Sustify

The Vision

Corporate sustainability commitments are everywhere, but finding verified sustainable vendors remains a manual, time-consuming process. Procurement teams spend weeks researching vendors, validating sustainability claims, and matching requirements to capabilities.

Sustify was founded to solve this problem using AI-powered matching. The core idea: use transformer-based embeddings to automatically match business sustainability requirements with verified vendor capabilities.

The Problem Space

Why Vendor Matching is Hard

Unlike traditional B2B marketplaces where requirements are structured (e.g., "I need 1000 units of Product X"), sustainability requirements are nuanced:

• "We need carbon-neutral shipping for our supply chain"
• "Looking for packaging suppliers using 100% recycled materials"
• "Need data center vendors with renewable energy commitments"

These requirements don't fit neatly into dropdown menus or category filters. They require semantic understanding of both what the business needs and what vendors offer.

The Verification Challenge

Sustainability claims are easy to make, hard to verify. "Greenwashing" is rampant. Businesses need confidence that vendor sustainability claims are legitimate.

Technical Approach

Transformer-Based Matching

The core matching engine used transformer-based embeddings (BERT-family models and OpenAI embeddings) to encode both:

• Requirement embeddings: What the business needs (free-text descriptions)
• Capability embeddings: What vendors offer (extracted from vendor profiles, certifications, case studies)

Matching happened in embedding space using cosine similarity, allowing semantic matches even when exact terminology differed.

Why Transformers?

Sustainability language is domain-specific. "Carbon neutral," "net zero," "carbon negative" have nuanced differences. BERT-based models fine-tuned on sustainability documents could capture these distinctions.

I experimented with:

• DistilBERT: Fast, lightweight, good baseline
• RoBERTa: Better performance on domain-specific terminology
• OpenAI text-embedding-ada-002: Excellent zero-shot performance without fine-tuning

Final production system used a hybrid: OpenAI embeddings for general matching + fine-tuned RoBERTa for sustainability-specific scoring.

Vendor Recommendation Pipeline

The recommendation flow:

1. Requirement Parsing: Extract key sustainability criteria from free-text requirements
2. Embedding Generation: Convert requirements to dense vectors
3. Candidate Retrieval: Approximate nearest neighbor search across vendor database
4. Re-ranking: Cross-encoder re-ranking for top-K candidates
5. Verification Scoring: Boost vendors with verified certifications (ISO 14001, B Corp, etc.)
6. Presentation: Ranked list with explanations (why was this vendor matched?)

Verification System

To combat greenwashing, I built a multi-tier verification system:

• Tier 1: Self-Reported: Vendor claims without external verification
• Tier 2: Document-Verified: Vendors upload certification documents (automated OCR extraction of cert numbers + manual review)
• Tier 3: Third-Party Verified: Direct API integration with certification bodies (e.g., B Corp API)

Recommendations prioritized Tier 3 and Tier 2 vendors. Tier 1 vendors were clearly labeled as unverified.

Product Development

User Research

Before building, I conducted 20+ interviews with:

• Corporate sustainability managers (target buyers)
• Sustainable product vendors (supply side)
• Procurement teams (secondary buyers)

Key learnings:

• Buyers want validated vendors, not just search results
• Vendors struggle with discoverability—getting in front of the right buyers
• Certification verification is a huge pain point (manual, time-consuming)

MVP Features

The initial MVP focused on core value propositions:

1. Free-text requirement submission: Businesses describe what they need in plain language
2. AI-powered vendor matching: Transformer-based recommendations
3. Verification badges: Clear indicators of vendor verification status
4. Direct messaging: Connect buyers and vendors

Pilot Phase & Learnings

Early Pilot Interest

I secured pilot interest from 3 companies:

• Mid-size retail company seeking sustainable packaging vendors
• Tech company looking for carbon-neutral shipping partners
• Food & beverage company sourcing sustainable ingredient suppliers

The AI matching worked well—pilot users reported that recommendations were relevant and saved them significant research time.

The Pivot Decision

Despite positive pilot feedback, I made the difficult decision to pivot away from Sustify in early 2024. Key reasons:

1. Marketplace chicken-and-egg problem: Needed vendors to attract buyers, needed buyers to attract vendors. Slow, capital-intensive to solve.
2. Sales cycle length: Corporate procurement decisions take months. As a solo founder, I couldn't sustain long sales cycles without funding.
3. Verification scalability: Manual verification didn't scale. Automated verification required partnerships with certification bodies (complex, slow to establish).

The technical matching system worked. The business model timing and execution were the challenges.

Technical Lessons Learned

1. Embeddings are powerful for fuzzy matching: Transformer embeddings handled the semantic matching better than any keyword-based system could have.
2. Fine-tuning pays off in specialized domains: Domain-specific fine-tuning (sustainability language) improved precision by ~20% over zero-shot models.
3. Explainability matters: Users wanted to know why a vendor was recommended. I added feature highlighting (which parts of the vendor profile matched which requirements).
4. Data quality > algorithm sophistication: Clean, structured vendor data mattered more than cutting-edge models. Garbage in, garbage out.

Business Lessons Learned

1. Solve a painful problem, not just an interesting one: The matching problem was interesting technically, but the pain point wasn't urgent enough to drive fast adoption.
2. Marketplaces are hard: Two-sided marketplaces require solving two distinct customer acquisition problems simultaneously. Very challenging as a solo founder.
3. Timing matters: Corporate sustainability was gaining momentum, but procurement budgets and processes hadn't caught up yet.
4. Know when to pivot: Continuing to push a slow-growing startup as a solo founder while starting full-time work wasn't sustainable. Better to pivot and apply learnings elsewhere.

What I'd Do Differently

With hindsight, several things I'd change:

• Start with one side: Build value for vendors first (help them get discovered) before worrying about buyer acquisition. Solve the chicken-and-egg sequentially, not simultaneously.
• Narrow focus: "Sustainability" is too broad. Focusing on a specific vertical (e.g., sustainable packaging only) would have been more tractable.
• Automate verification from day one: Manual verification was a bottleneck. Should have prioritized API integrations with certification bodies earlier.

Where the Technology Went

While Sustify as a business pivoted, the matching technology wasn't wasted. The transformer-based matching approach I developed influenced my later work on OneSKU (hybrid retrieval for vendor catalogs) and Channel AI (semantic search over structured data).

Lessons about embedding-based search, verification systems, and handling fuzzy requirements carried forward into subsequent projects.