Grand View Research analysts value the global edge AI market at $24.91 billion for 2025, projecting it to reach $118.69 billion by 2033, with North America generating over 36% of that expenditure. While numerous edge AI initiatives succeed during controlled trials, maintaining identical performance across hundreds of live locations featuring varying operating environments, network quality, and lighting conditions requires substantial effort.
As a Senior Technical Project Manager at Anicca Data Science Solutions and a doctoral researcher in cyber engineering, Siddharth Pani has spent more than five years bridging that divide. He transitioned computer vision from theoretical concepts into practical application at McDonald’s drive-thru lanes, where his Enhanced Vehicle Detection (EVD) system currently operates across 120 U.S. sites. Furthermore, he has authored multiple peer-reviewed publications covering AI, cybersecurity, digital transformation, IoT, blockchain, and edge computing across critical industries.
We discussed the reality of the edge AI surge from the perspective of the servers, data centers, and restaurants where inference takes place, focusing on what differentiates a lasting deployment from one that is quietly deactivated.
Siddharth, forecasts often portray edge AI as inevitable and effortless. Having deployed a vehicle-detection system across 120 American McDonald’s drive-thru locations, what do you believe market optimism overlooks?
Market forecasts measure financial expenditure rather than technical difficulty. A model functioning inside a controlled laboratory environment differs fundamentally from one operating across 120 commercial restaurants. Labs offer stable camera angles, lighting, and network connections. Deploying that exact model into an active drive-thru exposes it to rain-obscured lenses, cameras physically displaced overnight, and network dropouts during peak meal rushes. Our objective extended far beyond achieving high test-set accuracy. Trimming three seconds from order processing times and boosting vehicle detection by approximately 20% only held value if those metrics remained consistent across structurally unique locations. Ultimately, financial investment follows the resolution of those operational challenges rather than preceding them.
The gap between testing and practical implementation is a challenge you faced prior to working with drive-thru cameras. During your time at IBM, you developed an automation tool that decreased daily manual tasks for your team by 20%. How did a career trajectory originating in financial infrastructure transition into computer vision for restaurants?
Working with payment systems revealed the unglamorous side of engineering. Financial settlements must be accurate without exception, and that strict demand for reliability shaped my professional philosophy. Transitioning to Anicca in 2021 and subsequently managing the Microsoft and McDonald’s accounts involved a complete shift in domain. However, the fundamental engineering question remained unchanged: can this system operate autonomously at scale without constant human oversight?
That same mindset influenced AniccaVision, the broader computer-vision platform you spearheaded for McDonald’s. Deployed across 21 restaurants, it handles customer-journey analytics, order accuracy, labor management, and drive-thru merge points. While many computer-vision initiatives transmit video feeds to the cloud for processing, you kept computation local to each restaurant and collaborated with NVIDIA to run multiple camera streams on a single server, reducing hardware expenditures by roughly 30%. Why was localized intelligence worth that complexity?
The decision was driven by latency and cost considerations. Drive-thru decisions require execution in well under a second, a threshold cloud round trips cannot guarantee during network congestion. Furthermore, uploading raw video feeds from every camera to the cloud incurs steep, continuously scaling expenses as store counts grow. Retaining processing locally on an onsite server resolves both hurdles. The primary challenge involved hardware constraints; through our collaboration with NVIDIA, we optimized the models to enable multiple camera feeds to share a single machine instead of requiring dedicated hardware per camera, cutting equipment costs by approximately 30%.
Regarding AniccaVision, you managed the automated deployment pipeline designed to scale the platform beyond its initial 21-restaurant test phase. This initiative was subsequently showcased at Embedded World in Germany, where Microsoft designated Anicca an official AKS Edge Essentials partner. From a non-engineering perspective, what did achieving that milestone entail?
It required mastering nearly every variable outside the core model itself. Prototypes rely on engineers being readily accessible, whereas scaled rollouts assume zero on-site technical support. Consequently, onboarding new locations had to be nearly autonomous, and monitoring systems needed to identify issues before onsite staff became aware of them. Collaborating with Microsoft on that deployment pipeline provided the necessary technical foundation, while our presence at Embedded World aligned our architecture directly with the platform it supports. Industry recognition is gratifying, but systemic repeatability is what drives commercial viability.
Today, you oversee Anicca’s Microsoft account. From late 2023 through mid-2025, you directed a 15-person team managing the Azure data-center expansion, overseeing GB200 onboarding alongside data-science and analytics tasks that identify infrastructure gaps for stakeholders. The team, currently numbering around ten, automates hardware acceptance and configuration, enhances fault detection, and maintains a 99.99% reliability standard. Although distinct from a restaurant drive-thru, what principles translate between these environments?
The operational stakes appear different, but the core engineering instincts are identical. Data-center expansions penalize minor errors severely, as a single faulty configuration file or misdiagnosed hardware failure can cascade across the entire system. To prevent this, we automated configuration generation and implemented robust error-notification and retry mechanisms, eliminating the need to babysit repetitive processes. Edge AI in restaurants presents analogous risks, substituting servers with cameras. Both domains reward a singular mindset: assume that rare failures will manifest at the most critical moment, and architect systems so that errors cost minutes rather than days.
Your academic research and delivery work appear closely aligned. Your published literature suggests that keeping inference directly on edge devices minimizes both the exposed attack surface and transmitted data volumes. Does this research actively influence your deployment practices, or do they operate independently?
They influence each other significantly more than people might expect. Writing those papers forced me to formalize principles I previously treated as workplace intuition. Selecting an onsite server over a cloud-based round trip in a restaurant setting stems from identical logic: reduced data in transit and a smaller defensive perimeter. Ultimately, the security rationale and the deployment rationale represent two expressions of the same principle.
As a Senior Member of the IEEE—the world’s largest technical professional organization—you were invited to review submissions for the IEEE APSIT 2025 conference. Evaluating the work of fellow engineers requires distinct professional standing. Does exposure to external engineering perspectives shape your approach to deployment challenges?
It has influenced my work more than anticipated. Engagement with the IEEE connects me to discussions outside my immediate projects, which is vital in a fast-moving field like edge AI. Reviewing for APSIT offered a different perspective, exposing where engineers explicitly state their operational assumptions and where they omit them. In practical deployments, those unstated assumptions frequently cause field failures. Consequently, this experience refined how I document assumptions on my own projects, particularly concerning environmental variability.
You have established a notable career advancing sophisticated technologies from theoretical concepts to scaled implementations. As edge AI expenditure approaches $100 billion in the coming years, what should investing companies anticipate most?
My candid perspective points to the unglamorous foundational work that rarely appears in corporate presentations. Capital will continue following algorithmic models, but enduring projects will be those that prioritized monitoring, field reliability, and disciplined deployment from day one. Personally, my goals include translating the frameworks established across Microsoft, IBM, and other projects into reusable assets for other teams, completing my doctoral research in cyber engineering, and advancing ongoing studies. If I could offer a single recommendation to engineers entering this domain, it would be to embrace the unglamorous 20% of the work that commences after the demonstration concludes. Ultimately, that is where both true value and operational difficulty reside.
