Twenty years of risk & data. Now shipping agentic AI.
Engels Consulting is led by Koen Engels — econometrician, engineer and investment manager, former founder and Managing Director EMEA at Aon PathWise, and CEO of EunoriQ. We help financial institutions and scale-ups move AI from slideware to production, with the controls a regulator would expect.
Where we add the most value.
AI strategy to production
From use case and business case to a running, monitored system — including the governance around it.
Agentic delivery
Setting up multi-agent development workflows (Claude, Codex, open models) that teams can trust and afford.
Risk & quantitative modelling
ALM, scenario generation, solvency and IFRS 9/17 models, validation and documentation.
Financial projections
P&L scenarios and decision support for boards, investors and M&A.
Built inside financial services.
The experience of our managing partner, Koen Engels, behind every engagement.
Founder & Senior AI / Data Scientist — EunoriQ
Designed and built Triangulate on Google Cloud with an agentic setup across Claude, OpenAI Codex and Hugging Face models with MCP. Built a systematic equities research platform (tree ensembles, neural nets, GLMs, econometrics) alongside the ADIA Lab structural-break challenge. Advised nexxbiz.io on AI for system migrations.
Managing Director EMEA — Aon PathWise
Founded and led the EMEA business unit for GPU-accelerated, cloud-based risk software. Owned P&L, strategy and partnerships; C-suite engagement; regular dialogue with AWS, Azure and NVIDIA. Delivered LLMs for document libraries and ML clustering.
Risk Modelling & Audit — PwC
Large-scale engagements for major Dutch insurers, DNB and cross-border M&A. Board advice on risk frameworks; Solvency II, IFRS 9/17; ML-based fraud and anomaly detection.
Risk Consulting & Software — Willis Towers Watson
Around 20 banks, pension funds and insurers. Key role in the ING / NN demerger; ESG, asset projections and ALM; research for the Dutch Society of Actuaries.
Achmea · Saxo Bank
Group-level longevity research and balance-sheet interest-rate hedging; client support for investment products and automated trading.
MIT · INSEAD · Stanford · VU · Delft
MIT AI Executive Academy; INSEAD Advanced Corporate Finance (M&A); Stanford Machine Learning & Deep Learning; MSc Econometrics and MSc Investment Management (VU); BSc Applied Physics (Delft).
What building with AI agents actually teaches you.
Field notes from running multi-agent development across our own products — the things we now set up on day one for clients.
Agents need a written memory
Chat history is not a system of record. Every project keeps an AGENTS.md and a handoff document, so any agent — or human — can resume after a context reset without re-reading thousands of transcripts.
One owner per resource
Parallel agents collide on ports, locks and long-running jobs. Explicit ownership — one server, one heavy model task, lock files with real process identities — prevents silent duplication.
Judgement is the merge gate
Agents may open and even auto-merge PRs, but only after a written decision on the ticket: what we see, reasoning, decision, follow-up. Green CI is necessary, not sufficient.
Context is a budget
Token cost scales with what agents re-read. Line budgets on large files, concise handoffs and targeted retrieval cut spend more than switching to a cheaper model — and stored agent histories can quietly grow to tens of gigabytes.
Match the model to the task
Frontier models for reasoning and code; small local models for OCR and classification where privacy, latency or cost matter. No vendor gets a default seat in the data path.
Receipts over memory
Every dataset and model run carries a manifest: source hashes, inputs, freshness and the commit that produced it. If you can’t trace an output to its inputs, you can’t defend it to a regulator.
Finished is not the same as good
A model run completing without errors says nothing about whether the model is any good. Quality needs its own checks: repeat runs with different random starts, and validation on periods the model has never seen.
Separate “we broke it” from “the world changed”
Every change is tested against saved copies of real examples, so a failure means our code broke. Separately, a nightly check runs against the live sources to spot when the outside world changes. Mixing the two teaches teams to ignore failing tests.
Agents never see secrets
Secret files such as .env and local data are git-ignored from the first commit, so an agent can never push them. Add isolated cloud configuration per process, no tokens in logs and read-only credentials by default: agents get the minimum to do the job.
Start every prompt the same way — and pay less.
Model providers such as OpenAI (including Codex) and Anthropic cache the beginning of your prompts. When a new request starts with exactly the same text as a recent one — at least about 1,000 tokens, roughly 750 words — that repeated part is billed at a significant discount and answered faster. So put fixed instructions, rules and tool definitions first, and the changing question last. One timestamp at the top is enough to lose the discount. Even better: automate your agent context so it always does this. Read OpenAI’s guide →
Let’s make your AI programme production-grade.
Advisory, interim leadership or hands-on delivery.