Every staffing company in India now has an “AI Hiring” page.
Every one of them claims they can find AI engineers. Most of them couldn’t tell you the difference between a fine-tuned LLM and a random forest classifier. They search “AI” on LinkedIn, find profiles that mention “machine learning,” and send them to you with a note that says “strong AI candidate.”
I know this because the clients who come to Supersourcing for AI hiring have usually tried two or three agencies first. The story is always the same. The agency sent profiles. The profiles looked good on paper. Then the technical interview revealed that the “AI engineer” had trained a sentiment analysis model in a Jupyter notebook during a weekend course — and had never deployed a model to production, never managed an inference pipeline, never dealt with model drift, and never built a data pipeline to feed a production ML system.
The client wasted 3-4 weeks. The role is still open. The AI project is stalled.
This is the hardest hiring problem in technology right now. And I’m going to tell you exactly why — and exactly how we solve it.
My name is Mayank Pratap. I co-founded Supersourcing. Here’s the credential that matters most for this conversation.
Google selected us for the AI Accelerator 2024.
Let me explain what that means. Google didn’t select a staffing company for the AI Accelerator because we’re good at sending resumes. Google evaluated our production AI capabilities — the AI systems our sister company EngineerBabu builds and deploys daily. Credit scoring engines that make thousands of real-time decisions. Fraud detection systems that catch sophisticated attacks. Healthcare AI that predicts patient outcomes. Production AI. Not demos. Not proofs of concept. Systems that run in production, make real decisions, and have real consequences.
That production AI experience is what our assessment team uses to evaluate AI candidates. When the Supersourcing team assesses an ML engineer, they’re evaluating against the standard of what production AI actually looks like — because the EngineerBabu team builds it every day.
No other staffing company in India has this combination. A Google-validated AI capability powering a staffing operation led by Vijay Kiran — who built engineering hiring at Flipkart ($35 billion, 10,000+ engineers) and two other unicorn companies.
CMMI Level 5. LinkedIn Top Startup India — twice. Wipro, Virtusa, and Impetus as direct client partners. Vijay Shekhar Sharma — founder of Paytm — backs us personally.
The AI Hiring Crisis — Why Every Company Is Struggling
Here’s the uncomfortable math.
Every company wants AI. Few companies can find the engineers to build it.
The global demand for AI/ML engineers has grown over 400% since 2020. Senior ML engineer salaries in San Francisco have crossed $400,000 in total compensation. Even at that price, positions stay open for months. Companies compete against Google, Meta, OpenAI, Anthropic, and every funded AI startup for the same limited talent pool.
The generative AI wave made it worse. Since 2023, every company — not just tech companies — wants LLM engineers, prompt engineers, RAG architects, and AI agents builders. A bank in Dubai wants generative AI. A hospital in London wants clinical AI. A logistics company in Singapore wants route optimization AI. A retailer in Sydney wants demand forecasting AI.
The demand is universal. The supply is not.
India changes the math. India has the second-largest AI talent pool globally after the US. The IITs — globally ranked among the top 50 engineering institutions — produce some of the world’s strongest AI researchers. IISc Bangalore, IIIT Hyderabad, and ISI Kolkata are globally recognised for machine learning research. Indian AI researchers are disproportionately represented at top conferences — NeurIPS, ICML, ICLR, CVPR.
But there’s a difference between AI research talent and AI production talent. India has both. Most countries have neither.
The production AI talent exists because India’s own technology ecosystem demands it.
Flipkart’s recommendation engine processes millions of personalised recommendations daily. Razorpay’s fraud detection evaluates every transaction in real-time. Swiggy’s delivery optimization runs inference at city-scale. These systems require ML engineers who don’t just train models — they deploy, monitor, retrain, and operate them in production.
Vijay Kiran hired the engineers who built these systems at Flipkart. He knows what production AI talent looks like. Not from reading about it. From evaluating tens of thousands of engineers and selecting the ones who could build AI at India’s most demanding scale.
Why AI Hiring Is the Hardest Hiring Problem in Technology
I’ve placed engineers across 80+ technology specialisations. SAP, Salesforce, data engineering, cloud, DevOps, full-stack, mobile. AI is the hardest. By far. Here’s why.
The Gap Between Notebooks and Production Is a Canyon
A data scientist who trains a model in a Jupyter notebook has completed maybe 10% of the work required to deploy AI in production. The remaining 90% is engineering.
Data pipelines that feed the model with clean, validated, correctly formatted data. Feature stores that serve the same features consistently across training and inference. Model serving infrastructure that handles thousands of predictions per second with sub-100ms latency. Monitoring systems that detect when the model’s accuracy degrades. Retraining pipelines that update the model without disrupting production. A/B testing infrastructure that compares the new model against the incumbent. Versioning systems that track which model is deployed where and when.
Most “AI engineers” on the market have notebook skills. They can train a model. They can tune hyperparameters. They can achieve impressive accuracy on a test set. They cannot deploy that model to production, keep it running, and ensure it continues to perform as the world changes around it.
The Supersourcing assessment separates notebook AI from production AI. Because the EngineerBabu sister company builds production AI daily — deploying models that make real credit decisions, real fraud assessments, real clinical predictions — the assessment team knows exactly what questions to ask. Not “explain backpropagation.” Rather: “your production model’s accuracy dropped 3% this week. Walk me through your debugging process.”
The candidate who’s only worked in notebooks can’t answer that. The candidate who’s operated production AI answers it from experience.
The Specialisation Explosion
“AI engineer” is as meaningless as “doctor” without a specialisation.
A Natural Language Processing engineer works with text — transformers, attention mechanisms, tokenisation, embedding models. A Computer Vision engineer works with images — CNNs, object detection, image segmentation, video analysis. A Generative AI engineer works with LLMs — fine-tuning, RAG architectures, prompt engineering, agent frameworks. An MLOps engineer builds the infrastructure that makes all of the above work in production. A Reinforcement Learning engineer designs systems that learn from interaction — robotics, game AI, recommendation systems.
These are different roles. Different skills. Different assessment criteria. A brilliant NLP engineer might know nothing about computer vision. A strong MLOps engineer might never have trained a model. A generative AI specialist might struggle with classical ML techniques.
Most staffing agencies search “AI/ML” and send whatever appears. The Supersourcing team assesses for the specific AI specialisation the client needs. A Generative AI assessment evaluates LLM fine-tuning, RAG pipeline design, hallucination mitigation, and token cost optimization. A Computer Vision assessment evaluates model architecture, training data pipeline design, and edge deployment. Completely different evaluations for the same “AI engineer” job title.
The Hype-to-Skill Ratio Is Broken
Since ChatGPT launched, hundreds of thousands of professionals have added “AI” to their profiles. Many took a weekend course. Some completed a Coursera specialisation. A few built a chatbot using the OpenAI API and now call themselves “AI engineers.”
The real AI engineers — the ones who’ve built production ML systems at companies like Flipkart, Google, Amazon, Razorpay — are buried under the noise. Finding them requires an assessment that can distinguish between someone who calls the ChatGPT API and someone who fine-tunes a model on proprietary data, builds the inference pipeline, deploys it with auto-scaling, monitors drift, and retrains when performance degrades.
Google AI Accelerator selection is the credential that separates our assessment from every other agency’s. Google evaluated production AI capability. Not API usage. Not demo building. Production systems. That standard is what the assessment team applies to every AI candidate.
How Vijay Kiran Assesses AI Talent — And Why It’s Different
Most agencies assess AI candidates the way they assess software engineers. Coding test. System design question. Behavioral interview. Done.
That framework misses everything that matters in AI hiring.
Vijay Kiran built the engineering hiring operation at Flipkart. Flipkart’s AI systems — recommendation engines, search ranking, demand forecasting, fraud detection, pricing optimization — are among the most sophisticated in Asia. The engineers who built these systems were hired through Vijay’s process. That process evaluates AI engineers on four dimensions that generic assessments miss.
Production thinking. The assessment presents real-world deployment scenarios. “Your model needs to serve 10,000 predictions per second with sub-50ms latency. The model is 2GB. Your inference servers have 16GB RAM. Walk me through your approach.” A notebook-trained candidate freezes. A production-trained candidate starts talking about model optimization, batching, quantization, and caching strategies — because they’ve solved this before.
Failure reasoning. “Your production model’s precision dropped from 92% to 84% over the last two weeks. What could cause this? What’s your investigation process?” This question reveals whether the candidate has operated a production model or only trained one. Drift detection, data quality investigation, feature distribution analysis, upstream data source changes — production AI engineers have a mental checklist. Notebook engineers don’t.
Architecture judgment. “The business wants real-time personalisation for 5 million users. Do you build a real-time inference system or a batch prediction system with caching? Why?” There’s no single right answer. The question reveals how the candidate thinks about trade-offs — latency versus cost, accuracy versus speed, complexity versus maintainability. This is architecture thinking applied to AI — and it’s the skill that determines whether an AI system works in production or collapses under real-world constraints.
Communication clarity. AI engineers work with product managers who don’t know what embeddings are, with executives who want to know when the AI will be “ready,” and with data teams who need to understand input requirements. The ability to explain AI concepts without jargon — and to set realistic expectations about what AI can and cannot do — is assessed explicitly.
What India’s AI Ecosystem Looks Like in 2026
India isn’t just a source of AI talent. India is building its own AI future.
The Indian government’s IndiaAI mission is investing ₹10,000+ crore in AI infrastructure, research, and compute capacity. India has overtaken the UK in AI research output. Indian AI startups raised over $3 billion in 2025 alone.
But what matters for hiring is the production AI experience base. India has it because India’s own companies demand it.
Flipkart’s recommendation system serves personalised results to 400+ million registered users. The ML engineering team that built and maintains it is world-class. Vijay Kiran hired many of them.
Razorpay processes payments for millions of businesses. Every transaction runs through fraud detection ML models in real-time. The engineers who built those models understand production AI at financial-services-grade reliability.
Swiggy’s delivery optimization uses reinforcement learning to route 2+ million daily deliveries. Zerodha’s trading platform uses ML for market surveillance. CRED uses ML for credit assessment. PhonePe’s payment intelligence runs inference at India-scale — which means handling more digital transactions than almost any country on Earth.
These companies have created a talent pipeline of ML engineers who’ve built, deployed, and operated AI at genuine scale. When they transition to new roles — whether at GCCs, startups, or through staffing partners like Supersourcing — they bring production experience that’s rare globally.
The cost advantage compounds the talent advantage. A senior ML engineer with production experience in San Francisco costs $350,000-$450,000/year in total compensation. The same calibre in India: $80,000-$140,000/year. The skills are equivalent. The systems they’ve built are equivalent in complexity. The cost difference is structural — driven by India’s cost of living, not by talent quality.
The Specific AI/ML Roles the Team Fills
Let me be precise. Not “we do AI hiring.” These specific roles, assessed against these specific standards.
AI Engineers — end-to-end AI system builders. Model training, feature engineering, inference pipeline design, API integration, production deployment. The generalists who can take an AI project from data to deployed system. Assessed for production deployment experience, not just model accuracy.
ML Engineers — focused on model development and optimization. Model architecture selection, hyperparameter tuning, training pipeline optimization, model compression for deployment. Assessed for model selection judgment — knowing when a simple XGBoost outperforms a deep learning model is as important as knowing how to build the deep learning model.
Generative AI Developers — LLM fine-tuning, RAG architecture, prompt engineering, AI agent frameworks (LangChain, LlamaIndex, CrewAI), vector databases (Pinecone, Weaviate, pgvector). The hottest specialisation in AI right now. Assessed for production RAG — not just calling an API, but building retrieval systems that are accurate, cost-efficient, and hallucination-resistant.
NLP Engineers — transformer architectures, text classification, named entity recognition, sentiment analysis, question answering, text generation. Assessed for real-world NLP — handling messy, multilingual, domain-specific text that’s nothing like clean benchmark datasets.
Computer Vision Engineers — CNNs, object detection (YOLO, RetinaNet), image segmentation, video analysis, OCR, medical imaging. Assessed for deployment constraints — mobile inference, edge computing, real-time video processing — because production computer vision is primarily an optimization problem.
MLOps Engineers — the infrastructure builders. ML pipelines (Kubeflow, MLflow, Vertex AI), model serving (TensorFlow Serving, Triton), experiment tracking, model monitoring, automated retraining, feature stores. Assessed for operational discipline — keeping ML systems running reliably is harder than building them initially.
Data Scientists — statistical analysis, experimentation design, A/B testing, business intelligence with ML augmentation. Assessed for business impact thinking — not just model metrics but “did this model change a business outcome?”
AI/ML Leads and Directors — technical leadership for AI teams. Architecture vision, team building, stakeholder management, AI strategy, vendor evaluation, build-vs-buy decisions. Vijay Kiran’s leadership assessment — built across three unicorn hiring operations — evaluates strategic AI thinking alongside technical depth.
For every role, the assessment is specialisation-specific. A generative AI evaluation looks nothing like a computer vision evaluation. An MLOps assessment tests different skills than an NLP assessment. This specificity is why clients come to Supersourcing after failing with agencies that assess all AI candidates with the same generic test.
How the Engagement Actually Works
Hour 0-2: AI-specific requirement mapping. Not “we need an AI engineer.” Rather: “we need an ML engineer who can build a real-time fraud detection system processing 50,000 transactions per minute with sub-100ms inference latency, deployed on AWS SageMaker, integrated with our Kafka event stream.” That level of specificity shapes the entire sourcing and assessment process.
Mayank or a senior team member leads this conversation. Not a recruiter reading from a template. A founder who runs both a staffing company and a product engineering company that builds production AI.
Hours 2-48: AI ecosystem sourcing. The team sources through AI research communities, ML engineering forums, Kaggle contributor networks, open-source ML project contributors, and AI-powered matching. Not a LinkedIn search for “machine learning.” Targeted outreach to professionals whose production AI experience matches the client’s specific requirements.
Google AI Accelerator 2024 powers the matching — identifying candidates whose experience patterns match the specific AI specialisation, deployment scale, and domain the client needs.
Hours 48-72: Specialisation-specific assessment. The real differentiator. A generative AI assessment presents RAG architecture challenges — retrieval accuracy trade-offs, chunking strategy decisions, hallucination mitigation approaches, cost optimization for LLM API calls at scale. A computer vision assessment presents real-world deployment constraints — model size versus accuracy, edge inference optimization, training data quality management. An MLOps assessment presents production reliability scenarios — model serving failures, drift detection, pipeline debugging.
Vijay Kiran’s framework governs the quality bar. Production thinking. Failure reasoning. Architecture judgment. Communication clarity. Four dimensions that separate production AI engineers from notebook-trained ones.
Hours 48-72: Delivery of 3-5 assessed profiles. Each with specialisation-specific assessment notes. Not “strong AI candidate.” Rather: “ML engineer with 4 years production experience, deployed real-time fraud detection on SageMaker serving 30K predictions/minute, reduced false positive rate from 12% to 3.5% through model architecture improvements, strong communicator, available in 2 weeks.”
Days 3-14: Interviews, offer, deployment. Full coordination. Compliance handled. CMMI Level 5 processes throughout.
Why Most AI Hiring Fails — The Three Patterns
Hiring notebooks, not production. The candidate built an impressive model in a Kaggle competition. The client hires them. Three months later, the model still isn’t in production — because the candidate doesn’t know how to build inference pipelines, manage model versioning, or handle production data that’s messier than competition datasets. The AI project is declared a failure. The real failure was the assessment.
Supersourcing’s assessment separates notebook capability from production capability. Every
AI candidate is evaluated on deployment experience, not just model-building skill. The EngineerBabu sister company builds production AI daily — that production standard is the assessment benchmark.
Specialisation mismatch. The client needs a generative AI engineer to build a RAG system. The agency sends a computer vision engineer who “also does NLP.” The skillsets are fundamentally different. The candidate struggles. The project stalls. The client blames AI for being overhyped when the real problem was a hiring mismatch.
The team’s specialisation-specific assessment prevents this. A generative AI evaluation is a fundamentally different assessment from a computer vision evaluation. If the candidate’s real strength is classical ML, the assessment catches it before the client sees the profile.
Confusing AI enthusiasm with AI capability. Since ChatGPT, everyone is excited about AI.
Excitement doesn’t equal capability. The candidate talks passionately about AI’s potential. They’ve read every paper. They attend every meetup. They have strong opinions about AGI timelines. They’ve never deployed a model to production.
Vijay Kiran’s assessment framework focuses on what the candidate has built, not what they believe about AI. “Show me a production system you deployed” separates the enthusiasts from the engineers. The filter works.
What Clients Get With Supersourcing for AI Hiring
Mayank Pratap leads every engagement. The client talks to a founder who runs both a staffing company and a production AI engineering company.
Vijay Kiran — Flipkart ($35B, 10,000+ engineers, one of Asia’s most sophisticated AI operations), three unicorn hiring operations — runs assessment. The AI quality bar is set by someone who hired the engineers that built recommendation engines serving 400 million users.
Google AI Accelerator 2024 — the only staffing company with this credential. Not a badge. A validation by Google of production AI capability.
Specialisation-specific assessment across every AI discipline: AI Engineering, ML Engineering, Generative AI, NLP, Computer Vision, MLOps, Data Science, and AI Leadership. Not generic “AI” screening.
48-72 hours to assessed shortlist. 12-14 days to deployed AI engineer. 30-day replacement guarantee on permanent placements.
CMMI Level 5 for enterprise procurement. LinkedIn Top Startup India — twice. Wipro, Virtusa, Impetus as vendor partners. Backed by Vijay Shekhar Sharma.
Every engagement model — remote contract, remote full-time, onsite, hybrid, leadership hiring. Contract for AI R&D exploration. Permanent for production AI teams. Leadership for Head of AI and VP ML roles.
Let’s Talk AI
If you need AI/ML engineers — generative AI developers, ML engineers, NLP specialists, computer vision engineers, MLOps engineers, or AI leadership — and your current approach is sending you “AI enthusiasts” instead of production AI engineers, email me.
[email protected]. The founder. The person who runs the only staffing company that Google selected for the AI Accelerator.
Tell me the AI specialisation, the deployment context, the scale requirements, and the timeline. The team will have 3-5 production-assessed, interview-ready profiles in your inbox within 48-72 hours.
No commitment. No contract. Just profiles. Assessed by the standard that Google validated. Mayank Pratap Co-founder, Supersourcing [email protected] | supersourcing.com
Google AI Accelerator 2024 · CMMI Level 5 · LinkedIn Top Startup India (Twice) · Talent Director: Ex-Flipkart, 3 Unicorn Hiring Ops · Wipro / Virtusa / Impetus Client Partner · Backed by Vijay Shekhar Sharma
