- Artificial Intelligence
Enterprise AI in Production: What 6 Real Deployments Taught Us
Enterprise AI in Production: What 6 Real Deployments Taught Us
AI projects in production look very different from AI projects in a pitch deck. This post draws on six live Auriga IT deployments across national highway infrastructure, omnichannel retail, B2B e-commerce, FMCG distribution, education technology and sports, and what each one taught us about what actually works once the system has real users on it.
The 60-second version
- AI works. But not in the way most project briefs describe it. The technology is rarely the constraint.
- Data readiness is the real limiting factor. Every project with a slow start had a data preparation problem underneath it.
- Distribution beats sophistication. A model that works on WhatsApp outperforms a more powerful model sitting behind a dashboard nobody opens.
- Triage is not an afterthought. Knowing that something happened is only useful if the right person finds out fast enough to act.
- This is a portfolio of six live deployments: NHAI, Ferns N Petals, VoomiSupply, Badho.in, Class Saathi and Machaxi. Each case study is linked below.
None of the six projects below started with a question about which AI model to use. They all started with an operational problem that had grown too large or too fast for the existing system to handle.
That framing - operational problem first, AI layer second - is the common thread in what worked. Most AI projects that underdeliver are not underpowered by the AI. They are under-prepared by the infrastructure around it.
Computer vision for national highways: detection is not the hard part
Client: National Highways Authority of India (NHAI). A 13-module AI traffic management system deployed on the Delhi-Mumbai Expressway, inaugurated by PM Narendra Modi on 12 February 2023.
NHAI · Computer Vision for 50,000+ km of National Highways
NHAI needed automated monitoring across the Delhi-Mumbai Expressway. Human operators scanning hundreds of simultaneous camera feeds could not reliably catch a wrong-way vehicle at 2am. Auriga IT built a 13-module system using YOLO-based computer vision, IoT sensors, ANPR cameras and radar, all feeding a unified command dashboard. Incident reporting time dropped by more than 8 minutes.
Omnichannel AI support starts with a unified customer record, not a model
Client: Ferns N Petals. India's largest omnichannel gifting brand, with 400+ stores, 100+ countries and 99% PIN code coverage. Five disconnected support channels unified via CygnusAlpha.
Ferns N Petals · Unified AI Customer Support
FNP was running support across five channels simultaneously. Agents switched between five tools per query. Customers repeated their problem on every channel switch. The 400+ franchise stores were invisible to the central team. CygnusAlpha unified all five channels, added a WhatsApp self-service bot, and introduced three-tier routing by intent, complexity and emotional weight.
Semantic search across 1.5 million SKUs is a data quality problem first
Client: VoomiSupply, Pennsylvania, USA. A B2B HVAC marketplace with 1.5M+ SKUs on Shopify. Keyword search was replaced with a custom semantic engine, and 80,000+ product types were enriched automatically.
VoomiSupply · Semantic Search for 1.5 Million SKUs
VoomiSupply's keyword search could not interpret buyer intent. With 1.5M+ SKUs and incomplete HVAC attribute data, where SEER, BTU and tonnage were often missing, buyers were hitting dead ends on every specification query. The replacement was a custom semantic engine on PostgreSQL with pgvector, automated enrichment pipelines, Cygnus Copilot for admin, and Cygnus Insights for real-time analytics.
WhatsApp AI for B2B India: the channel matters more than the model
Client: Badho.in. India's B2B FMCG distribution platform, serving 800,000+ retailers, 10,000+ distributors and 1,000+ brands. Badho.AI is an LLM-powered WhatsApp chatbot built and deployed by Auriga IT.
Badho.in · AI WhatsApp Support for 800,000 Retailers
Before Badho.AI, support was informal WhatsApp chats with human agents. Queries about orders, wallet balances and payouts were slow and untracked. With 800,000+ retailers, the support load was unsustainable. Auriga IT built an LLM chatbot integrated with the WhatsApp Business API and Hasura GraphQL, giving users instant access to live platform data, 24/7, with no app required.
Adaptive learning AI: session frequency is a model performance variable
Client: TagHive Inc. (Samsung-funded). The Class Saathi app, with 500,000+ free MCQ quizzes, the Saathi AI adaptive engine, and a Flutter cross-platform build for Android, iOS and desktop.
Class Saathi · AI-Powered Personalised Learning Engine
TagHive needed an adaptive system that identifies student weak areas, adjusts question difficulty, and keeps students engaged long enough to improve outcomes. Saathi AI analyses performance after every session and adjusts the topic and difficulty distribution of practice quizzes automatically. The gamification layer - daily missions, badges, streaks - is integral, not a bolt-on. The AI only runs on sessions students actually complete.
Facial recognition attendance: onboarding is the bottleneck, not inference
Client: Machaxi. A sports and fitness platform operating across multiple centres. Facial recognition attendance built with Euclidean Distance algorithms and ML identity verification.
Machaxi · Facial Recognition Attendance System
As Machaxi's network grew, manual check-in became a daily bottleneck. Players re-registered at each new centre. Parents had no attendance visibility. Managers were drowning in data entry. Auriga IT built a facial recognition system in which existing users check in in 3-4 seconds, new users register in under 20 seconds, parents receive real-time notifications, and managers get automated daily reports with zero manual input.
What these six AI projects in production have in common
Six industries, six technology stacks. The underlying pattern is consistent enough to be worth naming.
The problem came before the AI
In every case, the brief started with an operational problem, not a technology. NHAI needed faster incident response. FNP needed agents to stop switching tools. The AI was the method, not the reason.
Data readiness was the real constraint
VoomiSupply's enrichment pipelines, Badho.in's Hasura GraphQL integration, Machaxi's face registration flow: every project had a data preparation phase that set the ceiling for what the AI could do.
Integration beats sophistication
A model connected to live data answering a query over WhatsApp outperforms a more powerful model behind a dashboard nobody opens. Distribution and integration are the multipliers.
The AI layer is rarely the hard part
Alert triage at NHAI. Channel unification at FNP. Onboarding UX at Machaxi. The difficult problems were operational or structural. Model performance was usually the easiest variable once the infrastructure was right.
Most AI projects that underdeliver are not underpowered by the AI. They are under-prepared by the infrastructure around it.
We now start every AI engagement with a data readiness audit before scoping any model or architecture work. We ask where the live data lives, who owns it, and how it gets out before we discuss what the AI will do with it. And we design for the channel the end user actually uses, not the one that is easiest to build on.
Frequently asked questions about AI in production
What does AI actually do in a highway traffic management system?
In Auriga IT's NHAI deployment, AI computer vision monitors hundreds of simultaneous CCTV feeds and detects wrong-way driving, stationary vehicles and congestion in real time, automatically generating alerts with incident type, location, severity and a live camera link within seconds. This reduced incident reporting time by over 8 minutes compared to manual monitoring across 50,000+ km of national highways.
Can WhatsApp replace a customer support app for B2B platforms in India?
For B2B distribution platforms in India, WhatsApp is often the right primary channel, not a workaround. Badho.AI serves 800,000+ retailers through WhatsApp with zero app dependency, delivering instant query resolution 24/7 via an LLM layer integrated with Hasura GraphQL and the Badho CRM as the live source of truth.
Why does semantic search perform better than keyword search for large catalogues?
Keyword search matches exact text. Semantic search maps buyer intent to the right product even when the query does not contain the exact product name. For VoomiSupply's 1.5M+ HVAC SKUs, "quiet 2-ton mini split for basement" returns zero results with keyword matching, but resolves correctly with a semantic engine on PostgreSQL with pgvector, because the engine understands what the buyer needs rather than only what words they used.
How long does facial recognition check-in take in a real deployment?
In Auriga IT's production deployment for Machaxi sports centres, existing users are identified and checked in in 3-4 seconds using Euclidean Distance algorithms and an ML identity library. New users register in under 20 seconds. Parents receive real-time notifications and centre managers get automated daily attendance reports with no manual input.
What industries has Auriga IT deployed AI in?
Auriga IT has shipped production AI across government infrastructure (NHAI national highways), omnichannel retail (Ferns N Petals), B2B e-commerce (VoomiSupply), FMCG distribution (Badho.in), education technology (TagHive / Class Saathi), and sports and fitness (Machaxi).
What is the difference between AI automation and AI agents in customer support?
AI automation handles rule-bound, deterministic workflows, like Badho.AI automatically resolving an order status query from a live database lookup. AI agents reason across multi-step tasks and contextual signals, like CygnusAlpha's routing engine reading intent, emotional weight and complexity to decide whether a message goes to automation, an AI-assisted agent, or a senior specialist.
What is the most common reason an enterprise AI project underdelivers?
In Auriga IT's experience across these six deployments, the most common cause is not model quality but data readiness. If the live data is incomplete, poorly described, or locked behind a system with no clean way to query it, the AI layer inherits that ceiling. Every project here with a slow start had a data preparation problem underneath it rather than a modelling problem.
Auriga IT builds AI that runs in production
If you have an operational problem and want to understand whether AI is the right tool, and what it would actually take to deploy, talk to the team that built these six systems.
Related content
Auriga: Leveling Up for Enterprise Growth!
Auriga’s journey began in 2010 crafting products for India’s [...]
