01
Real-Time AI Audit
A fixed-price diagnostic: where the latency goes, what breaks, what it costs.
// Real-time AI engineering — v0.1
Real-time AI systems, engineered for latency, operated for uptime.
We build the real-time layer under voice and video products — transport, telephony, turn-taking — and budget the latency through the phone leg. After launch we keep it measured: evals on every model change, cost per minute tracked.
global footprint — live6 sites / 4 regions
- London · Fintech — -82% manual review
- Boston · Healthtech — 0 hallucinated citations
- Berlin · E-commerce — 47% tickets auto-resolved
- Singapore · Logistics — -31% stockouts
- Los Angeles · Studio HQ — home base
- Tokyo · Delivery — distributed team
x 0.42 · y -0.18 · z 0.91spatial mesh
15+systems shipped
6+ yrsin production
~4 wksto a first slice
99.9%uptime SLA
// Trusted by teams in fintech, health, and logistics
- Northwind Health
- Vantage Pay
- Harbor Commerce
- Meridian Freight
- Halcyon Bio
- Seabridge Markets
- Lumira Labs
- Aster Energy
// What we do
Assess, build, rescue, run — plus the two things underneath. We measure what we claim, and show you the measurements.
02
Voice Agents in Production
Voice agents that survive real calls — latency, turn-taking, telephony, evals.
03
Real-Time Rescue & Migration
Inherit a broken or orphaned real-time stack, stabilise it, move it somewhere maintainable.
04
Real-Time Ops
Keep it working: the right counters, alerts that fire early, incident response in stated hours.
05
RAG & Knowledge Systems
Retrieval grounded in your data, with evals that prove the answers are real.
06
Compliance-Grade Recording
Consent, redaction, retention and an audit log for regulated conversations.
// How we work
From prototype to production, in four moves.
01
Discovery
We map the problem, the data, and the eval that defines "done".
02
Prototype
A working slice in weeks — real model, real data, measured.
03
Production
Hardened, observable, evaluable. Shipped where users live.
04
Scale
Cost, latency and reliability tuned as load and scope grow.
// Selected work
Shipped systems, measured outcomes.
Anonymized for confidentiality. The numbers are the point.
Fintech
Document intelligence
Extraction and review pipeline over dense financial documents.
-82%manual review
Healthtech
Clinical RAG
Grounded answers with citation-level evals on clinical sources.
0hallucinated citations in eval
E-commerce
Support agent
Tool-using agent resolving tickets end-to-end with guardrails.
47%tickets auto-resolved
Logistics
Demand forecasting
Forecasting models feeding replenishment decisions daily.
-31%stockouts
// Toolchain
The stack we build on.
- Languages06
- Frontend & mobile06
- Backend05
- Data05
- Cloud & DevOps06
- Queues & streaming04
- Realtime & desktop05
- AI08
- Testing04
- Design05
- Workflow05
See the full stack59 tools · 11 domains
// Engagements
Start small, scale as it proves out.
One path, three steps — a paid Sprint to de-risk, a fixed-scope build, then ongoing operation as it grows. Priced on outcomes, not hours.
01 · Entry
Start hereArchitecture Sprint
$2–4kfixed · 1–2 wks
De-risk before you build. We map the system, choose the architecture, define what “done” and “fast enough” mean, and prove the risky part with a working POC. Credited to the build.
- Target architecture & success metrics
- Scope, plan & costed fix list
- Fixed price, credited against the build
02 · Build
Build
from$12kfixed scope or pod
Ship the product — designed, built, and delivered in your repo and conventions, with evals and observability from day one.
- Turnkey fixed scope, or a dedicated pod
- Evals & observability from day one
- Documented and yours — no lock-in
03 · Operate
Where it growsRun & Scale
from$1.5k/ month
Keep it running and improving after launch — SLA, monitoring, cost control, and steady iteration as you scale.
- Incident response in stated business hours
- Uptime, performance & cost monitoring
- Limits re-measured as you scale
+ Specialized tracks — deeper engagements for real-time video, AI, and data-intensive products, when your domain needs it.
Not a pick-one menu — most teams start with a Sprint and grow into ongoing operation. Every engagement ships documented, evaluable, and yours.
Not sure which? Get an estimateMost start with a short, paid discovery sprint to de-risk scope and shape the plan, then iterative delivery in two-week increments with a working system at the end of each. You see progress continuously — no big-bang reveal at the finish.
We typically ship a first working slice in about four weeks. The discovery sprint front-loads the unknowns, so build starts against a clear target rather than a moving one.
We treat model output as untrusted input: retrieval grounding for factuality, evals wired in from day one (including citation-level checks), observability on every stage, and sanitization before anything is rendered or stored. Reliability is measured, not assumed.
Yes — entirely. We work in your repository, match your conventions, and hand over documented, evaluable systems with no proprietary runtime or lock-in. What ships is yours.
Almost always. We build on mainstream tooling (TypeScript/React/Next, Go, Postgres, the major model providers) and adapt to yours rather than imposing ours — the pull requests should read like your team wrote them.
The discovery sprint is a fixed fee; delivery is then scoped from what it surfaces — typically a fixed monthly rate per increment. No open-ended hourly billing.
// Insights
From the blog
Real-time videoHow many cameras fit on one box — and the ceiling that hides until you cross it
Sizing a box for a camera fleet has two answers and people only compute one. Here are measured cores per stream for the software path, the hardware ceiling that stops you at eight sessions while the GPU sits at 8% utilisation, and what crossing it looks like — which is nothing at all.
Real-time videoYour SFU bill is a transcode bill — the 24/7 publish nobody ordered
One multiplication tells you whether your media bill is a bandwidth problem or a transcoding problem — and it is almost always the second. Modelled at 1000 cameras against a published price list, the difference between two configurations is a six-figure monthly line, and the fix has a price nobody quotes.
Real-time videoVideo is black on iPhones and fine everywhere else — the level your camera declares
Android plays it, desktop plays it, the iPhone shows a black rectangle and reports no errors. The cause is one number in the SPS, the tell is two counters in getStats, and the fix costs no CPU at all.
// Manifesto
Less marketing, more engineering. We measure twice and ship what survives production — observable, evaluable, and yours.
// Let's build
Start a project.
Tell us where you are. We reply within a day with a concrete next step.