Service
Artificial Intelligence
Building a Smarter, Connected Tomorrow with AI-first and Agentic AI
Highlights
- Tata Consultancy Services (TCS) evaluated NVIDIA Nemotron 3.5 Super VL under early access for a multimodal field-audit workflow for telecom.
- Multimodal artificial intelligence (AI) transforms fragmented field evidence into reliable, evidence-based engineering assurance while maintaining human oversight.
- AI-powered telecom site audits improve audit accuracy, accelerate approvals, reduce manual effort, and minimize rework across network operations.
- By combining telecom domain expertise with multimodal AI, TCS enables scalable audit automation, stronger traceability, and consistent quality assurance.
On this page
Can multimodal AI automate telecom field audits?
Telecom field audits sit at the intersection of engineering intent and physical execution. Field operations teams capture photographs of cabinets, antennas, networking, and power supply units before and after a work order. Audit teams then determine whether the deployed site conforms to approved drawings, equipment schedules, and installation standards. The process is essential, but it is often manual, fragmented, and difficult to scale across modernization programs.
TCS evaluated NVIDIA Nemotron 3.5 Super VL for a multimodal field-audit workflow, benchmarking its performance directly against comparable models. The evaluation focused on whether a domain-adapted vision-language model could identify customer-specific equipment in field imagery, correlate visual evidence with engineering design documents, detect installation variances, and produce structured, traceable audit records.
The evaluation revealed two primary findings:
- Model capability and speed: As an open model deployed on dedicated infrastructure, NVIDIA Nemotron 3.5 Super VL delivered an 89.1% accuracy (compared to 73.9% for comparable models) with an average inference latency of 5.04 seconds per task versus 29.70 seconds for a comparable model setup—achieving about 5.9x latency reduction while significantly reducing reasoning error rates on complex schematics.
- Systemic trustworthiness: Model capability alone does not create a trustworthy audit process. Operational value emerged when multimodal reasoning was combined with customer-specific adaptation, authoritative catalog retrieval, deterministic validation, evidence references, confidence thresholds, and qualified human review.
The problem TCS set out to explore
A typical telecom site field operations workflow can include hundreds of images and multiple engineering design artifacts. Reviewers must identify equipment, read partially visible labels, reconcile naming differences, distinguish active assets from equipment marked for removal, and link downstream circuits to the correct parent cabinet. The evidence may be distributed across equipment schedules, layout drawings, AC/DC schematics, and installation photographs.
TCS focused the evaluation on a practical question: Could a multimodal model help engineers move from scattered-field evidence to a traceable first-pass audit without treating model output as absolute engineering truth?
The experiment tested a connected workflow rather than an isolated image-classification task:
- Recognize customer-specific equipment and relevant visual attributes in field imagery.
- Normalize observed equipment to an approved customer taxonomy and product catalog.
- Connect photographs, drawings, schedules, and site metadata.
- Compare the observed implementation with the approved engineering design.
- Return structured findings with evidence, confidence scores, and a clear route to human review.
Evaluation principle: Automate evidence collection and comparison while routing ambiguous, low-confidence findings to qualified engineering review.
Illustrative domains include radio and antenna systems, transmission and internet protocol (IP) networking, fiber and cabling, power and battery systems, cabinets, safety/compliance, and AC/DC power schematics.
How TCS structured the evaluation
TCS used Nemotron 3.5 Super VL as part of a broader audit system, not as a stand-alone model. The workflow was designed around three connected inputs: field imagery, approved engineering information, and site metadata (see Figure 1).
- Evidence ingestion and page routing- The pipeline ingests field operations photographs, cabinet interiors, nameplate crops, engineering drawing packs, equipment inventories, site metadata, and installation standards. Document processing recovers page text, layout regions, tables, and evidence coordinates. Page routing separates equipment schedules, layouts, AC/DC schematics, and other relevant page families so downstream extraction can apply the right rules.
- Domain adaptation for customer equipment- TCS adapted the model to target asset domains using automated labeling pipelines employing NVIDIA Physical AI Data Factory (PAIDF) assets. Training targets included equipment type, manufacturer, model family, canonical product ID, evidence spans, and confidence levels.
A critical design choice is to train the model to identify and normalize equipment rather than memorize transient file paths or document locations. Predicted product IDs are resolved through an authoritative catalog maintaining datasheets, manuals, installation guides, and revision history. This separation allows underlying documents to change without requiring model retraining. - Audit reasoning and validation- At inference time, the model returns ranked equipment candidates with confidence and supporting visual or optical character recognition (OCR) evidence. A normalization layer cleans aliases, validates part-number formats, and checks the customer taxonomy. The audit engine then compares the observed asset and configuration with the approved design, generating pass/fail status, missing-equipment alerts, installation deviations, and evidence references.
Comparative performance evaluation:
Nemotron 3.5 Super VL vs. comparable models: TCS conducted a comparative benchmark on the field cell site equipment and detailed design dataset to evaluate model performance across structural extraction, classification, and complex circuit reasoning tasks.
1. Overall metric comparison
Evaluation metric |
Comparable models |
NVIDIA Nemotron 3.5 Super VL |
Super VL advantage |
Accuracy (Micro) |
73.9% |
89.1% |
+15.2% |
Accuracy (Macro) |
74.7% |
82.2% |
+7.5% |
Precision |
73.9% |
89.1% |
+15.2% |
Recall |
73.9% |
89.1% |
+15.2% |
F1 Score (Micro) |
73.9% |
89.1% |
+15.2% |
F1 Score (Macro) |
74.7% |
82.2% |
+7.5% |
Parse success rate |
100.0% |
100.0% |
0.0% |
Average latency per task |
29.70 seconds |
5.04 seconds |
-24.66 sec (~5.9x faster) |
2. Breakdown by task difficulty
When categorizing tasks by complexity, Nemotron 3.5 Super VL demonstrated marked superiority on medium-difficulty tasks requiring higher-order multi-step reasoning across visual schematics.
Task difficulty |
Comparable model’s accuracy |
Nemotron 3.5 Super VL accuracy |
Accuracy Delta |
Easy |
71.1% |
85.6% |
+14.5% |
Medium |
80.5% |
97.6% |
+17.1% |
3. Key task level details
- Complex state classification: On tasks requiring precise visual grounding to determine equipment status and spatial topology for the active or isolated states of critical infrastructure, Nemotron 3.5 Super VL achieved perfect 100.0% accuracy, eliminating the visual hallucination errors seen with the comparable model (66.7%).
- Detailed sub-system identification: For identifying and validating specific infrastructure components such as power delivery and distribution subsystems, Nemotron reached up to 100.0% accuracy, providing a 20% to 22% improvement over the baseline model depending on the subsystem.
- Standard data extraction parity: For standard metadata extraction and highly structured documentation regions, both models demonstrated strong baseline capabilities, achieving parity at up to 100% accuracy.
What TCS learnt from the field evaluation
The evaluation surfaced five key lessons central to applying multimodal models in engineering operations:
- Image understanding alone is not enough: Connecting imagery with drawings, schedules, and site context is far more useful than treating each photograph as an isolated visual task.
- Domain adaptation drives operational relevance: Custom equipment imagery, OCR variants, and hard negatives were vital for teaching the model customer taxonomies and stable asset IDs.
- OCR and evidence quality shape outcomes: Preserving image regions and bounding boxes allows reviewers to distinguish model uncertainty from poor source photos.
- Authoritative retrieval should sit outside the model: A governed external catalog ensures datasheets and engineering specs remain current without needing model retraining.
- Human review is a control, not a fallback: Confidence thresholds and automated exception routing focus field engineers on high-ambiguity or safety-critical cases. Approved reviewer corrections feed directly into subsequent model-refinement cycles
Business impact and strategic road map
- Faster deployment acceptance: Automated evidence organization accelerates work-order approval timelines.
- Targeted engineering effort: Field engineers shift focus from manual image checking to resolving flagged exceptions.
- Lower remediation costs: Early variance detection prevents costly site revisits during rollout waves.
- Enhanced vendor governance: Clear evidence tracing provides objective auditing across third-party contractors.
Toward agentic field operations
The next phase moves toward an agentic compliance loop: Design à Deployment àAudit à Network digital twin. Connecting work orders, repository designs, and live network assets enables autonomous field operation audit, automated remediation recommendations, and seamless sync with digital twin systems at scale.
Foundation for evidence-rich compliance platform
The evaluation demonstrates that domain-adapted multimodal open models can significantly elevate field-audit workflows. By pairing NVIDIA Nemotron 3.5 Super VL with deterministic normalization, external retrieval, and human review controls, TCS achieved an 89.1% audit accuracy while reducing per-task processing latency to 5.04 seconds—outperforming frontier alternatives like comparable models in both speed and visual precision. This capability lays the foundation for scalable, evidence-rich compliance platforms across telecom and industrial infrastructure.