Research

Writing about problems we have built for.

Case studies sell outcomes. This is where we write about the work itself: the constraints, the decisions, and what we learned along the way. Each piece goes deeper than a blog post and more specific than a whitepaper.

11 min read

Beyond the receptionist: where private clinics lose time and money

How MediConcierge, the AI patient front door, absorbs repeat enquiries and surfaces previously invisible work, sitting in front of the clinic’s existing practice-management system. Part two of the MediConcierge series.

Private clinic economics are tightly coupled to receptionist capacity. The person at the front desk decides who gets booked, who gets lost, which enquiries get followed up, and which get silently abandoned after-hours. When that capacity is overloaded, which it almost always is, the commercial cost shows up as lost leads, missed appointments, empty slots, and a clinic owner who has no visibility into where the friction actually is. This paper describes how MediConcierge, the AI patient front door, absorbs first-touch enquiries and surfaces the work the clinic could not previously see. It is built to sit in front of the practice-management system the clinic already runs rather than replace it. Companion to the clinic-safe AI patient concierge paper; this one covers the operational-commercial side rather than the safety engineering.

  • Healthcare
  • Clinic Operations
  • Patient Front Door
  • AI SaaS
  • Front-desk Operations
Read the paper
11 min read

Persona drift, tone drift, and the quieter cousins of agentic misalignment

Small-scale misalignment in production LLM systems, and the controls that catch it, drawn from voice roleplay training and healthcare conversational AI.

Anthropic’s research on agentic misalignment describes dramatic failures in high-autonomy settings: LLMs choosing blackmail, leaking confidential data, or ignoring safety instructions to preserve goal pursuit under simulated threat. Production LLM products experience a quieter, commoner cousin of the same dynamic. Buyer personas in voice roleplay drift back toward helpful-chatbot mode after eight turns. Medical concierges want to tell users what a symptom could mean even when explicitly told not to. Scoring models rate polite-but-off-target responses more favourably than human graders would. This paper describes the four categories of small-scale misalignment we encounter most often across our products, the multi-layer controls that catch them, and what production experience suggests about how the Anthropic findings connect to everyday LLM engineering.

  • AI Safety
  • LLM Controls
  • Production AI
  • Conversational AI
  • Misalignment
Read the paper
9 min read

Inside MediConcierge: a clinic-safe AI patient concierge

The conversational AI layer in MediConcierge, our AI patient front door for UK private healthcare, and where the safety engineering concentrates.

UK private healthcare clinics are a strange commercial environment for AI. They handle repetitive inbound questions well suited to conversational tools, but they also operate under UK GDPR, Caldicott principles, and a duty of care that treats a sloppy AI response as a safeguarding issue. This paper focuses on MediConcierge, the AI patient front door that handles first-touch enquiries and booking, sitting in front of the clinic's existing practice-management system. It tracks leads and patients lightly so staff can follow up. We cover the concierge specifically because it is the component that touches patients directly, which concentrates most of the interesting safety engineering.

  • Healthcare
  • Conversational AI
  • Multi-tenant SaaS
  • UK GDPR
Read the paper
11 min read

Voice AI for roleplay training: latency, silence, and the illusion of a real conversation

What it takes to make a voice roleplay partner realistic enough that trainees engage with it as practice rather than as a test. The problem appears in sales, HR, clinical communication, and anywhere else a conversation can make or break the relationship.

Some conversations are too important to practise on the people they are actually for. Sales reps cannot ramp on real customer calls without losing deals. HR managers cannot rehearse a termination meeting on an actual employee. Junior clinicians cannot practise breaking bad news on real patients. A voice-first AI practice partner can fill that gap, but only if the experience feels real enough that the trainee engages with it as a conversation rather than as a test. This paper describes how we built a voice AI roleplay training platform, first deployed for wholesale sales at a UK consumer-goods brand and designed to generalise to HR, clinical communication, and other domains. It covers the three problems at the heart of making the system work wherever it is applied: latency in the voice pipeline, silence detection for turn-taking, and persona stability across multi-turn conversation.

  • Voice AI
  • Conversational AI
  • Roleplay Training
  • Latency
  • LLM
Read the paper
10 min read

Calibrated colour matching on a phone camera: CIE LAB for field inspection

Replacing visual quality inspection with objective measurement, and the operational savings that follow.

Field hygiene inspection teams run most quality inspection by eye. That produces three recurring problems: different technicians grade the same sample differently, lighting varies from site to site in ways that bias the assessment, and verbal grades produce no audit trail. This paper describes LineSpec, the system we built to grade cleaning quality by colour for field hygiene inspection, with line-cleaning and draught-dispense hygiene as the first vertical. It uses phone-camera colour measurement in CIE LAB space, with CIEDE2000 as the difference metric, in an offline-first PWA. The technical core is calibration: getting a phone camera to produce colour measurements consistent enough to trust under real-world field conditions.

  • Computer Vision
  • Field Operations
  • Colour Science
  • PWA
  • Measurement
Read the paper
RSE Labs billboard

More to come.