Horn Software ArchitectsHorn Software Architects

Flagship Case Study — Document Intelligence

TenderMatch: Engineering a Nationwide Document Intelligence Pipeline

How HSA isolated a document intelligence bottleneck and built a continuous monitoring and extraction pipeline across 101 tender sources.

Summary

The Project in Brief

TenderMatch is a production document intelligence pipeline engineered by HSA. It continuously monitors 101 government and institutional tender sources across South Africa, capturing new publications within minutes to hours of release. A two-stage classify-then-extract pipeline converts raw, unstructured tender documents into structured, schema-validated records.

Monitored Nodes

101 Sources

Ingestion Time

Minutes to Hours

Active Pipeline

2,000+ Tenders Tracked

In the bidding industry, speed to qualification decides margin. Before TenderMatch, tender documents had to be found, read, and checked against statutory registration and compliance requirements by hand, one publication at a time.

That manual review created a document bottleneck. While one bid was being read, other new publications went unreviewed, creating a real risk of missed deadlines and qualification errors.

The Bottleneck

The Human Document Limitation Throttling Bid Capacity

Discovery

Isolating the Data Flow Friction

Before proposing any software, HSA mapped the bidding cycle end to end to find where time was actually being lost:

  • 1

    Observation: Tenders were being found and collected manually across dozens of separate government and institutional sites, with no consistent way to know what had just been published.

  • 2

    Measurement: Reviewers were manually searching each document for the same recurring compliance fields (CSD registration, tax status, local content requirements) scattered throughout the text.

  • 3

    Isolation: The true constraint was not the final bid writing stage; it was source monitoring and compliance-field extraction speed.

Architecture

The Ingestion and Validation Engine

A multi-stage pipeline designed for secure, low-latency processing of unstructured documentation.

1. Continuous Source Monitoring

All 101 sources are monitored continuously, so new publications are captured within minutes to hours of release rather than found by manual search.

2. Classify, then Extract

Each document runs through a two-stage pipeline: first classified, then processed for structured field extraction, rather than a single pass that has to do both at once.

3. Structured Records

Extracted fields are written into structured, schema-validated database records that the client's team can search and act on directly.

Challenges

Overcoming Real-World Noise

Challenge 1: Inconsistent Source Formats

101 sources means 101 slightly different ways of publishing a tender notice. HSA engineered the ingestion layer to normalize this variation rather than hand-writing a separate parser for every source.

Challenge 2: Preventing LLM Data Hallucinations

Relying purely on a language model to identify compliance parameters risks hallucinated fields (e.g. an invented deadline). We resolved this by validating extracted output against strict schemas and expected formats before it is written to the database, rather than trusting model output directly.

Throughput Scaled:

New tenders are now captured within minutes to hours of publication instead of being found by manual search, letting the client track over 2,000 tenders on an ongoing basis.

Consistent Compliance Checking:

Schema validation on every extracted record means compliance fields are checked the same way every time, instead of depending on which reviewer read a given document.

Outcomes

Measurable Operational Leverage

Lessons

Systems Thinking is Sector-Agnostic

TenderMatch is just one instance of a document and information constraint. The methodology is sector-agnostic: the exact same event-driven architecture and validation gates apply when a logistics firm struggles to parse customs documents, or a manufacturing company manually verifies accounts payable delivery lines. We target the information bottleneck first, then build the software that eliminates it.

Does your business have a similar constraint?

We do not build software to follow templates. We audit operational loops, isolate constraints, and build the custom systems that remove them.