Field notes
How enterprise AI actually reaches production.
Ten long-form notes from our own engagements, with sources you can check. Written by Gulshan, founder of Xelec.
Edge
9 Sept 2026
Digital twins from IoT sensors and cameras: modelling a site you can replay
A twin is not a 3D render. It is a time-series model of a real place, accurate enough that replaying yesterday explains what went wrong.
Engineering
10 min read
Reliability
27 Aug 2026
Evaluating AI you cannot fully test: golden sets, guardrails and drift
You cannot enumerate the inputs. You can still have a regression suite, a release gate and an alarm that fires before your users complain.
Engineering
9 min read
Edge
6 Aug 2026
NVIDIA Jetson Orin Nano builds: optimisation, thermals, camera pipelines and fleet updates
Getting a model onto a Jetson is an afternoon. Getting a hundred of them to run for three years in a warm cabinet is the engineering.
Engineering
10 min read
Reliability
2 Jul 2026
Retrieval that survives production: chunking, ranking, citations and audit trails
A retrieval demo needs good answers. A retrieval system needs permissions, provenance, refusals and a way to prove what it saw.
Engineering
11 min read
Buying AI
15 Jun 2026
Presales demos that close: build a narrow working slice, not a click-through
Buyers stopped believing mockups. A demo that runs on their data, with guardrails, moves deals faster than a polished deck.
Buyers
7 min read
Cost
19 May 2026
Cutting Claude and GPT bills: where the money actually leaks
Most large inference bills are not a pricing problem. They are a routing, caching and context problem.
Buyers
9 min read
Models
8 Apr 2026
Distilling a frontier model into an industry-specific small model
For one narrow job, a small model trained on your own traces can beat the model that taught it — cheaper, faster, and yours.
Engineering
10 min read
Compliance
26 Mar 2026
Data residency and GDPR for AI systems: where your documents and prompts really go
The honest answer to 'where does our data go' is a diagram, not a reassurance. Here is how to produce one.
Buyers
8 min read
Compliance
4 Mar 2026
What the EU AI Act actually asks of you, in plain procurement language
Not every AI system is high-risk. Knowing which of yours are — and what each tier owes — is a two-day exercise, not a legal saga.
Buyers
9 min read
Buying AI
11 Feb 2026
Why AI pilots stall before production — and the six checks that predict which ones ship
Most enterprise AI pilots die between the demo and the change request. The failure is almost never the model.
Buyers
8 min read
Questions on any of these, or want the working version applied to your data? Write to gulshan@xelec.in.
Next step
Reading is cheaper than a stalled pilot. Talking is cheaper still.
Thirty minutes with an engineer who has shipped this. If there is no fit, we say so on the call.