Orange · Gillette · Adani Power · GBM
Enterprise Conversational AI Portfolio
Over six years at Cedex Technologies, I led delivery of over 50 conversational AI projects for global enterprise brands across four major accounts — Orange (France Telecom), Gillette, Adani Power, and GBM — spanning every major conversational channel, multiple languages, and the full range of enterprise use cases from customer service automation to internal knowledge management.
Quick facts
- Client
- Orange · Gillette · Adani Power · GBM
- Role
- Project Manager
- Category
- Conversational AI · Enterprise
- Year
- 2016–2022
- Duration
- 6 years · 50+ projects
50+
Bots delivered
15+
Voice apps
4
Enterprise brands
6
Years of delivery
Background & context
Conversational AI was a rapidly evolving space between 2016 and 2022. Platforms changed, capabilities expanded, and enterprise expectations grew. What started as novelty chatbots on Facebook Messenger evolved into sophisticated multi-channel, multi-language conversational programmes integrated with CRM systems, knowledge bases, and live agent platforms. The ability to deliver consistently across this evolving landscape required both technical depth and programme discipline.
The challenge
What made this hard.
01
Platform fragmentation across enterprise channels
Each enterprise client had different channel requirements — some wanted Facebook Messenger, others WhatsApp Business, Alexa Skills, Google Actions, or web-embedded chat. Each channel had its own API, its own design constraints, and its own testing requirements. Building and maintaining expertise across all of them simultaneously was a significant organisational challenge.
02
Multi-language NLP at production quality
Orange required French and Arabic support. Gillette needed Hindi alongside English. Adani Power operated across multiple Indian languages. Each language required separate NLP training data, separate entity recognisers, and language-specific conversation design. Off-the-shelf multilingual models of the era performed inconsistently — quality required custom training in each language.
03
Integrating conversational AI into enterprise systems
Enterprise chatbots that cannot access real data are toys. The projects that delivered genuine business value all required deep integration with CRM systems, product databases, order management systems, and live agent platforms. Building and maintaining these integrations — across multiple clients, each with different backend systems — was a significant ongoing engineering challenge.
The approach
How we solved it.
Built a reusable conversation design framework
Rather than designing each bot from scratch, I developed a conversation design framework — a structured approach to intent taxonomy, entity design, fallback architecture, and persona definition — that could be applied consistently across projects. This reduced conversation design time significantly and improved quality consistency across the portfolio.
Platform-specific expertise centres
I structured the team into platform specialisations — Dialogflow/Google, IBM Watson, Alexa/Voicebot, and Rasa/Open-source — with cross-pollination through regular knowledge-sharing sessions. Clients benefited from deep platform expertise rather than generalist knowledge.
Continuous training pipeline for all deployed bots
Every bot in production had a monthly retraining cycle — reviewing unhandled queries, adding new training examples, and updating entity dictionaries. This continuous improvement approach meant bots got measurably better over time rather than stagnating post-launch.
QuizMaster — building for virality
The QuizMaster bot, built for general audience engagement, applied the same gamification principles learned from Filmykaant — daily questions, public leaderboards, social sharing. It reached Facebook Top 10 status and was selected for FBStart, validating the virality playbook across a different content vertical.
Impact & outcomes
What we delivered.
50+ production chatbots and 15+ voice apps delivered
Across six years and four major enterprise accounts, maintaining consistent quality and on-time delivery throughout.
QuizMaster recognised as Facebook Top 10 Chatbot
The second Cedex-built bot to reach Facebook Top 10 status, selected for the FBStart programme alongside Filmykaant — an unusual distinction for a single team.
Multi-language support across four languages
Production-quality conversational AI in English, Hindi, French, and Arabic — each with custom NLP training data and language-specific conversation design.
Orange Telecom: automated 40% of first-line customer queries
The customer service automation programme for Orange's French-speaking customer base deflected a significant proportion of routine queries from live agents, reducing cost per contact and improving resolution speed.
Tools & technologies
Lessons learned
What this taught me.
The most important skill in conversational AI delivery is expectation management. Clients who understand what the bot can and cannot do reliably will adopt it. Clients who expect it to handle everything will reject it when it fails on edge cases.
Reusable frameworks compound. The investment in a shared conversation design methodology paid back with every successive project — not just in delivery speed but in the ability to onboard new team members quickly.
Voice and text are fundamentally different design problems. Voice requires much shorter responses, much cleaner error recovery, and much more conservative scope. Teams that design voice bots like text bots consistently produce poor user experiences.
Next case study
HDFC Bank Developer Portal →