsing	pdfplumber + pypdf
MCP Server	Python MCP SDK · httpx
Frontend Hosting	Firebase Hosting
Backend Hosting	Google Cloud Run

Project Structure

prism/
├── backend/
│   ├── main.py                  ← All FastAPI routes (single file)
│   ├── mcp_server.py            ← MCP server — exposes tools to Claude Desktop
│   ├── architecture-design.html ← Canvas-drawn system architecture diagram
│   ├── requirements.txt
│   ├── Dockerfile
│   ├── supabase_schema.sql      ← Core 15 tables (run in Supabase SQL editor)
│   ├── supabase_rag_schema.sql  ← RAG tables: document_chunks, semantic_cache, agent_memory
│   ├── supabase_eval_schema.sql ← Eval tables: agent_traces, eval_feedback
│   ├── agents/
│   │   ├── orchestrator.py      ← LangGraph multi-agent graph + SSE streaming
│   │   ├── tools.py             ← All LangChain tools (DB + RAG)
│   │   ├── guardrails.py        ← 4-layer input/output safety pipeline
│   │   ├── router.py            ← Smart model selection by question complexity
│   │   ├── tracer.py            ← Local trace logging to Supabase
│   │   └── rag/
│   │       ├── embedder.py      ← fastembed singleton
│   │       ├── chunker.py       ← Smart/section/paragraph chunking
│   │       ├── vector_store.py  ← PGVector upsert + search via Supabase RPC
│   │       ├── retriever.py     ← Hybrid BM25 + cosine retrieval + Cohere rerank
│   │       ├── cache.py         ← Semantic cache at 0.92 threshold
│   │       ├── memory.py        ← Cross-session memory extract + retrieve
│   │       └── pipeline.py      ← ingest_text() orchestrator
│   ├── eval/
│   │   ├── ragas_eval.py        ← RAGAS metrics (faithfulness, relevancy, precision) + MLflow logging
│   │   ├── runner.py            ← Benchmark runner (20 Q&A pairs, all 5 agents) + MLflow logging
│   │   └── benchmark.json       ← Ground truth Q&A dataset
│   ├── dvc.yaml                 ← DVC pipeline (benchmark + ragas stages, reproducible eval)
│   └── .env                     ← GROQ_API_KEY · GEMINI_API_KEY · SUPABASE_URL · SUPABASE_SERVICE_KEY
│
└── frontend/
    ├── public/
    ├── firebase.json
    ├── .firebaserc
    ├── playwright.config.ts      ← E2E test config (Chromium, auto-start dev server)
    ├── tests/
    │   └── e2e/
    │       ├── auth.spec.ts      ← Login + role selection tests
    │       ├── navigation.spec.ts← Sidebar navigation tests (all 9 pages)
    │       ├── copilot.spec.ts   ← AI Copilot panel tests
    │       └── rfi.spec.ts       ← RFI Register tests
    └── src/
        ├── App.tsx               ← Router · localStorage auth · session/project state
        ├── components/
        │   ├── Sidebar.tsx       ← Collapsible nav — construction + PM groups
        │   ├── Navbar.tsx        ← Page title · AI status chip · user badge
        │   └── AICopilot.tsx     ← Streaming AI Copilot — SSE token stream, tool chips, 👍👎 feedback
        └── pages/
            ├── Login.tsx         ← Demo role picker + auth state init
            ├── construction/
            │   ├── Dashboard.tsx        ← Project overview, risk cards, key metrics
            │   ├── Documents.tsx        ← Upload, classify, build FAISS index
            │   ├── Projects.tsx         ← Project CRUD — DB-persisted + demo projects
            │   ├── RFIRegister.tsx      ← RFI lifecycle with AI response generation
            │   ├── ChangeOrders.tsx     ← CO waterfall + AI clause assessment
            │   ├── Obligations.tsx      ← Contract deadline tracker
            │   ├── PunchList.tsx        ← Deficiency tracking + Unit Turnover checklist
            │   ├── Submittals.tsx       ← Ball-in-court + AI pre-review
            │   ├── DailyLog.tsx         ← Structured log + AI narrative + delay detection
            │   ├── Scheduling.tsx       ← Gantt + Kanban + NL task entry — DB-persisted
            │   ├── Analytics.tsx        ← Cost forecast · delay risk · schedule optimization
            │   ├── Intelligence.tsx     ← Blueprint CV · 3D vis · materials forecasting
            │   ├── Safety.tsx           ← PPE detection · hazard scoring from site photo
            │   ├── GenerateDocs.tsx     ← One-click document generation
            │   ├── Workforce.tsx        ← Crew management + daily rate tracking
            │   └── KnowledgeGraph.tsx   ← Entity-relationship graph visualization
            └── property/
                ├── LeaseAbstraction.tsx ← PDF lease parsing + approval submission
                ├── TenantRisk.tsx       ← AI credit scoring + auto-escalation
                ├── NOIForecaster.tsx    ← NOI/cap rate + approval gate
                ├── MaintenanceOrders.tsx← Work orders + AI diagnosis + dispatch approval
                ├── CAMReconciliation.tsx← Pro-rata CAM + cap enforcement + approval
                ├── PMDailyLog.tsx       ← Property daily activity log
                └── Approvals.tsx        ← Supervisor approval queue (all types)

Supabase Schema

Run these SQL files in the Supabase SQL editor in order:

backend/supabase_schema.sql — core 15 tables
backend/supabase_rag_schema.sql — RAG tables + PGVector RPCs
backend/supabase_eval_schema.sql — traces + feedback tables

Requires pgvector extension — enable in Supabase Dashboard → Database → Extensions → vector.

Construction Tables

rfis — RFI Register

Column	Type	Notes
id	text PK
session_id	text
subject, description, response	text
status	text	Open / Responded / Closed
priority	text	Low / Medium / High / Critical
submitted_by, assigned_to	text
date_submitted, date_due, date_responded	text
days_open	integer

change_orders — Change Order Register

Column	Type	Notes
id	text PK
session_id	text
title, description, reason	text
amount	numeric
status	text	Pending / Approved / Rejected
submitted_by	text
assessment	jsonb	AI assessment result

obligations — Contract Obligations

Column	Type	Notes
id	text PK
session_id	text
description, due_date, responsible_party, type	text
completed	boolean

punch_items — Punch List

Column	Type	Notes
id	text PK
session_id	text
description, location, trade, category	text
priority	text	Low / Medium / High / Critical
status	text	Open / In Progress / Resolved
ball_in_court	text	Contractor / Owner / Architect / Inspector

submittals — Submittals Register

Column	Type	Notes
id	text PK
session_id	text
title, spec_section, submitted_by, reviewer	text
status	text	Pending Review / Approved / Rejected / Revise & Resubmit
date_submitted, required_date, review_deadline	text
compliance_score	numeric
flags	jsonb	Red flags from AI review
ai_review	text	Full AI review narrative

daily_logs — Construction Daily Log

Column	Type	Notes
id	text PK
session_id	text
date, weather, work_performed, delays, incidents	text
crew_count	integer
labor_hours	numeric
narrative	text	AI-generated formal narrative
delay_claims	jsonb	Detected delay claims
weather_impact	text

workers — Workforce

Column	Type	Notes
id	text PK
session_id	text
name, trade, role, company, phone, email	text
status	text	Active / Inactive
start_date	text
daily_rate	numeric

projects — Project Registry

Column	Type	Notes
id	text PK
session_id	text
name, client, value, phase	text
status	text	On Track / At Risk / Delayed
completion	integer	0–100
rfi, workers	integer
start_date, end_date	text

schedule_tasks — Scheduling

Column	Type	Notes
id	text PK
session_id	text
name, phase, assignee	text
start_day, duration, progress	integer
status	text	Upcoming / In Progress / Done / Delayed
priority	text	low / medium / high / critical

Property Management Tables

leases — Lease Abstraction

Column	Type	Notes
id	text PK
session_id	text
property_address, tenant_name, permitted_use	text
lease_start, lease_end	text
monthly_rent, security_deposit, ti_allowance	numeric
rent_escalation, cam_cap, renewal_options, termination_clause	text
cam_included	boolean
ai_summary	text	Plain-English lease summary
status	text	Pending Approval / Active / Rejected

tenants — Tenant Risk

Column	Type	Notes
id	text PK
session_id	text
name, unit, email, phone	text
move_in, lease_end	text
monthly_rent	numeric
risk_score	text	AI score string
risk_level	text	Low / Medium / High
risk_details	text	AI assessment narrative
status	text	Active / Reviewed / Action Required

maintenance_requests — Maintenance Orders

Column	Type	Notes
id	text PK
session_id	text
unit, tenant_name, category, description	text
priority	text	Low / Medium / High / Emergency
status	text	Open / In Progress / Resolved / Dispatch Approved
assigned_to	text
date_submitted, date_due, date_resolved	text
ai_diagnosis	text	AI root cause and recommendation
estimated_cost	numeric

noi_reports — NOI Forecaster

Column	Type	Notes
id	text PK
session_id	text
property_name, report_period	text
gross_potential_rent, vacancy_loss, other_income	numeric
effective_gross, operating_expenses, noi	numeric
cap_rate	numeric
ai_summary, ai_recommendations	text

cam_reconciliations — CAM Reconciliation

Column	Type	Notes
id	text PK
session_id	text
property_name, reconcile_year	text
total_cam_pool, total_leasable	numeric
tenants_data	jsonb	Per-tenant allocation breakdown
ai_summary	text

pm_daily_logs — Property Daily Log

Column	Type	Notes
id	text PK
session_id	text
property_address, date	text
activities, maintenance_notes, tenant_interactions, occupancy_notes	text
ai_summary	text

unit_turnovers — Unit Turnover Checklist

Column	Type	Notes
id	text PK
session_id	text
unit	text	Unit identifier
checked_items	jsonb	Array of completed checklist item IDs
notes	text	Free-text notes
completed	boolean
completed_at	timestamptz
updated_at	timestamptz	Auto-updated on save

approvals — Human Approval Queue

Column	Type	Notes
id	text PK
session_id	text
type	text	lease / maintenance / cam / noi / tenant
reference_id	text	FK to the source record's id
title, description	text
status	text	Pending / Approved / Rejected
requested_by	text
review_notes	text	Reviewer's notes
reviewed_by	text
created_at	timestamptz
reviewed_at	timestamptz

user_settings — Per-User Preferences

Column	Type	Notes
uid	text PK	Firebase/auth UID
session_id	text
notifications	boolean	In-app alerts
auto_risk	boolean	Auto-scan on upload
email_alerts	boolean	Email overdue alerts
updated_at	timestamptz

API Reference

Session

Method	Endpoint	Description
POST	`/session/new`	Create session — returns `session_id`

Documents & Q&A

Method	Endpoint	Description
POST	`/construction/upload-classify`	Upload + auto-classify documents
POST	`/construction/build-project`	Build FAISS index + generate dashboard
GET	`/construction/dashboard/{session_id}`	Get project dashboard
POST	`/construction/ask`	Cross-document Q&A with citations

AI Analysis — Construction

Method	Endpoint	Description
POST	`/construction/contract-risk`	Clause-level contract risk analysis
POST	`/construction/cost-forecast`	EAC, CPI, SPI, variance forecast
POST	`/construction/predict-delays`	Phase-level delay probability
POST	`/construction/optimize-schedule`	AI schedule compression
POST	`/construction/nl-task`	Parse natural language task
POST	`/construction/subcontractor-score`	Subcontractor performance scorecard
POST	`/construction/meeting-intelligence`	Extract actions/decisions from meeting PDF
POST	`/construction/spec-compliance`	Check submittal against spec
POST	`/construction/safety-analyze`	PPE + hazard analysis from site photo
POST	`/construction/analyze-blueprint`	Blueprint/drawing computer vision analysis
POST	`/construction/generate`	Generate RFI / delay notice / CO / weekly summary
POST	`/construction/knowledge-graph`	Build entity-relationship graph from documents

RFI Register

Method	Endpoint	Description
POST	`/construction/rfi-register/create`	Create new RFI
GET	`/construction/rfi-register/{session_id}`	List all RFIs
POST	`/construction/rfi-register/respond`	AI-generate RFI response
PUT	`/construction/rfi-register/update`	Update RFI status or response

Change Orders

Method	Endpoint	Description
POST	`/construction/co-register/create`	Create new change order
GET	`/construction/co-register/{session_id}`	List all change orders
POST	`/construction/co-register/assess`	AI assessment of change order
PUT	`/construction/co-register/update`	Approve / reject / update

Contract Obligations

Method	Endpoint	Description
POST	`/construction/obligations/extract`	Extract obligations from contract
GET	`/construction/obligations/{session_id}`	List all obligations
PUT	`/construction/obligations/complete`	Toggle obligation complete

Punch List

Method	Endpoint	Description
POST	`/construction/punch/create`	Create punch item
GET	`/construction/punch/{session_id}`	List all punch items
POST	`/construction/punch/ai-categorize`	AI categorize open items
PUT	`/construction/punch/update`	Update status / ball-in-court

Submittals

Method	Endpoint	Description
POST	`/construction/submittals/create`	Create new submittal
GET	`/construction/submittals/{session_id}`	List all submittals
POST	`/construction/submittals/ai-review`	AI pre-review with compliance scoring
PUT	`/construction/submittals/update`	Update status

Daily Log — Construction

Method	Endpoint	Description
POST	`/construction/daily-log/create`	Create daily log entry
GET	`/construction/daily-log/{session_id}`	List all daily logs
POST	`/construction/daily-log/ai-summary`	Generate narrative + detect delay claims

Projects

Method	Endpoint	Description
POST	`/construction/projects/create`	Create new project
GET	`/construction/projects/{session_id}`	List all projects
PUT	`/construction/projects/update`	Update project status / completion

Scheduling

Method	Endpoint	Description
POST	`/construction/schedule/create`	Create new schedule task
GET	`/construction/schedule/{session_id}`	List all schedule tasks
PUT	`/construction/schedule/update`	Update task status / progress

Workforce

Method	Endpoint	Description
POST	`/construction/workers/create`	Add worker
GET	`/construction/workers/{session_id}`	List all workers
PUT	`/construction/workers/update`	Update worker details
DELETE	`/construction/workers/delete/{id}`	Remove worker

Property Management — Leases

Method	Endpoint	Description
POST	`/pm/leases/abstract`	AI extract lease terms from PDF
GET	`/pm/leases/{session_id}`	List all leases
PUT	`/pm/leases/update`	Update lease status

Property Management — Tenants

Method	Endpoint	Description
POST	`/pm/tenants/add`	Add tenant + AI risk score
GET	`/pm/tenants/{session_id}`	List all tenants

Property Management — Maintenance

Method	Endpoint	Description
POST	`/pm/maintenance/create`	Create maintenance order + AI diagnosis
GET	`/pm/maintenance/{session_id}`	List all maintenance orders
PUT	`/pm/maintenance/update`	Update status / assignment

Property Management — NOI

Method	Endpoint	Description
POST	`/pm/noi/forecast`	Run NOI forecast + AI recommendations
GET	`/pm/noi/{session_id}`	List all NOI reports

Property Management — CAM

Method	Endpoint	Description
POST	`/pm/cam/reconcile`	Run CAM reconciliation + AI summary
GET	`/pm/cam/{session_id}`	List all CAM reconciliations

Property Management — Daily Log

Method	Endpoint	Description
POST	`/pm/daily-log/create`	Create property daily log
GET	`/pm/daily-log/{session_id}`	List all property daily logs
POST	`/pm/daily-log/ai-summary`	Generate AI summary

Property Management — Unit Turnover

Method	Endpoint	Description
POST	`/pm/turnover/save`	Upsert checklist state for a unit
GET	`/pm/turnover/{session_id}`	List all saved unit turnovers
GET	`/pm/turnover/{session_id}/{unit}`	Get specific unit checklist state

Approvals (Human Guardrails)

Method	Endpoint	Description
POST	`/approvals/request`	Submit record for approval
GET	`/approvals/{session_id}`	List all approvals
PUT	`/approvals/review`	Approve or reject — cascades to source record

User Settings

Method	Endpoint	Description
GET	`/user/settings/{uid}`	Get user preferences
PUT	`/user/settings/save`	Save preferences (upsert)

AI Copilot — Multi-Agent

Method	Endpoint	Description
POST	`/agent/chat`	Single-turn agent response (blocking)
POST	`/agent/stream`	SSE streaming agent response — yields `meta \| token \| tool_start \| tool_end \| done \| error`

RAG — Document Ingestion & Search

Method	Endpoint	Description
POST	`/rag/ingest`	Chunk, embed, and upsert a document into PGVector
GET	`/rag/docs/{session_id}`	List all ingested documents
DELETE	`/rag/docs/{session_id}/{doc_name}`	Remove a document from the vector store
POST	`/rag/search`	Semantic search across ingested documents

Evaluation & Feedback

Method	Endpoint	Description
POST	`/eval/feedback`	Store 👍/👎 rating on an agent response
GET	`/eval/feedback/{session_id}`	List all feedback for a session
GET	`/eval/traces/{session_id}`	List agent trace logs for a session

Testing

E2E Tests (Playwright)

25 tests covering auth, sidebar navigation, AI Copilot, and RFI Register.

cd frontend
npm run test:e2e           # headless, all tests
npm run test:e2e:ui        # Playwright UI mode (interactive)
npm run test:e2e:report    # open last HTML report

Against production:

$env:PLAYWRIGHT_BASE_URL = "https://your-app.web.app"
$env:CI = "true"
npx playwright test

Eval Pipeline (MLflow + DVC)

cd backend

# Run benchmark only
python -m eval.runner --session_id demo --mlflow

# Run RAGAS only
python -m eval.ragas_eval --session_id demo --mlflow

# Run full pipeline (both stages, tracks data versions)
dvc repro

# View metrics in MLflow UI
mlflow ui   # open http://localhost:5000

MCP Setup

1. Start the PRISM backend

cd backend
venv\Scripts\activate
uvicorn main:app --reload --port 8000

2. Add to Claude Desktop config

Find your config file at:

C:\Users\<YourName>\AppData\Local\Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude\claude_desktop_config.json

Add the mcpServers block:

{
  "mcpServers": {
    "prism": {
      "command": "C:/Users/<YourName>/prism/backend/venv/Scripts/python.exe",
      "args": ["C:/Users/<YourName>/prism/backend/mcp_server.py"],
      "env": {
        "PRISM_API_URL": "http://localhost:8000"
      }
    }
  }
}

3. Restart Claude Desktop — fully quit via system tray → Quit, then reopen

4. Verify — Settings → Developer → PRISM should show green/connected

Demo Access

No sign-up required. Choose a role at the login screen:

Role	Profile
Project Manager	Marcus Rivera — full access to all modules
Site Engineer	Sarah Kim — field and scheduling focus
Quantity Surveyor	Ahmed Hassan — cost and contracts focus
Safety Officer	Priya Nair — safety and compliance focus

Local Development

Prerequisites

Python 3.11+
Node.js 18+
Supabase project (free tier works)
Groq API key (free tier available)
Gemini API key (free tier available)

1. Run the Supabase schema

Open the Supabase SQL editor for your project and run backend/supabase_schema.sql — creates all 15 tables and disables RLS.

2. Backend

cd prism/backend

python -m venv venv
venv\Scripts\activate        # Windows
source venv/bin/activate     # Mac/Linux

pip install -r requirements.txt

Create backend/.env:

GROQ_API_KEY=your_groq_key_here
GROQ_MODEL=llama-3.1-8b-instant
GEMINI_API_KEY=your_gemini_key_here
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_SERVICE_KEY=your_service_role_key_here

# Optional
COHERE_API_KEY=your_cohere_key_here
LANGCHAIN_API_KEY=your_langsmith_key_here
LANGCHAIN_TRACING_V2=true

uvicorn main:app --reload --port 8000

3. Frontend

cd prism/frontend
npm install
npm run dev

Open http://localhost:5173

Deployment

Frontend — Firebase Hosting

cd frontend
npm run build
firebase deploy --only hosting

Backend — Google Cloud Run

cd backend
gcloud run deploy prism-backend `
  --source . `
  --region us-central1 `
  --platform managed `
  --allow-unauthenticated `
  --set-env-vars "GROQ_API_KEY=...,GROQ_MODEL=llama-3.1-8b-instant,GEMINI_API_KEY=...,SUPABASE_URL=...,SUPABASE_SERVICE_KEY=..." `
  --memory 4Gi `
  --timeout 300 `
  --concurrency 10

Use 4Gi — fastembed loads ONNX models at startup and will OOM at 2Gi.

Environment Variables

Variable	Where	Description
`GROQ_API_KEY`	backend/.env	LLaMA 3.1 / 3.3 / 4 Scout via Groq (LLM + vision fallback)
`GROQ_MODEL`	backend/.env	Default fast model: `llama-3.1-8b-instant`
`GEMINI_API_KEY`	backend/.env	Vision AI — blueprints, safety photos, video walkthroughs
`SUPABASE_URL`	backend/.env	Your Supabase project URL
`SUPABASE_SERVICE_KEY`	backend/.env	Service role key (bypasses RLS)
`COHERE_API_KEY`	backend/.env	Optional — enables Cohere Rerank for RAG precision
`LANGCHAIN_API_KEY`	backend/.env	Optional — enables LangSmith tracing
`LANGCHAIN_TRACING_V2`	backend/.env	Set to `true` to activate LangSmith
`HUGGINGFACE_API_KEY`	backend/.env	CLIP · BLIP · BART · BERT-NER · RoBERTa via HF Inference API — required for HuggingFace provider in Settings
`VITE_API_URL`	frontend/.env.production	Backend URL (Cloud Run URL in prod)

Data Persistence Model

Feature	Storage	Fallback
RFIs, Change Orders, Obligations, Punch, Submittals, Daily Logs	Supabase	None (requires session)
Projects	Supabase + DEMO_PROJECTS constants	Demo always shown
Schedule Tasks	Supabase	Static INITIAL_TASKS
Leases, Tenants, Maintenance, NOI, CAM, PM Daily Logs	Supabase	Demo data if empty
Unit Turnover	Supabase (auto-save on toggle, debounced 800ms)	None
Approvals	Supabase	None
User Settings	Supabase + localStorage	localStorage only if Supabase unreachable
FAISS Vector Index	In-memory per session	Rebuild on next upload
Session State	localStorage (frontend) + in-memory (backend)	New session on refresh

Known Limitations

Scanned PDFs need OCR pre-processing — ilovepdf.com works well
FAISS index is in-memory — redeployment clears active sessions (documents need re-upload)
PGVector RAG persists across restarts; FAISS legacy Q&A does not
Large PDFs (100+ pages) take longer on first embed
Gemini free tier hits daily quota quickly — video analysis and blueprint CV automatically fall back to Groq vision (frame extraction)
fastembed ONNX models load at cold start — Cloud Run needs 4Gi memory minimum
MCP session persistence — each Claude Desktop conversation needs an active session ID; sessions reset on backend restart
RAGAS eval requires documents to be ingested first (/rag/ingest) — run on a session with real documents for meaningful scores

Built with FastAPI · React · Supabase · LangGraph · FAISS · Groq · Gemini · fastembed · RAGAS · MLflow · DVC · Playwright · Firebase · Cloud Run · MCP

Name		Name	Last commit message	Last commit date
Latest commit History 21 Commits
backend		backend
frontend		frontend
.gitignore		.gitignore
README-MCP.md		README-MCP.md
README.md		README.md
architecture-design.html		architecture-design.html
architecture.html		architecture.html

Folders and files

Latest commit

History

Repository files navigation

◈ PRISM

AI Construction & Property Management Intelligence Platform

What PRISM Does

Platform Overview

Screenshots

Features

Construction — Document Intelligence

Construction — Contract & Compliance

Construction — Project Administration (AI-augmented)

Construction — Cost & Schedule Intelligence

Construction — Safety & Field

Property Management — Lease Abstraction

Property Management — Tenant Risk Scoring

Property Management — NOI Forecaster

Property Management — Maintenance Orders

Property Management — CAM Reconciliation

Property Management — Unit Turnover

Human Guardrails — Approval Queue

AI Copilot — Streaming Multi-Agent Chat

Guardrails — Safe AI at Every Layer

RAG Pipeline — Document-Grounded Answers

Evaluation & Observability

MCP Integration — Talk to Your Project in Claude Desktop

Tech Stack

Project Structure

Supabase Schema

Construction Tables

Property Management Tables

API Reference

Session

Documents & Q&A

AI Analysis — Construction

RFI Register

Change Orders

Contract Obligations

Punch List

Submittals

Daily Log — Construction

Projects

Scheduling

Workforce

Property Management — Leases

Property Management — Tenants

Property Management — Maintenance

Property Management — NOI

Property Management — CAM

Property Management — Daily Log

Property Management — Unit Turnover

Approvals (Human Guardrails)

User Settings

AI Copilot — Multi-Agent

RAG — Document Ingestion & Search

Evaluation & Feedback

Testing

E2E Tests (Playwright)

Eval Pipeline (MLflow + DVC)

MCP Setup

Demo Access

Local Development

Prerequisites

1. Run the Supabase schema

2. Backend

3. Frontend

Deployment

Frontend — Firebase Hosting

Backend — Google Cloud Run

Environment Variables

Data Persistence Model

Known Limitations

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages