Taqdeer

AI-powered construction estimation SaaS

A bilingual Arabic/English multi-tenant SaaS that turns construction RFPs into structured, priced Bills of Quantities, combining AI document analysis, semantic catalogue search, assembly-aware matching and deterministic pricing.

Role: Co-Founder & Full-Stack Engineer
Timeframe: 2025 – Present

Visit taqdeer-co.com ↗

The problem

Construction estimation requires engineers to interpret project documents, identify work items, match them against supplier or internal catalogues, and prepare structured quantity and cost estimates. The work involves complex terminology, multiple document formats, inconsistent descriptions, and scope lines that often map to several catalogue components at once.

The objective

Turn Arabic and English RFP documents (Excel, Word, PDF) into reviewable, priced Bills of Quantities, with AI doing the interpretation and matching, and the application keeping every financial calculation deterministic and auditable.

My role

Founded the product and built it end to end as the sole engineer.
Designed the system architecture, data model, AI pipeline and security model.
Built the FastAPI backend, the Next.js front end, and the catalogue ingestion layer.
Owned delivery: CI/CD, containerization, production deployment and monitoring.

Technical approach

Documents are parsed into structured work items through FastAPI and Anthropic Claude using structured outputs, so the model returns typed data rather than free text.
Work items are matched to a priced catalogue with pgvector semantic search on PostgreSQL, and every match carries a confidence score for human review.
An assembly-aware engine decomposes a single scope line into the multiple catalogue items an engineer would specify, returning quantities and confidence scores.
Pricing is deliberately excluded from model context. Costs, markups and totals are computed in application code so financial logic stays deterministic.
Supplier rate sheets are normalised into one internal schema: item codes, bilingual descriptions, units, material and labour cost, markup and brand.
Multi-tenancy is enforced in the backend with role- and capability-based access control, JWT authentication with TOTP 2FA, scoped integration API keys, and per-tenant usage metering and billing.

Key features

AI document analysis

Excel, Word and PDF RFPs processed through FastAPI and Claude with structured outputs.

Semantic catalogue search

PostgreSQL + pgvector retrieve relevant catalogue items by semantic similarity, with confidence-scored matches.

Assembly-aware matching

One scope line decomposed into the several catalogue items it actually requires, with quantities.

Deterministic pricing

Costs, markups and totals stay in application logic, never delegated to the model.

Catalogue ingestion

Supplier rate sheets normalised into a common bilingual schema.

Secure multi-tenancy

RBAC with capabilities, JWT + TOTP 2FA, scoped API keys, per-tenant metering and billing.

Bilingual, RTL-aware product

RFP intake, Kanban project pipeline, BoQ review and approval, and financial dashboards.

Production delivery

Docker, GitHub Actions CI/CD, Vercel and Render deployment, and monitoring.

Interactive demo

SIMULATIONRuns entirely in your browser. Scope lines, catalogue, match scores and prices are fictional; the real product uses live models and customer catalogues.

Sample scope line

RFP language

Confidence threshold for review0.75Markup12%

Apply 15% VAT

1
Document input
Done
Supply and build 200 mm hollow concrete blockwork walls, plastered both sides, as per drawings. Quantity: 120 m2
Language: EnglishQuantity: 120Unit: m²

AI interpretation

Done

Schema-bound structured output (typed fields, not free text). No prices in the model context.

{
  "work_type": "masonry",
  "element": "wall",
  "material": "hollow concrete block",
  "thickness_mm": 200,
  "finish": "cement plaster",
  "finish_faces": 2,
  "quantity": 120,
  "unit": "m²"
}

3
Semantic retrieval
Done
Top catalogue candidates by vector similarity (pgvector). Candidates below the threshold are flagged for review.
- MS-201
  Hollow concrete block wall, 200 mm
  0.94
- PL-110
  Cement plaster, 15 mm
  0.88
- MS-205
  Hollow concrete block wall, 150 mm
  0.81
- MR-020
  Masonry mortar
  0.77
- PT-301
  Emulsion paint, two coats
  0.52

Assembly-aware matching

Done

One scope line decomposed into the catalogue items an engineer would specify.

Item	Why	Qty	Confidence
MS-201Hollow concrete block wall, 200 mm	Primary element	120.0 m²	0.94
PL-110Cement plaster, 15 mm	Plaster on both faces	240.0 m²	0.88
MR-020Masonry mortar	Mortar per m² of wall	2.16 m³	0.77

5
Deterministic pricing
Done
Rates come from the catalogue and every total is computed in application code, so results are repeatable and auditable.
Code Qty Rate (SAR) Amount (SAR)
MS-201 120.0 m² 60.00 7,200.00
PL-110 240.0 m² 23.00 5,520.00
MR-020 2.16 m³ 260.00 561.60
Subtotal 13,281.60
Markup 12% 1,593.79
VAT 15% 2,231.31
Total (SAR) 17,106.70
All lines above the review threshold.

Code	Qty	Rate (SAR)	Amount (SAR)
MS-201	120.0 m²	60.00	7,200.00
PL-110	240.0 m²	23.00	5,520.00
MR-020	2.16 m³	260.00	561.60
Subtotal	13,281.60
Markup 12%	1,593.79
VAT 15%	2,231.31
Total (SAR)	17,106.70

Outcome

A production-deployed, bilingual AI SaaS built and shipped by one engineer.
Demonstrates end-to-end AI engineering, semantic retrieval, relational data architecture, secure backend development, bilingual front-end engineering and production deployment.

Source code and customer data are private. Diagrams on this page are original illustrations of the architecture.

Next case studyEnterprise BI & Digital Transformation →Let's connect