I don't believe in AI magic.
I believe in the mathematics. That isn't a tagline — it's a filter. It tells me which problems are worth solving, which architectures will survive production, and which AI initiatives are theatre dressed as transformation.
Architecture decisions I've owned in production:
- Zero-shot vs. per-series training: Chose cross-learning TFT so new SKUs were forecastable on day one without retraining — 90+ markets, 25K+ SKUs
- Retrieval-grounded LLM vs. fine-tuned: Insisted on RAG + deterministic validation layer before any figure reached a planner — zero hallucinated numbers in production
- State-aware async agent runtime: Designed checkpointing + self-recovery so agent fleets fail gracefully, not catastrophically
- Decoupled from vendor runtime: A2A + MCP abstraction layer — same workflow runs on Gemini, Bedrock, Vertex AI or Azure OpenAI without rewriting agent logic
- Governance as first-class citizen: Trust interceptors, RBAC, HITL escalation designed into the runtime — not added after the compliance team asks
How I got here
I came up through mathematics. A Boeing-sponsored PhD from IIT Bombay in Optimization & Operations Research — stochastic programming, large-scale combinatorial optimization, simulation under uncertainty — gave me something rare in this field: the ability to prove what I build, not just demo it. Before that, an M.Tech from NIT Tiruchirappalli and a B.Tech in Mechanical Engineering from WBUT.
Then quantitative finance — WorldQuant and a Geneva family office — where feedback is instant and brutal. A wrong model costs money the same day. It installed an evaluation discipline I never lost. Then production ML for a US credit card company, healthcare NLP for CVS Health inside a HIPAA environment, and nearly four years at Unilever.
At Unilever I didn't join an existing AI team. I built the entire forecasting and decision-intelligence function from scratch — hiring the 30-person org, defining the architecture, choosing the model strategy, managing P&L, and governing three Big 4 firms (EY, Deloitte, BCG) concurrently. The results — 4–5pp accuracy step-change, £2.5M+ revenue uplift, 20% fewer stockouts, 200+ daily users across 90+ markets — came from the mathematics. Not the slide deck.
Now at Cognizant I work on what comes after the model. I architect enterprise agentic AI platforms — autonomous agent fleets that are observable, self-healing and governed for GDPR / HIPAA / EU AI Act — including the semantic ontologies that let Google Gemini agents reason consistently across heterogeneous enterprise data, with multi-cloud orchestration across AWS Bedrock, GCP Vertex AI and Azure OpenAI.
What I care about
Most teams building agents are solving for capability. I'm solving for industrialisation. Trust and evaluation are the product, not an afterthought. The hard part is not the model — it's proving, to a regulator and a skeptic, when a system is right, when it is unsure, and when it should stop and ask a human.
I am wary of fluent chain-of-thought that only reads like reasoning. I design for the moment a model is confidently wrong — deterministic guardrails, human-in-the-loop escalation, audit trails, and evals that reflect the real distribution, not a curated demo. This is what governance-by-design means in practice.
What I've shipped and how
Zero-shot forecasting platform
TFT cross-learning model so new markets and SKUs are forecastable without retraining. Databricks feature store, MLflow, Airflow orchestration, Bayesian safety-stock, real-time inference into SAP IBP. Built end-to-end from scratch at Unilever — £2.5M+ uplift.
RAG-grounded LLM co-pilot
Hybrid retrieval + cross-encoder reranking, deterministic validation guardrail before any output reaches a user, HITL approval for plan actions. 200+ planners daily, zero hallucinated figures. The change management was the hard part, not the model.
Production agentic platform
Semantic ontologies for Gemini agent reasoning across heterogeneous enterprise data. A2A + MCP abstraction makes the workflow cloud-portable. State-aware async runtime with checkpointing, RBAC, HITL escalation and full audit trails.
Clinical NLP pipeline
BERT/GPT NER + ICD-10 classification inside a HIPAA environment. AUCPR-validated from the start, full audit trail per prediction, human coder always able to trace a decision back to source text. −35% claims processing time at CVS Health.
Teaching & recognition
- ◆AIM 40 Under 40 (2026)Analytics India Magazine — India's highest AI industry recognition for enterprise impact
- ◆Mentor · Analytics VidhyaGuiding senior data scientists on production-grade RAG, LangGraph state machines and zero-defect AI infrastructure
- ◆100+ publications · 13K+ followersMedium, Substack, Towards AI as DrSwarnenduAI — agentic AI, LLMOps, production ML, forecasting mathematics
- ◆Keynote speakerMachine Learning Developers Summit (MLDS) · International Forecasting Forum
Background, briefly
Education — PhD, Mechanical Engineering & Operations Research (Optimization), IIT Bombay (Boeing-sponsored, 2014–2019) · M.Tech, Production Engineering, NIT Tiruchirappalli (2011–2013) · B.Tech, Mechanical Engineering, WBUT Kolkata · MicroMasters, Supply Chain Analytics, MITx on edX · EPAT (Algorithmic Trading), QuantInsti
Certifications — Supply Chain Design · Supply Chain Management Tools & Techniques · Supply Chain Dynamics · CSCA01 Supply Chain Planning & Strategy (edX)
Core skills — Agentic AI · LLM Orchestration · Multi-agent systems · MCP · A2A · RAG · Temporal Fusion Transformer · Bayesian optimization · Monte Carlo simulation · Stochastic programming · MLflow · Databricks · Airflow · SAP IBP · AWS Bedrock · GCP Vertex AI · Azure OpenAI · GDPR · HIPAA · EU AI Act
Languages — English · Hindi · Bengali
Location — Bengaluru, India · Open to relocation (global) · Open to remote
Contact — swarnenduiitb@gmail.com · +91 88792 20084