Why Strong AI Models Still Fail in Production: The Three Gaps and Operator Solution

Models That Work in Demos Often Fail in Production

As model capabilities grow, the real blocker for business deployment is no longer whether a model works, but whether it can run stably ten thousand times, be reused across teams, and continuously produce measurable results. At the Lance Meetup in Shanghai on September 12, Volcano Engine algorithm engineer Zeng Qian used the LAS multimodal operator system as a case study, alongside three large-scale scenarios—embodied intelligence, e-commerce material generation, and video ad placement—to show how a unified operator system moves strong model capabilities into enterprise production systems.

The Three Gaps of the "Last Mile"

Productionizing models requires crossing three gaps:

Model Capability Gap : Large models have input format limits, struggle with long-video understanding, and suffer low success rates in video generation. A full long video cannot be processed in one call; it needs frame extraction, shot detection, audio separation, and semantic compression. Complex PDF parsing (titles, tables, formulas, page numbers, image-text relationships) also cannot be solved with a single model call.

Engineering Pipeline Gap : Real data processing mixes CPU preprocessing, GPU inference, rule validation, storage, and task management. Rewriting glue code for every new scenario makes it hard to reduce launch cycles and maintenance costs.

Quality Certainty Gap : Video generation is stochastic; a usable asset may need many retries. Returning a single generation result is insufficient for batch delivery. Input validation, automatic evaluation, targeted repair, retry, and human confirmation must be built into the process.

LAS addresses these gaps by encapsulating differences and limits into operators: validate and preprocess inputs, record parameters and state during execution, auto-evaluate outputs, and apply targeted fixes instead of rerunning entire tasks.

Production-Grade Operators: Four Layers and 150+ Operator Matrix

LAS has accumulated over 150 operators covering video, audio, text, document, and image modalities, supporting embodied intelligence, autonomous driving, e-commerce marketing, and ad production. The operator service landscape has four layers:

Application Layer : Store inspection, financial report parsing, embodied intelligence pre-training, highlight clipping, e-commerce materials, ad generation, overseas expansion, etc.

Service Layer : Three entry points sharing the same capabilities—batch-processing API, CLI/Skills for Agent invocation, and real-time Studio WebUI.

Operator Matrix :

Video: understanding, editing, generation, repair, splitting.

Audio: transcription, splitting, language identification, denoising; ASR supports 99 languages .

Text & Document: 176 language identification , structured parsing, semantic splitting.

Image: generation, cropping, quality scoring.

Foundation Layer : Dataset management, compute queues, monitoring/alerting, security, billing. Decoupling lets upstream businesses ignore model differences while downstream systems stably consume results.

Quantity is not the goal; an operator reused across multiple businesses is more valuable than dozens sitting unused in docs. Reuse depends on stable contracts and actual processing results .

Three Unifications Defining a Production-Grade Operator

Unified I/O : Every operator defines input media, file limits, parameters, output structure, status, and error meanings. Example: video shot-detection operator outputs shots, characters, scenes, props, and timeline; training-data operator outputs traceable dataset versions and quality results. Upstream need not adapt to model differences; downstream stably consumes results.

Unified Calling Protocol : Short tasks use synchronous calls; long videos, batch documents, and generation tasks use async protocol. Tasks have traceable IDs, explicit states, idempotent semantics, failure reasons, retry support, and historical version retention—serving both low-latency online interaction and offline batch processing.

Unified Quality Standard : Getting the API running is not enough. Input side validates format, size, duration, and content usability. Processing logs model, parameters, versions, and intermediate results. Output side uses rules, model evaluation, or human review to judge pass/fail.

An operator is not just a model wrapper; it is a stable contract combining model capability, pre/post-processing, resource scheduling, and quality requirements into a reusable production unit.

Workflow Orchestration: Encoding Business Strategy into the Pipeline

With a solid operator system, workflow orchestration embeds business goals, quality standards, and exception handling into the processing chain:

Evaluation decides next step: pass → return; fail but repairable → locate problematic segment, adjust parameters, local retry; brand/safety/compliance undecidable → human review.

Human review conclusions feed back into evaluation standards, gradually reducing repeated audits.

Orchestration also optimizes compute: format conversion, splitting, frame extraction, text cleaning run on CPU; generation, understanding, enhancement run on GPU. System allocates resources per node, controls concurrency, splits batches, concentrating expensive compute on inference-heavy steps. LAS's AI-native distributed compute engine supports batch concurrency; single operator handles thousands of concurrent tasks. Beyond concurrency numbers, what matters is failure traceability, task recoverability, and cost controllability.

Three Real-World Cases Showing Scale in Action

1. Embodied Intelligence Training Data Processing: Ego Video into Lance, VLA Model Reads Directly for Training

Training data processing is a typical high-throughput, long-cycle, multi-modal mixed task. Large data volume, many stages, high failure/rewrite cost—ideal for testing "production-ready" operator systems. Raw data becomes training assets through four layers working together:

Application layer targets pre-training corpora, multi-modal understanding, video generation, speech/audio, image-text documents.

Standard operator layer provides text cleaning/deduplication, video understanding/generation, speech processing, document parsing, semantic splitting, quality assessment, detection/estimation, language identification, image processing.

Compute orchestration layer handles batch job submission, CPU/GPU tiered scheduling, concurrency and batch splitting, failure retry and recovery, dataset version management, lineage and run records.

Data asset layer powered by Lance achieves image-text and vector in same table, incremental label columns with version rollback, zero-copy random read and direct training read.

In embodied intelligence, operators first transform first-person or multi-camera ego video into structured annotations: captions, event states, action sequences, object trajectories. Lance then stores video, annotations, and vectors in one table, using incremental columns, version rollback, and training direct-read to turn annotations into continuously iterable training assets. Annotations undergo consistency and confidence scoring; low-confidence and difficult samples enter human review, forming a "pre-label → review → re-check → loop" human-in-the-loop cycle. In this scenario, core annotation task accuracy reached over 90% , and annotation-to-delivery cycle shortened from month-level to week-level .

2. E-Commerce Viral Material Replication: Reuse Structure, Not Pixels

"Viral" refers to breakout short videos on platforms. The pipeline extracts shot structure, composition, script rhythm, and transition logic from historical viral videos, preserves the narrative skeleton, and swaps products, characters, and backgrounds to quickly generate differentiated placement versions for multiple SKUs, platforms, and accounts. The chain has six steps:

Viral Deconstruction : Fine-grained video understanding and shot-detection operators identify subject, composition, shot order, selling-point expression, and rhythm, extracting transferable structured scripts.

Prompt Generation : Convert fragmented creativity into structured instructions adapted to video models, or directly produce shot scripts ready for generation.

Material Pairing : Build batch task matrix linking source video with multiple product images, character images, background images; one configuration generates multiple versions.

Generation & Editing : Video editing enhancement replaces subject, character, background; outputs different aspect ratios, resolutions, platform specs; or generates video directly from shot scripts.

Evaluation & Repair : Full quality check on replacement boundaries, temporal consistency, frame quality; failed results auto-rewrite parameters and retry by problem type.

Versioning & Placement : Retain candidate versions, quality tags, run records; qualified materials flow directly into placement and performance feedback.

After deployment: material production efficiency up 100x+, manual cost down 80%+, lottery success rate up 50%+ . More important than numbers is organizational change—every team member shifts from repetitive labor to "strategist," refocusing on product selection, creativity, and placement analysis. The speaker summarized: "What's truly copied is not pixels, but the viral's structured expression; what truly boosts capacity is not a single generation, but the complete closed loop of deconstruct, generate, evaluate, repair, place."

3. Video Ad Placement: Connecting Long-Form Content Across Generation and Distribution

The challenge in video ad placement is not single-shot quality but stably filtering truly high-converting highlight clips from hundreds or thousands of materials, and rapidly producing multi-language, multi-platform versions. LAS strings content production, highlight extraction, placement material generation, and overseas localization into one chain: first redraw, shot-detect, splice, dub; then identify emotional peaks, conflict points, plot hooks; batch-generate versions by platform duration and aspect ratio; finally transcribe, translate, subtitle, and distribute in multiple languages. This flow delivers 10x+ single-episode production efficiency , minute-level per-material generation , placement material success rate up 50%+ , supports 30+ languages , flexibly adapting to diverse overseas needs.

Model + Operator + Orchestration = Scale

Two years of practice yield one insight: even the strongest model, if every business integration requires rewriting adapter code, cannot become an organizational asset that is continuously reusable. Models provide capabilities; operators provide stable contracts that business systems can depend on; orchestration embeds validation, evaluation, retry, and human confirmation strategies into the pipeline. Together they sustain large-scale production.

Scale cannot be judged by "feels faster." Look at first-pass rate and rework rate, effective throughput under batch processing, cost per usable output, and whether end-to-end delivery cycle truly improves. A single great generation result ≠ stable batch capacity; lower per-call price ≠ lower cost per usable material. Repeated retries, full regenerations, and manual rework all count toward real production cost.

Moving from "capability usable" to "production usable" requires an operator system that can be continuously reused, composed, measured, and evolved. Letting every model improvement convert faster, more stably, and at lower cost into business value is the problem LAS's multimodal operator system aims to solve.