PondoLens: AI red flags on public spending, explained for every citizen

Problem Statement

The Philippine government spends trillions of pesos every year, and a large share goes through public procurement under the Government Procurement Reform Act (RA 9184) and the New Government Procurement Act (RA 12009). The data is technically public. PhilGEPS publishes notices and awards, and COA posts annual audit reports for agencies and LGUs. In practice, the data is spread across portals, PDFs, and inconsistent formats, so very few citizens or even local councils can analyze it. Problems like splitting contracts to stay under bidding thresholds, the same supplier winning over and over, or items priced far above market are usually found only years later by COA or investigative journalists, after the money is already spent.

Proposed Solution

PondoLens is a web platform that ingests PhilGEPS award and bid data and parses COA annual audit report PDFs. A rules-plus-ML engine scores every contract and supplier for risk patterns:

• contract splitting (many similar contracts just under thresholds within a short period)
• repeat winners and low competition (single bidders, the same small group of bidders)
• price outliers versus similar items across agencies and the DBM-PS price list
• links to COA audit observations

An LLM layer turns each flag into a short, neutral explanation, for example: “Nine contracts for ‘office supplies’ worth ₱480k–₱499k were awarded to the same supplier within 30 days. This pattern can indicate splitting to avoid public bidding.” Every flag comes with source links. Users can search by LGU, agency, or supplier, follow their town, and ask questions in natural language (“Magkano ang ginastos ng aming bayan sa flood control noong 2024?”). PondoLens shows red flags, not accusations, and gives agencies a channel to respond to each flag.

Target Market

• Primary users: Citizens and civic groups, investigative journalists (e.g., PCIJ, Rappler, local press), and student researchers.
• B2G buyers: COA (to prioritize audits), the Office of the Ombudsman, DBM-PS, internal audit units of agencies, and sangguniang bayan/panlalawigan members who oversee about 1,700 LGUs.
• B2B/others: Compliance teams of suppliers and banks (due diligence), and multilateral donors funding local projects.

How AI Is Used

• Document AI and LLM extraction: COA reports are long PDFs with tables and narrative findings. The LLM pulls out structured observations (amount, issue type, agency, year). This can’t be done manually at national scale.
• Entity resolution: Supplier names show up as “ABC Const.”, “A.B.C. Construction Corp.” and so on. Embedding-based matching combines them, which is essential before repeat-winner analysis can work.
• Anomaly detection: Isolation Forest or statistical outlier models on unit prices and award patterns. Fixed rules alone miss new schemes.
• Semantic item matching: Embeddings compare differently worded items (“laptop i5 8GB” vs “notebook computer core i5”) so overpricing can be detected.
• LLM explainer and Q&A: Turns risk scores into citizen-friendly, cited explanations in Filipino or English. Without it, the tool remains a dashboard only analysts can read.

PLDT / Smart / ePLDT / Rokid Angle

• ePLDT: Big-data hosting and analytics in sovereign Philippine data centers. This is attractive to COA and the Ombudsman, who need data to stay in the country.
• PLDT Enterprise: A B2G transparency and analytics product to pair with government connectivity and cloud deals. Can be white-labeled for provincial governments.
• Smart: SMS or app alerts for citizens following their LGU (“New ₱12M road contract awarded in your town”).
• Rokid: Low relevance. A possible future use is AR site inspections where auditors see contract details over a physical project.

Demo Scenario (3 minutes for judges)

1. Open with a sample headline about a flood control project that turned out to be overpriced, and ask: “What if we could have seen this coming from public data?”
2. Type a sample municipality into PondoLens. A town scorecard loads showing total awards, competition level, and top suppliers.
3. A red flag appears: “Possible contract splitting – 9 contracts, same supplier, just below threshold.” Click it to see the timeline chart, PhilGEPS links, and the plain-Filipino explanation.
4. Show the price-outlier view: a sample item priced 3× the median of comparable purchases elsewhere.
5. Show a linked COA observation that was extracted automatically from a PDF, with the page cited.
6. Ask the chatbot: “Sino ang pinakamadalas manalo sa mga kontrata dito?” It answers with a chart and citations.
7. Close with: “Not accusations, just questions every taxpayer deserves to ask.”

30-Hour Build Plan

Suggested stack: Python (pandas, scikit-learn), PostgreSQL + pgvector, Gemini or other LLM for extraction and explanation, Unstructured/pdfplumber for PDFs, Streamlit or Next.js + Recharts. Public data / APIs: PhilGEPS open data and award notices, COA annual audit reports (coa.gov.ph), DBM Procurement Service price lists, data.gov.ph, Open Contracting Data Standard (OCDS) red-flag methodology, PSA OpenSTAT for population and income-class normalization.

Risks & Privacy Notes

• Defamation and false accusations: Use neutral “red flag” wording, show methodology, cite sources, give agencies a right to reply, and have humans review flags before publishing.
• Data quality: PhilGEPS fields can be incomplete. Show a data-completeness indicator and avoid flagging when data is too thin.
• Data Privacy Act 2012: Mostly public, institutional data. Individual names (e.g., sole proprietors, officials) are processed only as needed for transparency of public funds, and no unnecessary personal profiling is done.
• LLM hallucination: Extractions are checked against the source text, every answer includes the page or record cited, and text-to-SQL runs read-only.
• Clause 24 / competition rules: Only public government data is used. No confidential data goes into AI tools.

AI Use Disclosure (draft)

PondoLens uses an LLM (e.g., Google Gemini) to extract structured findings from COA PDFs, generate plain-language explanations, and answer questions in natural language. It uses sentence-transformer embeddings for supplier and item matching, and scikit-learn for anomaly detection. AI coding assistants helped with data-cleaning scripts and UI scaffolding. The risk rules, thresholds, and final wording of flags were designed and reviewed by the team, and no flag is presented as a finding of wrongdoing.

Rubric Scoring