Inspiration
Tender evaluation is a time-consuming process. Procurement teams working on construction and development projects may need to review 3–15 bids, each containing hundreds of pages of technical specifications, pricing tables, annexes and contractual conditions.
Comparing these documents manually takes considerable time and creates opportunities to overlook missing requirements, pricing inconsistencies and potential risk indicators.
I wanted to build a workflow that makes this process more structured, traceable and efficient, while keeping confidential bid documents within the user's NAS environment rather than requiring a separate cloud procurement platform.
What it does
Tender Bid Evaluator is an AI-powered procurement evaluation Skill designed for UGREEN NAS and Uliya.
It transforms a folder of tender documents into a structured evaluation report containing:
- Compliance matrix: Checks each bidder against mandatory and scored requirements extracted from the original tender documentation.
- Evidence traceability: Every extracted fact and compliance finding references its original document and page.
- Price normalization: Compares prices on a consistent VAT basis, identifies arithmetic inconsistencies and separates fixed prices from provisional sums and optional items.
- Risk indicators: Screens for potentially abnormal pricing, shared contact details, identical errors, document similarities and other cross-bid patterns.
- Clarification questions: Drafts bidder-specific questions, including requests to explain unusually low prices.
- Re-evaluation: When a bidder submits a revised offer, the Skill can re-analyze that bidder and identify changes against the previous report.
The Skill does not select a winning bidder or accuse companies of collusion. Its findings are advisory and leave the final decision to the procurement committee.
How I built it
The project is implemented as a modular Uliya Skill using the registered UGREEN NAS capabilities.
It orchestrates three NAS experts:
- File Agent (
ug-filesystem): Inventories the tender folder and individual bidder folders without modifying source documents. - Document Agent (
ug-document): Extracts requirements, prices, contractual conditions and bidder information, with document-level evidence references. - Knowledge Agent (
ug-knowledge): Retrieves previous evaluation reports and saves approved reports.
A key architectural decision is to analyze every bidder independently. Each bidder-specific delegation receives access only to that bidder's folder. Cross-bid comparisons use the facts returned by these isolated analyses, not the original documents.
The workflow also includes explicit approval checkpoints before criteria are scored and before the final report is saved.
The implementation consists of SKILL.md and supporting reference files defining evaluation criteria, price checks, similarity indicators, report structure, disclosure and test scenarios.
Challenges and design decisions
One of the main challenges was designing a multi-step workflow where every operation depends on verified results from the previous step.
I focused on several safeguards:
- Preventing information from one bidder's documents from entering another bidder's analysis.
- Distinguishing verified facts, missing evidence and AI-generated interpretation.
- Normalizing prices without silently including optional items or provisional sums.
- Treating document similarities and price anomalies as indicators rather than proof of misconduct.
- Preventing reports from being saved before user approval.
Another important challenge was working within the actual Uliya capability catalog rather than assuming that arbitrary scripts, external APIs or custom integrations would be available.
What I learned
This project helped me understand how to design AI agent workflows around real operational constraints rather than simply prompting an LLM to summarize documents.
In particular, I learned the importance of capability boundaries, structured delegation, evidence traceability, isolation of confidential information and explicit approval controls.
Verification and limitations
The Skill passed the Uliya Skill Kit static checker with zero errors and zero warnings, using capability catalog version 2.2.0.
Runtime execution on a physical UGREEN NAS has not yet been performed. The included synthetic test scenarios cover compliance extraction, price normalization, similarity indicators, unreadable documents, bidder isolation and approval handling.
The main remaining validation work is to test document extraction, page-level citation accuracy and the complete workflow on the actual NAS environment.
Tender Bid Evaluator is an advisory decision-support tool, not legal advice or an automated procurement decision-maker.
Built With
- analysis
- automation
- bid
- compliance
- evaluation
- nas
- processing
- procurement
- risk
- ugreen
- uliya
- workflow