Frequently asked
The questions procurement, legal and IT ask first.
Answered plainly, so nobody waits three weeks for an email. Deeper answers link to the field note that covers them.
Security and data residency
What your security reviewer asks before anything else.
01Where does our data go, and can it stay in the EU?
Yes — every product deploys into an EU or US region of your choice. Extract and Edge also run fully on-premise, nothing trains shared models, and residency is contractual.
02Which parts of the system hold our data?
Four: model inference, the retrieval index, logs and telemetry, and any human review queue. We pin a region for each one and give you a data-flow diagram naming every processor.
03Are embeddings treated as personal data?
By default, yes. Embeddings are derived from your documents and can carry personal data, so we keep the index inside your tenancy and in your region unless your counsel documents a reason not to.
04Do you train on our data?
No. Your data is used for your systems only. Where we fine-tune or distil a model on your traces, that model is yours and stays in your environment.
05What do you log, and for how long?
Query, caller, retrieved sources, model and prompt versions, output and the human decision that followed. Retention is set per store during discovery — prompts, outputs, traces, index and evaluation sets are configured separately.
06Can this run without internet access?
Yes. Xelec Edge runs distilled models on Xelec-designed hardware, offline-first, syncing when a network appears. Extract also has a fully on-premise deployment.
07What security paperwork can you sign?
DPA with a sub-processor list, residency commitments, retention terms and a security questionnaire response. We support DPIA preparation and work to SOC 2 aligned practices.
Regulation and governance
EU AI Act, GDPR and the evidence auditors ask to see.
01How do you handle the EU AI Act and GDPR?
We classify each system against the Act's risk tiers and build the documentation, logging and human oversight that tier requires. For GDPR we support DPIA preparation, retention controls and a signed DPA. Engineering support, not legal advice.
02Is our system high-risk under the AI Act?
Usually not. Internal drafting help and search over your own documents typically sit in the minimal-risk band; recruitment, credit, essential-services and safety-component uses are where high-risk obligations bite. We run a two-day classification per system.
03Are we the provider or the deployer?
It depends on whether you place the system on the market or use one. The distinction changes which obligations are yours, so we settle it in discovery — part of your evidence then comes from your model vendor's documentation rather than your own code.
04How do you prevent hallucinated or wrong answers?
Enforced citations, refusal below a confidence threshold, golden datasets gating every prompt and model change, and drift monitoring after go-live. Where errors are costly, low-confidence cases route to a human by design.
05Can you prove what the system did six months ago?
Yes, if the audit trail was built in from the start — that is why we do it on day one rather than retro-fit it. Every answer can be replayed with the passages, versions and thresholds that produced it.
06Do you give legal advice?
No. We build what your counsel's classification calls for, and we tell you plainly when a question belongs with a lawyer rather than with us.
Pricing, pilots and timelines
What it costs, how long it takes, and what happens if it fails.
01What does a paid pilot cost, and what do we get?
Fixed scope, fixed fee, typically four to eight weeks. You get a working system for one use case, the evaluation harness and datasets, and the findings either way. The discovery sprint prices it precisely.
02How fast can we be in production?
Clean document workflows have gone live in around eleven days; assistants over large permissioned corpora take six to ten weeks. Custom hardware follows Xelec's board timelines. We tell you which bucket you are in during discovery.
03What happens after the pilot?
The production gate is agreed in writing before work starts. Pass it and we scale or hand off with runbooks; fail it and you keep the findings, datasets and a written recommendation.
04Why do so many AI pilots never reach production?
Almost never because the model was too weak. It is missing ownership, no measured baseline and no written gate — organisational gaps that are far cheaper to fix before a pilot than after one.
05Can you build a demo our sales team can run?
Yes. Presales demo builds are a deliberately narrow working slice with seeded data and guardrails, usually two to three weeks. It is a real system, not a click-through mockup, and it upgrades into a pilot if the deals land.
06Who owns the IP?
You do — work product, models trained on your data, prompts, pipelines and documentation. Our products stay licensed to you, and we never reuse your data for other clients.
07Do you work with teams outside the US and Europe?
Our delivery is built around US and European hours, regulation and procurement. We take engagements elsewhere selectively when the use case fits our products.
08How do we start?
Book a discovery call. If there is a fit, we scope a two-week discovery sprint or go straight to a fixed-fee paid pilot.
Models, cost and lock-in
Model choice, inference spend and owning your own weights.
01Are we locked into one model provider?
No. The architecture is model-agnostic and every deployment records cost and quality per workflow, so swapping is an evidence-based decision. Where economics demand it, we distil a smaller model you own.
02How do you cut our Claude or GPT bill by 60%?
By measuring first: we baseline cost and quality per workflow, then route each request to the cheapest model that still passes the quality gate, add semantic caching and prompt compression, and distil a small model where volume justifies it. The 60% figure is our target, proven or disproven on your own traffic during the pilot.
03What is the biggest source of waste in a large model bill?
Context bloat — whole documents sent where three retrieved passages would do. After that: uncached repeated prompts, frontier pricing on classification work, and non-urgent jobs run through real-time instead of batch endpoints.
04Why a custom small model instead of just calling GPT or Claude?
For one narrow, high-volume job a distilled domain model is usually cheaper, faster and more consistent — and you can host it in your region, on-prem or on a device. We keep frontier models for the broad, low-volume work and benchmark both before you commit.
05What do we need before you can distil a model for us?
A narrow task with a stable output contract, a few hundred expert-labelled evaluation cases, and logged traces from the workflow. With no traces we start by capturing them — including the reasoning, not just the answers.
06Why do answers get worse over time even without a release?
Hosted models shift under you, and your inputs shift too. We monitor input drift, output drift and human override rates, which is the earliest honest signal that something changed.
Hardware, edge and IoT
What Xelec's hardware capability adds, and what it costs.
01What does Xelec, the parent company, actually contribute?
Hardware capability most AI vendors lack: embedded design, sensor integration, board bring-up, thermal validation and manufacturing. When inference belongs at the point of work, we design the device and the model together.
02Do you take hardware work without any AI in it?
Yes. Xelec has run IoT, embedded and electronics product development as standalone engagements for years. AI is an option on those programmes, not a prerequisite.
03Who owns the schematics, PCB files and firmware?
You do. Schematics, layout, BOM, firmware source and test fixtures are handed over. Any reused internal library is licensed to you perpetually and named in the contract.
04Can you take a device from prototype to volume manufacturing?
Yes — DFM review, pilot build, test fixtures and certification preparation, then volume with your contract manufacturer or ours. Sourcing and obsolescence continue as a sustaining engagement.
05Do you build on NVIDIA Jetson?
Yes — Jetson Orin Nano and Jetson-class modules, with model optimisation and quantisation, carrier-board and thermal design, camera pipelines and signed fleet updates with rollback.
06Why does edge performance drop after deployment?
Usually thermal throttling in a sealed enclosure, or a camera pipeline copying frames off the GPU. We validate at maximum ambient in the real enclosure and log clock and thermal state as production telemetry.
07Can you build a digital twin of our site?
Yes — sensors and cameras feeding a live model of the place, with an event history you can replay and simulate against. We validate the twin against a hand-counted shift and publish the error before anyone simulates anything.
Working with us
How engagements run, and who you actually talk to.
01Who will we be working with?
The engineers who build it. Discovery calls are taken by someone who has shipped this class of system, not by an account manager, and you keep the same team through the pilot.
02How do you handle handover?
Source, infrastructure definitions, evaluation datasets, runbooks and a walkthrough with your team. You can run it without us; many clients keep us on a sustaining engagement because they prefer to.
03Can you work alongside our existing vendors and platform team?
Yes. We fit your cloud, identity and CI conventions rather than importing ours, and we document interfaces so another team can own the system later.
04What if you think our idea is wrong?
We say so on the call, with the reason. Turning down a badly scoped pilot is cheaper for both of us than delivering one.
05Can I ask a technical question before committing to anything?
Yes — mail gulshan@xelec.in. Questions about architecture, cost or residency get answered directly, and no demo is required first.
Next step
Question not answered here?
Send it. Security and legal questions get answered before any demo.