Founder profile
AI Solutions Lead & Forward Deployment Engineer · Founder, Quantamix Solutions B.V.
I spent eighteen years proving that risk models did what they claimed, inside banks where a regulator would eventually ask. I now build AI systems and hold them to the same standard.
Market risk, model validation and capital calculation at the Reserve Bank of India, De Nederlandsche Bank, ING, EY, Rabobank and ASN Bank. FRTB, IRRBB, IFRS9, A-IRB and EBA stress testing, under ECB and DNB supervision. You learn quickly there that a model nobody can explain is a model nobody can use, however good the backtest looks.
Ten systems in production. A governance SDK on PyPI and the VS Code Marketplace. GraphRAG copywriting running inside Amazon Ring. A document-intelligence platform in Frankfurt that keeps its data in the EU. A seven-agent CRM delivered under contract in Italy. The common thread is a knowledge graph underneath, and a record of how each answer was reached.
Three European patent applications and four papers, with the benchmarks written down: constrained F1 of 0.756 against an ungoverned baseline of 0.328 on MultiGov-30. The claim is deliberately narrow — that a governed graph of agents beats an ungoverned one, by this much, on this benchmark. Anyone can check it, which is the point.
Research
Three European patent applications and four papers on graph-based AI governance. Each one states what was measured, on which benchmark, against what baseline — so the claim can be checked rather than taken on trust.
Trust-Aware Multi-Signal Document Retrieval with Graph-Based Compliance Scoring
Governed Graph-of-Agents Reasoning over Knowledge-Graph Topologies with Semantic SHACL Validation
Predictive Subgraph Generation, Confidence-Gated Graph Write-Back and Structured Reasoning
Production Lessons toward the PCT Open Standard: Batch-Mode Cryptographic Commitment
DRACE — the development-domain governance analogue of TRACE
Full filing detail sits on the patent notice page.
Experience
Act one built the regulatory instinct: eighteen years where an unexplainable model was an unusable one. Act two applies it to systems that have to answer the same question about themselves.
Products
Each one shipped, with the scale and the stack stated. Where a figure comes from a client engagement rather than an instrumented measurement, the wording says so.
AI dev-intelligence + governance SDK
Turns any codebase into a governed knowledge graph so AI agents reason over architecture instead of re-reading files each session.
AI governance / audit trail
The documents and documentation path for governance and audit trails — the enforcement layer for written evidence.
AI adoption intelligence (SaaS)
Captures AI-tool usage, maps it to a 95-friction / 8-layer taxonomy, and runs cascade-propagation analysis with forecasting and causal inference.
GenAI marketing (GraphRAG)
AI copywriting and campaign optimisation grounded in a brand knowledge graph, with agentic usage-rights automation.
Regulated document intelligence
GDPR and EU-AI-Act-aligned document intelligence with confidence-gated human review and a tamper-evident audit trail.
Brand-intelligence AI (SaaS)
Extracts a brand into a reusable brand-memory knowledge graph once, then generates against it.
Graph-native local inference
A GraQle plugin that lets a local model on Ollama approach cloud-model accuracy by routing through the graph.
Vertical CRM + AI
A CRM with a seven-agent AI reasoning layer for real-estate and hospitality investment advisory, delivered under contract in Italy.
AI market-intelligence + graph
The CrawlQ-brand distribution of the graph substrate, with a market-intelligence surface on top.
Public position
Between May and June 2026 I ran a public working thread on what an EU AI Act audit trail actually has to contain. It was not a marketing campaign. Senior practitioners disagreed with me in the open, several of them corrected me, and the resulting pillar page names eleven contributors because the argument was theirs as much as mine.
No product is “EU AI Act compliant”. Compliance is a determination a regulator makes about a deployed system in context. A substrate can be aligned with named articles, and that is the strongest honest claim available.
The substrate proves a replayable path. It does not gate or authorise an action. Conflating the two turns an evidence layer into an execution control, which is a different product with different failure modes.
GraQle governs the reasoning path through code; TraceGov governs the documentation path through written evidence. Same vocabulary and IP, two enforcement surfaces — not one merged package and not two unrelated products.
Articles 14 and 26 ask who was accountable, not merely what happened. Seeing a system is not the same as binding a decision to a person, and that gap is a genuinely open problem.
In scope for the work: Articles 4, 12, 13, 14, 15, 25 and 50. Explicitly out of scope: Articles 5, 53, 55 and Annex VII.
The full argument, with every contributor named, is on the EU AI Act audit-trail stack pillar.
Capability
Grouped by domain rather than listed flat, because the combination is the point — the risk column is why the AI column is governable.
Education & credentials
MSc Risk Management — Honours
Duisenberg School of Finance
2010 — 2011
MSc Financial Risk Management
Vrije Universiteit Amsterdam
2010 — 2011
Masters
Public administration and policy
2002 — 2004
Questions
The questions people actually arrive with, answered plainly.
Harish Kumar is the founder of Quantamix Solutions B.V. in the Netherlands and works as an AI Solutions Lead and Forward Deployment Engineer. He has twenty-two years of experience across two fields: eighteen years of quantitative and regulatory risk inside tier-1 banks and two central banks, and the generative AI transformation work that followed. He holds three European patent applications and has published four papers on graph-based AI governance.
He builds AI systems that can prove what they did. That means knowledge-graph substrates, retrieval-augmented generation, multi-agent orchestration and the audit trail underneath them — deployed into regulated environments where an unexplainable decision is an unusable one. He has shipped ten production systems, including a governance SDK on PyPI and the VS Code Marketplace, GraphRAG copywriting inside Amazon Ring, and an EU-resident document-intelligence platform running in Frankfurt.
He spent eighteen years in market risk, model validation and regulatory capital. That covers the Reserve Bank of India, De Nederlandsche Bank, ING, EY, Rabobank and ASN Bank, working on FRTB, IRRBB, IFRS9, A-IRB, CRR and EBA stress testing under ECB and DNB supervision. At Rabobank he owned the IFRS9 calculation engine across a loan portfolio above €400 billion.
Three European patent applications — EP26162901.8 (TAMR+), EP26166054.2 (CogniGraph, a divisional with fifteen claims) and EP26167849.4 (PSE) — plus papers on SSRN and Zenodo. The headline published result is a constrained F1 of 0.756 on the MultiGov-30 benchmark against an ungoverned baseline of 0.328, at 99.4% governance accuracy.
He works on the audit-trail and governance substrate that EU AI Act obligations rest on, across Articles 4, 12, 13, 14, 15, 25 and 50. The careful wording matters: no product is “EU AI Act compliant”, because compliance is a determination a regulator makes about a deployed system in its context. A substrate can be aligned with named articles, and that is the strongest honest claim available.
By email at harish.kumar@quantamixsolutions.com, or by booking a diagnostic briefing through the Quantamix Solutions website. He is based in Uithoorn, North Holland, and works with organisations across the Netherlands and the wider EU. The published work is on Zenodo and SSRN, and the code is on GitHub.
Python and TypeScript day to day; AWS Bedrock, Neptune, Lambda, DynamoDB and Athena on the cloud side; Neo4j and Neo4j Aura for graph work; Azure OpenAI, Google Cloud, BigQuery and Looker across the other two clouds; Next.js and React on the front end. On the AI side: GraphRAG, retrieval-augmented generation, the Model Context Protocol, multi-agent orchestration, embeddings and model routing.
That is the whole of what I do. If the audit trail is the part keeping the project from shipping, it is worth a conversation.