25+ years building high-throughput, low-latency platforms for global clients — order management, payments and real-time analytics at national scale. Now working on AI agents, retrieval systems and streaming infrastructure, with every claim backed by a repository you can read and a demo you can run.
Twenty-five years across domains that share almost no vocabulary — investment banking at Goldman Sachs, insurance at Wellpoint, networking at Cisco, retail for H&M in Stockholm and Maplin in London, government integration for the UK Post Office and DWP in Preston and Manchester, headless commerce on Hybris for Sodimac in Chile, and customs clearance on mainframe in Cochin. From COBOL and mainframe through J2EE, microservices, streaming and cloud-native platforms — the domains changed; the engineering problems rhymed.
Since 2023 I have re-tooled deliberately for the AI era — certified in Oracle and Databricks Generative AI and Oracle AI Vector Search, alongside a BS in Data Science at IIT Madras. My interest is not in prompt engineering but in the systems layer underneath agents: retrieval that cites its sources, control flow that can be reasoned about before it runs, and inference infrastructure that behaves under load.
That combination is my edge. Agentic systems fail for the same reasons distributed systems have always failed — partial failure, unbounded work, silent truncation, state that drifts from reality. Having debugged those at scale for two decades is more useful than any framework.
Goal-directed agent planning, tool orchestration, Oracle & Databricks Generative AI, Oracle AI Vector Search, n8n, KNIME.
RAG pipelines with source citation, embedding models and their failure modes, vector stores, corpus design and provenance.
OLAP/OLTP at 5G scale — Druid, SingleStore, MongoDB, Spark Scala, Hadoop; query performance and index design.
Apache Kafka — multi-broker clusters, TLS and SASL/SCRAM, ACLs, replication, rolling restarts and certificate rotation.
AWS (Certified Solutions Architect), GCP, Docker, Ansible (Red Hat Certified), GPU inference hosting, CI/CD and monitoring.
Java/Spring, GoLang, Python, React, Node.js — cloud-native microservices, API design, and delivery leadership for 100+ engineers.
Each of these was built independently — to explore a problem or to demonstrate a capability, not as a delivered client engagement. Where there is a live demo, it is a real deployment you can drive yourself.
A production-shaped RAG system over a licensed primary-source archive: a five-stage ingestion pipeline (fetch → normalise → chunk → embed → serve) processing 3,014 documents and 106.2M characters into a 153,856-vector index, served through a FastAPI WebSocket API with a browser client.
Architecture decisions that make it operable, not a notebook: indexing is checkpointed and resumable, with chunks deterministically shuffled under a fixed seed so any interrupted run still yields a representative sample rather than one skewed era. Chunk sizing is bounded to the embedding model's real token window after discovering it truncates silently at 256 tokens — a failure that corrupts an index with no error raised. Retrieval (~30 ms) is streamed to the client ahead of generation (seconds), so the response is useful before the model is called.
Governance and auditability are enforced in the pipeline, not the prompt. Every answer carries source citations. Synthesis is constrained to retrieved passages and forbidden from first person, so the system never speaks as the author. Fine-tuning was rejected on record: it teaches cadence, not knowledge, and produces unfalsifiable fabrications. Ingestion honours robots.txt crawl-delay, uses only the rights-holder's published API, and applies provenance rules that exclude repudiated material.
A listing endpoint timing out on a $lookup over millions of documents that
filtered after the join, compounded by $skip pagination that
regenerates every preceding page. Reproduced against 4M orders and 2M customers with the
WiredTiger cache constrained to hold a 7:1 data-to-cache ratio, then four strategies
benchmarked. Page 500: >11 minutes → 10 ms. Documents examined per document
returned: millions → 1. The recommendation was explicitly not to upgrade the tier.
A three-broker Apache Kafka 4.1.2 KRaft cluster provisioned by five idempotent Ansible roles, deployed and verified on a live VPS. A message sent through the browser demo crosses a TLS + SASL/SCRAM-SHA-512 listener, replicates across all three brokers, and is read back by a consumer group. Includes zero-downtime rolling restart and certificate rotation playbooks, ACL enforcement, and the KRaft controller-listener trap documented in the README.
An agent built on Embabel 1.0.0 that treats inbound triage as a goal-planning problem.
Plan mode derives what the agent would do — extractLead → enrich → draft → send
— with no LLM call and no cost, by searching achievable plans over declared preconditions.
Drop confidence below 0.75 or mark the mail as spam and send becomes unreachable,
leaving only escalate. No branching code was written to produce that behaviour.
The interesting half is free precisely because it is not inference.
A conversational voice stack: faster-whisper large-v3-turbo for speech recognition served on CUDA with int8/float16 quantisation, Kokoro for speech synthesis, and Pipecat orchestrating the turn loop. Each service is a standalone HTTP endpoint verified independently before the pipeline is wired together — the only way to attribute latency honestly. Includes silence-probability gating so background noise is not transcribed as speech.
Independent engineering work across AI agent design, RAG systems and Kafka infrastructure, plus application development for a Dubai-based startup. Parallel deep-dive into generative AI, vector search and agentic planning, anchored by the IIT Madras BS Data Science programme.
Led ten scrum teams (100+ engineers) delivering six 5G OLAP/OLTP applications — real-time and batch, standalone and non-standalone modes. Owned SoW milestone delivery and client-director relationships, and drove the migration to real-time analytics on Druid.
Technical leadership for cloud-native, microservice-based platforms handling high-volume transaction routing and settlement — multi-party integration, message-based order flow, and payment rails with strict ordering and reconciliation guarantees.
Architected a multi-tenant SaaS product — real-time transaction monitoring, identity and access control, policy engines and risk scoring — and stood up its infrastructure and security posture from scratch.
Onshore technical manager for Falabella, Sodimac and Tutoos. Built the transaction processing system — gateway integration, tokenization, idempotent retry and reversal handling at retail scale — alongside headless commerce on SAP Hybris.
Delivered for Goldman Sachs, Prudential, UK Post Office (London), IKEA, H&M (Stockholm) and UK DWP (Preston) — distributed order management, high-reliability message integration and large-scale reporting platforms.
Transaction integration, banking platform engineering, Cisco intranet applications, a US retirement-savings platform and mainframe customs clearance systems in Cochin — from COBOL to J2EE.
Integration Lead — UK Post Office
Integration Lead — DWP, UK Govt.
Microservices Architect — Falabella, Sodimac
Order Management Lead — H&M
My consulting practice — AI agents, retrieval systems and cloud infrastructure, delivered end-to-end from architecture through to production.
Visit AI Agentic Enterprises →Open to consulting engagements and leadership roles in cloud infrastructure, AI agent systems, retrieval and distributed data platforms.