Kriyalytics (2025–)
A multi-agent vendor rationalisation system: 7,000 → 5,000 vendors in the pilot
7,000 → 5,000vendors · early pilot result
A propose-and-check multi-agent system that consolidates vendor options without breaking the contracts the business already holds.
My role: Architect and delivery lead; led the team that built the pipeline.
Executive Context#
A global materials-science manufacturer had grown to approximately 7,000 vendors covering more than 3 million SKUs. The buying habit was simple: for each request, ask vendors for codes and quotes, then finalise a contract. Each request added another vendor and another contract, so many suppliers now competed on near-identical products. And because every request opened a fresh contract, the volume discounts stayed thin.
The programme goal was $2.5M in savings. At this scale, any saving above about $50K a month is material. Procurement leadership drove the work with finance backing.
The Actual Problem#
The brief asked for fewer vendors. The real problem was visibility: equivalent materials sat under different catalogue names, so no one could see which suppliers were genuinely interchangeable, or which contracts made a swap impossible.
Diagnostic Approach#
- AI Necessity: Catalogue research and specification matching were prerequisites for optimisation; procurement needed checked options rather than an autonomous purchasing decision.
- Data Sufficiency: Supplier-submitted catalogues, websites and PDFs had to be validated before treating their stated coverage as available supply.
- Methodological Fit: Specialised agents handled normalisation, research and validation. A tool-calling optimisation agent used a purpose-built optimisation MCP server.
- Infrastructure Readiness: The pipeline ran on Databricks with APIs in a constrained AWS environment.
Strategic Intervention#
We built a pipeline of five agents: normalisation, vendor research, validation, optimisation and sense-checking. The optimisation and sense-check agents iterated on candidate options.
The first proof of value arrived in about three weeks, when the proof of concept identified the common vendors supplying the same product. The full pilot ran about six weeks.
The pilot also exposed a constraint that the original brief had missed: existing contracts could not simply be replaced. We added the sense-check agent to account for contract timing, delivery delays and missing data.
Decision Ownership#
The Procurement Manager and Head of Purchase finalised quotes from the recommended vendors. The system proposed and checked options; people made the purchasing decision.
Outcome#
| Metric | Before | Early pilot result | Context |
|---|---|---|---|
| Vendors | ~7,000 | ~5,000 | Existing contracts limit how fast this can fall |
| Monthly saving | Baseline | ~$100K/month | Against a $2.5M programme goal |
| Analyst time | Baseline | 15 analyst hours/week freed | Preliminary pilot figure |
| Time to first value | — | ~3 weeks | Proof of concept found the common vendors |
| Pilot length | — | ~6 weeks | Run on the pilot scope |
These are preliminary pilot results; the full outcome is still pending, and procurement retains the final decision. The $2.5M goal is not yet reached and the programme is still running. Consolidation continues as contracts expire.
Strategic Takeaway#
An optimisation result has to respect the contracts under which the business actually operates. Agentic research helps prepare options; it does not remove procurement accountability.
Want to find the same kind of logic leak?
Start with a Clarity Call. We will look for the point where data, model choice, and operating decision stop matching.
Book a Clarity Call