Ten years turning ambiguous business and regulatory requirements into governed, production-ready AI. I own solution architecture for enterprise compliance and document-intelligence systems — defining classification taxonomies, retrieval architectures, and LLM integration patterns, and naming the trade-offs that make them scale, survive an audit, and ship.
I think people hire architects for outcomes, not for buzzword bingo. So here’s the work, grouped by what it actually decides.
Translating ambiguous business and regulatory requirements into reference architectures. Technology selection, integration patterns, schema and API design, scalability, and production-readiness.
Designing and formally evaluating retrieval strategies — RAG, PageIndex, hybrid, reranking — with evaluation harnesses that make the approach-selection call defensible, not vibes-based.
Audit-ready systems with classification taxonomies, PII redaction, model risk assessment, and explainability — aligned to GDPR, EU AI Act, and enterprise security KPIs.
Reference implementations that transfer cleanly across AWS Bedrock, Azure AI, and GCP. Model lifecycle, drift detection, CI/CD, capacity planning, monitoring.
Four principles I bring to every system.
Every choice sacrifices something. I name what I gave up and why — that’s the line between picking a tool and architecting a system.
Compliance, auditability, and explainability are designed in from day one — not bolted on the week before the audit.
Formal quality gates and evaluation harnesses gate every prompt and retrieval change. Nothing ships on “looks good.”
Patterns that transfer cleanly across Bedrock, Azure, and GCP — so a cloud choice is never a one-way door.
Patterns proven across telecom, healthcare, pharmaceuticals, and retail.
Pulled from the resume. Each one has a working artifact behind it — happy to walk through eval results, retrieval configs, or the bits that almost shipped sideways.
“Claude reads 300+ pages of regulatory text and cites its sources.”
End-to-end production RAG for a Tier-1 enterprise engagement: Claude Sonnet on Bedrock, Qdrant for hybrid retrieval, FastAPI orchestration, RAGAS gates — the decision framework other compliance workflows now follow.
“Hundreds of asset classes, structured outputs, a multi-week process killed.”
LLM classifier on Claude/Bedrock for a regulated enterprise client: structured output validation, confidence scoring, rationale generation, and a human-in-the-loop review queue for audit defensibility. Replaced a multi-week manual categorization with near-real-time.
“A job-hunt copilot that actually reads the listings.”
Personal project, v0 global hackathon submission. Next.js + Neon Postgres + Apify scrapers on Vercel, with retrieval over my own work history and a small evaluator for fit. Currently being dogfooded on my own search.
“$10K and a year of validation.”
Winner of the AWS Global GenAI Hackathon. Claude on Bedrock with dynamic, context-aware question generation and RAG-grounded transcript evaluation against role competencies. Prompt orchestration, structured-output scoring, quality validation across both generated and evaluated content.
“A knowledge graph that explains itself to a pharmacist.”
Neo4j knowledge graph with LLM-based entity linking and a RAG retrieval interface for complex pharmacological queries, built for a large US healthcare deployment. Every response grounded to a source document with inline citation and auditor-visible provenance. Azure OpenAI + Azure AI Search.
“Multi-tenant B2B prospecting, wrapped in MCP.”
Multi-tenant B2B prospecting SaaS exposed as an MCP server. FastAPI + Anthropic MCP SDK. The clients pull leads through Claude — the agent does the boring part, the human does the call.
More builds, less depth. Same person.
The ingestion layer downstream compliance and retrieval workflows stand on.
LLM-light document-to-structured-content service. Powers downstream compliance, classification, and retrieval workflows for a Tier-1 enterprise. The unglamorous infrastructure that everything else stands on.
Compliance review across docs, slides, spreadsheets, and images.
Extended a text-only compliance pipeline to handle PPTX, XLSX, embedded images, and OLE objects using Claude Vision. The pipeline now reads anything an auditor would actually open.
Multi-agent RAG over personal LinkedIn graph data.
Ingests a LinkedIn connection graph, enriches it, and surfaces second-order paths and warm introductions. Pitched to the Vercel AI Accelerator. The “LinkedIn GPT for individuals” idea.
Conversational restaurant booking over the phone.
Takes restaurant reservation calls end-to-end — handles availability, party size, dietary notes, and confirmation, then writes to the booking system. The first time hearing it pick up was uncanny.
Lead qualification agent on a partner agent platform.
Qualifies inbound leads against an ideal customer profile — pulls firmographic and intent signals, scores, routes. Built on the Nasiko A2A platform as part of their agent ecosystem.
Travel chatbot grounded in YouTube creator transcripts.
Niche RAG that treats travel YouTube creators as the knowledge base. Extracts transcripts, builds a hybrid retriever, answers travel questions with cited timestamps. The internet’s actual travel guide, finally queryable.
Car wash management system, no-code first.
End-to-end ops for a small car wash — bookings, customer ledger, daily reports. Built to test how far Lovable + Supabase can take you before you reach for code. Further than expected.
Started in analytics, ended up in production GenAI. The titles shifted; the work — turning messy real-world data into systems people can trust — didn’t.
Started in sports analytics and operational analytics; built predictive models, ETL pipelines, and early ML systems across sports, real estate, and infrastructure verticals.
Marketing Mix Modeling, real-time campaign optimization, user segmentation. First serious production ML work — the kind that has SLAs.
Transitioned into GenAI leadership. Built production RAG, knowledge graphs, and MLOps infrastructure. Won the AWS Global GenAI Hackathon during this period.
Leading enterprise RAG delivery, LLM evaluation frameworks, and AI governance for Tier-1 clients. Architecting parallel cloud-native reference implementations across AWS Bedrock and Azure OpenAI.
I lead the Bangalore chapter of The AI Collective — the largest practitioner community in India for production AI engineers. Since March 2025, we’ve shipped 40+ events, hosted speakers from across the ecosystem, and built the kind of room where people argue about eval harnesses and then go get dinner.
I publish small, sharp agent skills — opinionated prompts and tools designed to make Claude better at one specific thing. A few are below; the rest are in drafts.
I read everything. Best for: AI / GenAI Solution Architect conversations at AI-native companies, production-RAG and governance consulting, community collabs, or just a good thread about retrieval trade-offs. Slow for: cold sales pitches.