Vague VenturesHuman sign-off
Nothing counts until your expert approves it. The AI proposes and drafts; a person with authority decides. No output reaches a client, a court, or the market unreviewed.
Vague Ventures AI Services Est. 2022
AI on your highest-stakes work, verified by your own experts.
Legal review, trading decisions, gas dispatch, clinical interpretation. We design and run the AI; your experts verify every output before it counts. A commercial, entrepreneurial practice built on nearly two decades in energy, industry, and finance.
Cologne · 50.94°N 6.96°E
GDPR-compliant
Human-verified output
Append-only audit trail
01 / Services
AI for work where mistakes are expensive.
Your partners, clinicians, and traders are expensive experts. Turning them into AI engineers is the slowest and most costly way to get the leverage. So we run the AI for you, on your highest-stakes, highest-volume work, with every output verified by a human before it counts. Your experts stay experts.
We deliver three ways: consulting alongside your team, embedded execution with it, or a fully delegated unit we organise and run for you.
Client names are withheld under privilege and NDA. References available in conversation. Each row is labelled by stage: live service, designed concept, or focus area.
01
Legal · live service
→
Legal review as a service
For a German Wirtschafts- und Gesellschaftsrechtskanzlei, an AI workflow that reviews, drafts, and adversarially stress-tests filings to a large-firm baseline. In a recent cross-border dispute it reconciled a large body of individual claims across the entire evidence corpus, at a volume no team could sustain by hand. Every finding is cross-checked against the record, and the responsible lawyer signs off before anything leaves the firm.
02
Media · live service
→
Full-lifecycle content, marketing & sales
A production system built for World Radio Switzerland, a Cedra Ventures portfolio company, that runs the whole content lifecycle. It curates a dozen multilingual sources, drafts on-air bulletins in the station's voice, spins out branded social campaigns, and feeds audience analytics back into what gets produced, marketed, and sold. A moderator clears every item before it airs or ships.
03
Trading · designed concept
→
Decision partner for energy trading
A conversational "second brain" for senior gas-trading leadership that has read every contract on the book and reasons across the full supply lifecycle, from decade-long pipeline imports and LNG agreements down to spot cargoes. A quant face computes over a signed-off schema of contract terms, so the highest-stakes numbers are arithmetic, not a guess. It drafts the board paper and never sits in the execution path.
04
Energy · designed concept
→
Verified dispatch for gas networks
A decision-and-assurance system for the 24/7 gas-transmission control room. For every candidate action (routing, compressor scheduling, a nomination) it simulates that action through a transient hydraulic model and admits only the options that verifiably respect every hard physical and network-code limit. Contained ML proposers stay behind a deterministic verifier that has the last word; every check writes to an append-only audit record.
05
Health · service + product
→
Clinical data & AI interpretation
Structured clinical intake and AI-assisted interpretation that surfaces actionable insight for surgical teams across their caseloads. It is grounded in regional clinical guidelines, with role-based access and a strict data boundary, so the tool informs the clinician and never stands in for clinical judgement. Offered as a service today, with a standalone product in development.
06
Energy · focus area
→
Data centres as an energy system
Treating a data centre as energy infrastructure makes the real constraint on the AI build-out clear: not chips, but megawatts, cooling, and where a site is allowed to land. We separate the durable layer (power, cooling, the shell) from the fast-refreshing compute inside it, and design the first to outlast several generations of the second. A growing focus of the practice across Europe and the Middle East.
01.1 / Advisory & governance
We work as entrepreneurs, not analysts.
The approach is commercial first: technology is aimed squarely at the decisions that make or lose money. Strategy, software, and capital are the means to a sharper commercial call, never the point in themselves.
A
Strategy & governance
Governance of complex, multi-stakeholder programmes across digital, energy, and security, raising synergies while protecting the accountability of each plant, unit, or asset.
B
Software & digital strategy
Structuring and implementing software strategies for industrial estates. A Siemens partner by experience, deliberately vendor-agnostic: the right solution at the right time.
C
Transaction & commercialisation
License-model economics, DCF and cost-parity analysis, vendor negotiation, and deal structuring, from software estates to single-asset acquisition vehicles.
D
LLM training & enablement
Teaching teams to use large language models well: where a workflow must be deterministic, structured, and verifiable, and where it can be probabilistic and generative. Built on original teaching material: a ten-lesson interactive lecture from tokens to training, and a decision framework that places each task by how specifiable it is and how costly an error would be.
02 / Data handling
Built for privileged, high-consequence data.
If we are touching your filings, your clinical records, or your book, the handling matters as much as the output. Six commitments hold on every engagement.
A full audit trail
Every recommendation, model call, and check is written to an append-only record, so any output can be reconstructed for review, handover, or a regulator.
Your data stays yours
Client material is not used to train foundation models. It is processed under GDPR, and the models, subprocessors, and where processing runs are fixed in writing, then returned or deleted on the terms we agree up front.
A scoped engagement
We start with a defined pilot and a clear exit, not an open-ended dependency. You keep the leverage whether or not you keep us.
Deployment on your terms
Models and data run where you require: our cloud, your private cloud, or on-premise, inside your own network boundary. Where processing happens is agreed and fixed, not assumed.
Least-privilege access
Role-based, need-to-know access to every workflow. Who can see and do what is set up front, and written to the same audit record as every model call.
03 / About
An AI-services practice, principal-led from Cologne.
Vague Ventures is led by Sahand Hagi, its Managing Director. The practice sits deliberately at the seams between fields that usually do not talk to each other, the plant engineer and the fund, the simulation vendor and the operator, and does the translation that turns technical fact into a commercial decision.
It is Sahand Hagi's own practice: the vehicle for nearly two decades of work at the intersection of energy, industrial technology, and finance, now pointed at putting frontier AI to work where the stakes are highest. His principal-investment and operating work runs separately, through Cedra Ventures, a Zurich syndicate of owner-operators where he is a founding partner. Read the full profile ↗
On the name
Vague is the French word for wave. The play on English vagueness is deliberate; the work is not.
français · /vaɡ/ · “wave” · the work is exact, verified, accountable
04 / Case study
A digital twin that made a safer plant computable.
Vague Ventures helped bring Linde Engineering and Siemens Digital Industries Software together to predict the hydrodynamic loads that could tear piping off its rack, fast enough to design the plants of the future. Our role was advisory and connective: framing the problem, bringing the two organisations together, and scoping the benchmark. The engineering results below are Linde's and Siemens', enabled by that collaboration.
10,000+
interacting hydraulic components modelled
~5,000 N
transient force from a single small disturbance
~8×
simulation runtime reduction achieved
4×
plant scale the method is built to absorb
The problem
In a chemical plant, a fluid network can hold more than 10,000 interacting components across many recirculating loops. A sudden event sends a pressure surge through it. Because surge is a momentum phenomenon, small errors in velocity or fitting losses make the whole analysis unreliable. Get it wrong and a pipe can leave its rack bridge. Around five such incidents are known over ten years at the Frankfurt Höchst industrial park alone.
The approach
The workflow builds an "as designed" digital twin, then uses steady-state balancing in Siemens Simcenter Flomaster to reach the "as operated" state, the network exactly as it sits the instant before a disturbance. One click launches the transient. It solves mass, momentum and energy conservation and returns the force on every pipe, fitting and valve. Even a small disturbance produced forces on the order of 5,000 N.
The result
Benchmark runtime cut roughly 8×: about four days down to half a day.
Next-generation plants are expected to be four times today's size, so raw speed became the constraint. Parallelisation alone could not help: 1D/0D physics does not parallelise well. The Simcenter Flomaster teams borrowed techniques from the gaming industry and paired them with leaner modelling. The heaviest deluge-priming benchmark fell from about four days to about half a day, with no loss of accuracy.
Source: figures as reported in the Linde × Siemens white paper.
Full white paper available on request.
05 / Products
Platforms we are building.
Where a service hardens into something repeatable, it becomes a product: the same verified, human-in-the-loop discipline, packaged.
P1
EdTech · learning science
→
Adaptive multi-subject learning
An adaptive tutoring engine that turns authored curriculum into browser and print exercises, governed by machine-checkable Science-of-Reading pedagogy and a computerized-adaptive difficulty model. A live corpus of about 3,000 exercises across English and maths, plus a reading-intervention module.
P2
Health · AI · in development
→
Clinical interpretation platform
The clinical intake-and-interpretation capability above, packaged into a standalone web and mobile product, guideline-grounded, with role-based access and a strict data boundary.
06 / Contact
Start with one paragraph.
Email a short description of the decision or workflow you want to sharpen. You will get a reply within one business day and, if there is a fit, a scoped pilot proposal, fixed in time and cost.
The person behind the practice
Sahand Hagi's work in business, industry, and politics, and the publications behind it.