I build AI that survives production —
not demos that survive a slide deck.
Boeing-sponsored PhD in stochastic optimization. I designed and shipped Unilever's zero-shot TFT forecasting platform from scratch — £2.5M+ revenue impact across 90+ markets and 25K+ SKUs. Now architecting production agentic platforms on Google Gemini at Cognizant. The math, not the hype.
Built AI at
"I don't believe in AI magic. I believe in the mathematics. It tells me which architectures survive production — and which are just theatre dressed as transformation."
Most teams solve for capability. I solve for industrialisation.
Systems that 200+ planners use every day, portable across AWS Bedrock, GCP Vertex AI and Azure OpenAI, governed for GDPR / HIPAA / EU AI Act. The decisions I care about: zero-shot vs. per-series training, state-aware vs. stateless agent runtimes, retrieval-grounded vs. parametric generation. I design to prove what I build — not just demo it.
Production agentic AI
Multi-agent orchestration, semantic ontologies for Gemini-scale deployments, LLM routing and MCP — with state-aware async runtimes, checkpointing and self-recovery.
Governance-by-design
Trust interceptors, RBAC, HITL escalation and deterministic validation guardrails built into execution — not retrofitted after the audit. Compliance is architecture, not paperwork.
Quant-grade forecasting
Temporal Fusion Transformer zero-shot forecasting, Bayesian optimization and Monte Carlo stress-testing — rooted in the stochastic programming from my Boeing-sponsored PhD.
The mathematics behind what actually works in production
I write about what breaks when you move AI from notebook to production — agentic systems, LLMOps, forecasting and the governance that keeps them honest. No hype. Just the math.
GPT-4 Is Just a Giant Markov Chain — And That's the Genius of It
Why statistical structure, not "intelligence", explains what large models can and cannot do. Start here to understand the math.
Forecasting · Production · 7 minStop Forcing Your Time-Series Models to Memorize the Universe
Why zero-shot, retrieval-grounded forecasting beats memorising every series — the thinking behind RAFT-style architectures.
Gradient Boosting · Math-first · 6 minWhy CatBoost Mathematically Outsmarts XGBoost
A first-principles tour of gradient boosting — ordered target statistics, and when LightGBM, XGBoost and CatBoost each win.
Foundations · Running thesis · GitHub ★22AI, Mathematically Explained
Modern AI is classical mathematics with better computers. A growing repository and essay series proving it from first principles.
Four systems. Four hard numbers.
Zero-shot demand forecasting across 30+ markets
Temporal Fusion Transformer platform under data sparsity, 25K+ SKUs. +4–5pp accuracy · £2.5M+ revenue · −20% stockouts · −50% manual effort.
Portable, governed agentic AI platform
Semantic ontologies and context substrates for Gemini agent fleets — self-healing, EU AI Act governed, no vendor lock-in.
RAG-grounded supply-chain co-pilot
Natural-language planning for 200+ planners. First LLM deployment at Unilever. Zero ungrounded numbers surfaced to users.
Clinical NLP for ICD-10 coding
BERT/GPT document intelligence with compliance-grade validation and full audit trails. −35% claims processing time.
From people I've shipped with
"Swarnendu is made for analytics. He is humble, eager to learn and open to listening, even though he already has very in-depth knowledge of the field. On top of all that, he is hard-working and very good at Python coding. He is a great asset to any analytics team."
Suvendu Tripathy — Associate Partner, Cognizant · Aug 2022
"I was constantly amazed not only by his technical skills, but by how well he communicated them to business users, and his ability to connect the dots to surface real insights. I couldn't recommend the experience more."
Mishtert T — Enterprise AI Leader, Cognizant · Jul 2022
Three role-targeted CVs — pick the role you're hiring for
ATS-friendly, text-selectable PDFs. Each is tailored in headline, profile and skills-order. Hiring for a strategy role, a delivery role, or a data & analytics function? One of these fits.
Director · AI Leadership
Enterprise AI strategy, agentic platforms, multi-cloud orchestration, AI governance and board-level influence.
Director · AI & Data Science Delivery
Hands-on delivery, production ML, LLM systems, MLOps and analytics execution at scale.
Head · Data & Analytics
Supply chain / manufacturing analytics, demand forecasting, data strategy and decision intelligence.