Niche systems
Work that does not fit the six main areas because it is too specific for them. Enterprise VPN appliances, Peppol e-invoicing, IP camera surveillance, real-time calling, middleware, custom firmware images, and security challenges. Most of these arrive because a generalist declined them.
Enterprise appliances
SSL VPN and identity products, where the configuration is the attack surface.
Ivanti Connect Secure and Ivanti EPMM, along with Pulse Secure before the rename, Cisco ASA and IOS, and Fortinet. The recurring problem on all of them is the same: an appliance ships with a configuration that gets you connected, and almost none of that configuration survives contact with the threat model the vendor later assumes.
Work covers admin interface exposure, certificate and key handling, authentication policy, split tunnelling scope, and the audit logging that has to be in place before anything else is changed. Also migration between appliances, where the thing that breaks is rarely the tunnel and usually the management plane.
On devices that have been publicly exploited, the first job is establishing what the logs show, because an appliance that was already an authentication bypass does not tell you in the config that it was.
Peppol and electronic invoicing
The EU standard for public sector e-procurement.
Peppol is a network of service providers rather than a protocol, which is the part people get wrong when they quote it. A sender looks up the receiver through a participant discovery service, resolves routing through a service metadata registry, and then exchanges the actual document over AS4. Getting any one of those three wrong produces an integration that works with the one provider you tested against and silently fails with everyone else.
The documents are EN16931 compliant, which in practice means UBL or CII syntax plus Schematron validation against the profile a particular buyer has published. Buyers are frequently government entities with their own profile, and a document that is valid EN16931 is not automatically valid for them.
Also worth naming because it comes up alongside it: EN16931 validation tooling, EDI-style business rules, and the PDF and HTML renditions a human recipient has to read.
Surveillance and IP camera systems
Cameras, recorders, and everything between the lens and the viewer.
Almost all of this is the same work as streaming video, pointed at a different source. An IP camera publishes RTSP, usually with a high bitrate main stream and a separate low bitrate sub stream that most recorders will only touch when the main one fails. ONVIF handles discovery and control, and the parts worth getting right are the ones where vendors interpret it differently.
The recorder is the interesting half. A Linux box ingesting dozens of streams needs enough IO and storage to survive a retention window measured in weeks, and the storage maths is where projects go wrong: a 4MP stream that looks affordable per camera becomes unaffordable at thirty cameras and thirty days. Motion-indexed recording cuts it substantially, and it only works if the detection runs.
Where it converges with everything else on this site: repackaging those streams for a browser, latency tuning so a guard sees a live view rather than a slideshow, and the fact that a camera fleet is a security problem long before anyone asks about it. Default credentials, firmware age, and cameras exposed to the internet are still the common case.
Real-time and calling applications
Voice and video where a few hundred milliseconds is the whole product.
Low latency calling is a different discipline from delivering video on demand. The buffer that makes VOD smooth is the thing that makes a conversation feel broken, so the work is in removing delay rather than absorbing jitter: adaptive jitter buffers tuned small, pre-encoded audio so the encoder is not on the critical path, and the ordering that keeps late packets from delaying good ones.
WebRTC covers this with ICE for candidate gathering, STUN and TURN for connectivity, and DTLS-SRTP for the media itself. Getting a call through symmetric NAT reliably is the part that takes the time, and TURN is the part that takes the money, so the two are designed together rather than separately.
SIP is the other half and it is less forgiving: INVITE, 200, ACK, BYE, and an SDP negotiation that has to agree on codecs in an order both ends will accept. Trunking, Asterisk PBX, and the header rewriting that happens when a carrier and an internal system disagree about what a contact is.
Middleware
The layer that exists so two systems can disagree politely.
Middleware is unglamorous and it is where a lot of integrations live. Protocol translation between things that were never designed to talk, queueing so a slow consumer cannot stall a fast producer, retry with backoff so a three second outage does not become lost data, and an idempotency story so a retried message does not become two rows.
Industrial and building systems are where the awkward protocols live: Modbus, MQTT, OPC-UA, BACnet, each with its own idea of what a device is and no interest in your schema. Bridging those into something a modern service can consume is most of the job, and it is work very few people want to take on.
API gateways sit in the same family: authentication, rate limiting, request shaping, and the audit trail that makes it possible to answer who called this endpoint at three in the morning six months later.
Custom firmware and operating systems
Images built from source, because the vendor stopped.
Custom OpenWrt images for routers whose models fell out of the vendor's build system, with the packages, the defaults, and the first-boot behaviour decided rather than inherited. Where a model has no published image at all this means building from source and carrying whatever patches the hardware needs.
Custom Linux distributions for a single purpose, including bootable Ubuntu and Debian images built from the ground up when the standard installer cannot produce what the deployment needs. The discipline is reproducibility: the same inputs should give the same image, because an image you cannot rebuild is an image nobody can patch.
Embedded boards run the same idea with less room. Device trees, custom kernels where an out-of-tree driver is unavoidable, and init systems sized for kilobytes rather than megabytes. See Linux and systems for the wider picture.
Workflow platforms
n8n and similar, including the parts that need writing rather than clicking.
n8n covers a great deal on its own, and the work starts where the built-in nodes stop: custom credentials, custom nodes against an API nobody has published an SDK for, and webhook handling that survives a burst and a retry at the same time.
Self-hosted deployments have their own concerns that a hosted instance never has. Execution state across a restart, queue mode so a long job does not block everything behind it, and credential encryption bound to a key that is backed up somewhere other than the database it protects.
See Automation and AI for the wider automation work, including the machine learning and vision work.
Security research and challenges
Authorised testing, and building the exercises other people learn from.
Penetration testing against systems where the authorisation is in writing before the first packet goes out. Web, network, and application targets, with the report mattering as much as the finding: a vulnerability nobody can act on is a scan result, not a test.
CTF challenge design is the other half of this. Building web, crypto, binary exploitation, and forensics challenges means the same thing twice: making them genuinely hard without making them unsolvable, and making sure there is exactly one intended path so the people who finish it learned something. That requires knowing where the shortcuts are, which means solving it yourself first, repeatedly.
Exploitation research sits alongside, on lab machines and virtualised targets. See Networking and security for the defensive side of the work.
If it is on this page and you need it
These are the projects that get passed over twice before they reach somebody. A camera system that two integrators declined, an appliance where the previous person would not touch the config, a Peppol integration that only worked with one provider. If something on this page sounds like your problem, it probably is, and the first conversation is enough to find out whether it is.