Delivering a lesson is a solved problem — a player, a quiz engine and a certificate PDF are a fortnight’s work for a competent team. The two problems that are not solved are holding attention and proving that learning happened. The last two years made the first newly possible and the second newly suspect.
Thirty minutes on one feature, before anybody writes code. We map what it would cost in the three currencies you already worry about — latency, safety review and integration — and you keep a one-page written assessment. If your problem is retention rather than AI, take the same slot as a learner drop-off review instead. Same half hour, same one-page artefact.
Video, text, a quiz engine, progress bars, a certificate, an email when a cohort starts. This is well-understood work with well-understood libraries, and any competent team ships it. If a vendor’s pitch is mostly this, you are being sold a commodity at a premium.
Newly possible: a model that can hold a real conversation changes what a tutor can be — if it answers fast enough, stays in subject, and never says the wrong thing to a child.
Newly suspect: the moment a learner has a chatbot in another tab, every assessment instrument written before 2023 is measuring something other than what it claims to measure.
We have shipped the hard version of the first one. Which means the difficult parts are not slideware to us: the latency budget between a child finishing a sentence and the character answering, what happens when the model says something wrong to an eight-year-old, keeping visemes locked to audio while the network jitters — and the fact that a child will not wait for you.
Not categories. Situations — described the way the person living through one would describe it on a first call.
The content is genuinely good — subject-matter experts wrote it, the pedagogy is defensible, the reviews are warm. And the completion rate is a figure that gets rounded in the board deck and never printed on the website. Nobody in the company can say with confidence which lesson loses people, or whether the ones who finish actually learned something the ones who quit did not.
The AI tutor is perfect on stage: clean audio, an adult speaking in full sentences, a laptop on good wi-fi. Put it in front of a real nine-year-old on a school network and the round trip stretches, the child talks over the character, the lips drift out of sync with the audio, and the illusion is gone in one turn. Everyone in the room now knows the demo was a demo.
Five years of courseware lives in SCORM packages inside a platform the company has outgrown. There is no clean export, the tracking data is a completion flag and a percentage, and every migration quote starts with the word rebuild. Meanwhile a large enterprise buyer is asking whether the product launches through their LMS over LTI — and the honest answer is not yet.
The certification means something commercially because it is hard. It is no longer hard. Item banks written before 2023 are answerable by a model in seconds, proctoring vendors are a reputational risk of their own, and text detectors return false positives on exactly the learners least able to contest them — non-native writers especially. The exam still has to be defensible to an accreditor next quarter.
Product has a working prototype and it is good. Legal will not approve it, because nobody can answer the only question that matters: what will the model say to a child in the worst case, and what happens next when it does. There is no red-team transcript, no escalation path, no logged record of what was said. So the feature sits, quarter after quarter, while the roadmap slides around it.
The dashboard reports logins, video starts and minutes in app. The Head of Curriculum wants to know whether the learners who scored well in March still know it in July, and whether the new module caused the improvement or the cohort did. The event pipeline was never designed to answer either question, and no amount of dashboard work will make it.
A child stops speaking. Five things have to happen before the character answers, and together they get about a second — under roughly half of one it reads as conversation; past a whole one, the child has already looked away. Drag any stage and watch the total move. This is the conversation we have on the first call, and it is the fastest way to tell a shipped system from a showreel.
Illustrative budget, not a benchmark of your stack or ours — the point is the shape of the problem and where the milliseconds actually go. Select a stage and use the arrow keys to change it; Home and End jump to that stage’s floor and ceiling. Reset puts the numbers back.
A learner speaks. Speech recognition streams partial hypotheses and has to decide, quickly and correctly, that the sentence has actually ended — children trail off, restart, and talk over the answer. An LLM produces a reply that has to stay inside the lesson and inside the subject. Text-to-speech turns it into audio, and an animated 3-D character speaks it with visemes locked to that audio, rendered in a browser on whatever laptop the school bought in 2019. Every one of those stages spends milliseconds out of a budget of about one second, and a child enforces it more strictly than any product manager.
The part nobody demos is the part that took the longest: what the model is allowed to say. Guardrails for a child-facing tutor are engineering, not a system prompt — layered filtering, subject and behaviour boundaries, an escalation path to a human, and a logged transcript a reviewer can actually read afterwards. That work is what turns a prototype that legal will not approve into a product that ships.
Very few agencies have taken one of these all the way through. We say that plainly and without superlatives, and the fastest way to check it is to ask us for the latency budget stage by stage.
A funded-trader training programme where the assessment is the product: passing the evaluation has direct financial consequence, so the evaluation logic, the live market data feeding it, and the record of every decision that produced a result all have to be exactly right, at scale, every day. We built the admin platform behind it — evaluations, funded-trader states, compliance search, billing, affiliates and the analytics on top — alongside the live market data and the operational tooling the team runs the business from.
It is on this page because it is the clearest thing we have to point at when someone asks whether we understand high-stakes assessment. It also appears on the FinTech & Banking page, from the market-data and payments angle.
A computer-vision layer built with OpenCV and MediaPipe that lets an interaction respond to what is happening in front of the camera rather than following a script. We present it as an engineering capability and nothing more: opt-in, processed on the device wherever the deployment allows, purpose-limited to the narrow signal it exists to produce, and never a score attached to a named learner.
What it is not, stated so nobody has to ask: it is not proctoring, not monitoring, not surveillance, and not an “attention score” for a person. We do not build those, and we would rather lose the work than write this paragraph loosely.
A character that talks is a demo. A character a curriculum team can direct is a product. Behind the character sits the unglamorous half: an authoring flow where a lesson is assembled step by step, each spoken line gets its own delivery and its own duration, and the audio can be regenerated until it sounds like a teacher rather than a public-address system. Every second visible here is a second inside the turn budget above.
Thirty minutes, an engineer on the call rather than a sales team. You leave with a one-page written assessment: the pipeline stages and where the milliseconds go, the guardrail and review work your compliance lead will demand, and the integration surface against the platform you already have. If retention is the real problem, take the same slot as a learner drop-off review — where learners actually leave, and what your current event data can and cannot prove about it.
Timelines are ranges and they are estimates from people who have been wrong before. Everything in the third column is an artefact you keep — in your repositories, your cloud account, your RFP responses — whether or not the engagement continues.
A learner speaks; a character answers in under a second, in character, inside its subject.
The pipeline end to end — streaming ASR with endpointing and barge-in, LLM orchestration, TTS, viseme-driven rendering — plus a measured latency budget per stage, a guardrail layer, and transcript logging you can hand to a reviewer without editing it first.
The work that unblocks a feature which has been sitting in review for two quarters.
Threat modelling for what a model might say, layered filtering, subject and behaviour boundaries, an escalation and human-review path, red-team transcripts you can read, and an honest written statement of the residual limits — the part most vendors leave out, and the part your legal lead will actually read.
A score you can defend to an accreditor, and an integrity design that does not depend on detection.
Item banking and versioning, blueprint coverage, exposure control, adaptive delivery where it earns its complexity, and a record of every decision that produced a score. Plus an integrity model built on task design rather than on AI-text detectors, which we do not sell and would not trust.
Launch inside your customers’ LMS, and stop being hostage to one platform.
LTI 1.3 with the Advantage services that matter — deep linking, names and role provisioning, assignment and grade services — xAPI statements into an LRS or cmi5 where that fits, QTI for item exchange, OneRoster for rostering, and SCORM 1.2/2004 ingest with a documented route off it.
Evidence of mastery and retention, not a dashboard of logins.
An event model designed around evidence rather than activity: mastery thresholds, retention measured at intervals rather than at the end of the lesson, cohort drop-off with causes attached, and a written list — agreed in advance — of the exact questions the pipeline will be able to answer once it is running.
Enrolment, entitlements, billing, cohorts, support tooling and internal dashboards.
The unglamorous half, built to hold thousands of concurrent learners and the people who administer them. This is most of the work in most education products, and it is the half that decides whether your team spends its week teaching or reconciling.
WCAG 2.2 AA against your real interactive content, not a scanner report.
An audit run with actual assistive technology through the paths that matter — a screen reader traversing an interactive exercise, a keyboard-only route through a timed assessment, captions and transcripts, line length and typography that a dyslexic reader can use, and a reduced-motion alternative to an animated character — then the fixes, then a conformance statement you can put in an RFP response.
Camera and audio signals that make an interaction responsive rather than scripted.
An opt-in, purpose-limited component with on-device processing wherever the deployment permits, and a written data-flow description your privacy counsel can review as-is. Not proctoring, not monitoring, and never a score about a named learner — if that is the brief, we are the wrong shop.
The screen a nine-year-old actually sees first: a name, a time, and one button. Everything this page is about — the streaming recogniser, the model, the voice, the rig, the guardrails, the transcript that a reviewer will read next week — sits behind it and is supposed to be invisible. That is the standard. If the learner can tell how any of it works, something upstream has gone wrong.
AI-text detectors are unreliable in a way that is well documented, and their false positives land hardest on non-native writers — the learners least equipped to contest an accusation. A product that stakes its integrity on one has bought a liability and called it a feature. The defensible answer is task design, not detection.
AI used to decide admission, evaluate learning outcomes, or monitor prohibited behaviour during a test sits in the high-risk category, with obligations landing from August 2026. If any of that describes your roadmap, the architecture decisions that matter are being made now, not then.
Neutral age screening for mixed-audience services, separate consent for disclosure to third parties, written retention policies, and an end to keeping children’s data indefinitely. These are architecture requirements before they are policy documents.
The European Accessibility Act and EN 301 549 pull e-learning products toward WCAG 2.2 AA in Europe, which turns accessibility from a values question into a procurement one. An RFP that asks for a conformance statement will not accept a scanner score.
The industry’s Student Privacy Pledge wound down in 2025, so “we signed the pledge” is no longer an answer to anybody. What answers now is architecture: what you collect, why, where it lives, who it is shared with, and when it stops existing.
| What you are thinking | Our answer |
|---|---|
| “AI tutors are a demo, not a product.” | Fair, because most of them are. Our AI video agent is live in a browser: an LLM, a streaming voice layer and a rendered character answering a child in real time. Ask us about the latency budget stage by stage — that conversation is the fastest way to tell a shipped system from a showreel, and it is a conversation a showreel cannot survive. |
| “We cannot risk what an AI says to a child.” | Neither can we. That is guardrail engineering, not a prompt: layered filtering, subject and behaviour boundaries, escalation to a human, logged transcripts a reviewer can read, and a written statement of what the system still cannot guarantee. We will tell you where the limits are before you sign anything. |
| “Our learner data is regulated.” | We build to FERPA, the amended COPPA Rule and GDPR children’s-data expectations by architecture — minimised collection, explicit purposes, real retention policies, and deployment inside your own environment where that is the right answer. Compliance is a property of your product, not of us: we hold no compliance certifications and we will not imply otherwise. We build what your counsel signs off. |
| “We already have an LMS.” | Good. We integrate through LTI 1.3 and report through xAPI. We are not selling you an LMS, we are not reselling anyone else’s, and we are not migrating you off one you like. |
| “We tried an outside dev shop and it failed.” | Usually the same failure: an account manager sitting between you and the engineers, and nobody senior accountable for the architecture. Here you talk to the people writing the code, every project is signed off by a senior architect before development starts, and the source, infrastructure-as-code and runbooks are in your repositories from the first week. Start with a small scoped piece and judge us on it. |
| “You are in Europe.” | Fluvius USA Inc is a US entity in Sacramento, California, and contracts on US paper. We work EU hours with a US-morning overlap, so a US team has live hours every working day. Behind it: 200+ clients, ten years of them, 100% Job Success on Upwork, 4.9/5 on Clutch, and many clients who have worked with us for 7+ years straight — which is the figure that matters here, because education products are multi-year relationships. |
| “This will take a year.” | A conversational tutor pilot is typically 10–16 weeks. An LTI integration is weeks, not quarters. And if a scope genuinely needs a year, you will hear that on the first call rather than in the third invoice. |
| “What if the first month goes badly?” | The first 30 days of any engagement are a trial period, and either side can end it in that window without notice and without a penalty. It is in the agreement rather than in a sales conversation. A month of real work tells both of us more than another round of evaluation would. |
| “Can you talk about us afterwards?” | Only at the level of a technology stack and the roles on the team, and nothing beyond that without your written consent. Confidentiality is mutual and it does not expire when the engagement does. |
The other end of this industry: an evaluation a learner passes or fails inside a live market, where passing releases real trading capital. Nothing here is a completion percentage. Every rule that decides the outcome, every tick of data it was decided against, and the record of both have to be right the first time — because a disputed result is not a support ticket, it is a claim.
What the model may say and what stops it, what happens when it says something wrong anyway, who is told and how fast, what is logged and for how long, what a parent is told, and what you will write down as the limits you cannot guarantee. In the order a review actually asks them, with a column for your own answer — so the meeting is a read-through rather than a discovery.
Bring the feature that is stuck — the tutor that dies on a school network, the exam a second tab defeats, the integration a customer keeps asking about, or the completion number nobody wants to say out loud. We price it in latency, safety review and integration, and you keep the written page whichever way you go.
A senior software engineering team that uses AI as an industrial-grade tool — advanced Claude Code workflows in experienced hands.
