AI in Production: Where It Actually Earns Its Place

Last Thursday, Qualabs hosted a Tech Meetup built entirely around one question: what's actually running in production right now? Nicolás Levy opened with three talks covering AI in production video workflows. What emerged from three different angles was striking: the teams that handle AI best treat it as one tool among many, with clear boundaries around where it adds value.

Sign Language, Built as a Production Pipeline

Alexander Leschinsky from G&L Systems opened with a problem broadcasters carry quietly: qualified sign language interpreters are scarce, and demand keeps growing as accessibility becomes a legal requirement.

The numbers tell the story:

• UK (BSL): 151,000 deaf users. Only ~1,137 interpreters — about 130 signers per interpreter.
• US (ASL): ~500,000 users. Only ~10,000 interpreters — about 50 signers per interpreter.
• Germany (DGS): ~200,000 users. Only a few hundred interpreters — an acute shortage.

G&L builds media pipelines for live events. Requests for sign language interpretation started arriving from parliaments, broadcasters, and corporate conferences — with no interpreters available. And in many jurisdictions, it's not optional anymore. It's law.

Alexander's team partnered with Signaps, grounded in linguistic research from the University of Surrey. The difference: sign language models trained on real interpreters, not generic animation. Two deaf-led organizations — the Royal Association of Deaf People and Action Deafness — hold shares, keeping the deaf community central to how the models get built.

The Honest Part

Getting the avatar to move convincingly turned out to be the easy part. Getting the grammar and translation right takes the real work — treating sign language as its own language, with its own rules.

"The hard part isn't getting the avatar right. It's getting the grammar right and treating sign language as what it is: a language of its own with its own rules."

— Alexander Leschinsky, G&L Systems

Here's what makes this real:

• Sign languages aren't translations. They're distinct languages with their own grammar, facial expressions, and historical evolution.
• You can't record glosses once and reuse them. German Sign Language uses the mouth differently than British Sign Language. The same sign for "potato" is different across languages.
• LLMs help strategically. They analyze text context to determine grammatical intent before selecting pre-recorded authentic signs and merging them live.

Current latency sits at 8-10 seconds, versus 3-4 seconds for human interpreters. For delayed content, teams can synchronize both feeds perfectly.

One Shared Language for Live Incident Triage

Nicolás Caballero from Qualabs presented the Live Pattern Generation and Recognition System (LPGS) — built for the problem every operations team knows: when a live incident hits, you need answers, fast, under pressure. When data comes from multiple providers, each speaking its own language, that's where it breaks.

The Problem

Live teams pull data from Conviva, Mux, and Akamai, each with metrics named differently. Cross-referencing them manually while a live event keeps running is hard enough. The hardest part is explaining why something broke.

"Your team has a playbook for every issue they've ever seen. But when something new happens at 3am, explaining why is the hard part — and that's where confidence breaks."

— Nicolás Caballero, Qualabs

The Solution

When a patient arrives with familiar symptoms, the doctor treats it right away. When they don't, a specialist reviews the same data and proposes a diagnosis. LPGS follows that logic:

1. Adapters translate every provider's metrics into one shared vocabulary.
2. Detection watches for known anomalies using deterministic rules.
3. Diagnosis: a known pattern gets an instant diagnosis; an unknown one escalates to an LLM.
4. Validation: a human expert approves before the new pattern gets saved.

Three guardrails keep LLM hallucinations in check: evidence-based reasoning only, flagging insufficient information instead of guessing, and human approval before anything reaches production.

"The worst thing an LLM can hand you at three in the morning is a confident wrong answer."

— Nicolás Caballero, Qualabs

Nico's demo showed two live scenarios — a CDN failure and a playback issue on Samsung Smart TVs — each resolving with a Slack alert in seconds. Work that used to mean pulling data from several dashboards by hand.

The payoff:

• Known issues resolve instantly.
• The LLM only gets called where it adds value.
• New providers plug in as one component.
• Quality climbs because every new pattern passes through human validation.

Every Clip Earns One of Three Verdicts

Emil Santurio, also from Qualabs, closed with a problem that sounds simple but breaks most systems: automate quality control for hundreds of highlight clips every weekend, without losing accuracy.

Every clip gets one of three verdicts: pass, fail, or ask a human. The real insight isn't "use AI or not?" It's "which tool answers which check?"

Three Lanes, Three Tools

Point one tool at all three and you lose both ways. Too little AI and mistakes slip through. Too much AI and you burn cost and trust on things code could handle.

"Every AI answer has to come back with a timecode and a measurement. That's what makes the system explainable instead of a black box nobody trusts."

— Emil Santurio, Qualabs

In practice: playability and decodability checks run 100% deterministic. Logo and scoreboard detection currently runs on OpenCV, not AI. Editorial decisions stay with human reviewers.

Key Lessons

1. Start deterministic. Decompose "quality" before touching any model.
2. Evidence required. Every AI answer must carry a timecode, frame, and measurement.
3. Build a labeling corpus early. Good clips, bad clips — you can't measure regression without it.
4. Put models behind adapters. Technology moves fast. Don't hardwire one in.
5. Design for trust. Every choice earns or spends operator confidence.

The Pattern That Kept Surfacing

Across all three talks, teams putting AI into production treat it as one tool among many, with clear boundaries around where it adds value.

• Evidence backs every decision. Humans need to understand why a system flagged something.
• Humans keep authority over judgment calls. Not because they're always right, but because context matters.
• Every model sits behind an adapter. Technology changes. Systems need flexibility.

Why This Matters

The industry has been asking "how do we add AI to video workflows?" for years. These three teams are asking a harder question: where does AI actually add value?

Sign language accessibility isn't an edge case — it's a legal requirement and a human right, built on linguistics and community partnership, not animation.

LPGS isn't about replacing operators. It's about making them faster and more confident, freeing them for cases that need judgment.

Quality control isn't automation for its own sake. It's scaling human judgment without losing it.

That's the shift: from "look what AI can do" to "look what we can build with AI when we're honest about its limits."

For Builders

• Start deterministic. Most problems are measurement, not intelligence.
• Evidence, always. Build trust, not magic boxes.
• Humans in the loop. As essential validation, not fallback.
• Cost-effectiveness first. The right tool might be code, a model, or a person. Choose per check.
• Build adapters, not hardwired solutions. Models change. Design for flexibility.

The real work isn't the AI. It's knowing where it belongs.

The Tech Meetup series is where Qualabs shares real learnings from building media workflows in production. If you're interested in joining us for the next one, let's talk.

Want to see what happened at the meetup? Watch the full recording here.