Skip to main content

Talks

A running list of conference talks, meetups, and community appearances — Kafka, vector search, Git, platform engineering, org change, whatever I was deep in at the time.

2026

When Semantic Search Breaks in Your RAG Systems: RAM Walls, Silent Failures, and the Architecture Decisions

The team shipped semantic search and it worked — then 100M vectors tripled the RAM bill overnight and filtered queries started silently returning zero results. The RAM math to run before any architecture decision, quantization that hits 32x compression at 95%+ recall, why filtered search is the #1 silent production failure, and when DiskANN, hybrid BM25+vector, or a specialized vector DB actually earns its place — trade-offs in plain English, with live demos.

Watch · Slides (PDF) · Write-up

Kafka & Time: The Hardest Dependency

Distributed systems are hard not because they span machines, but because they span time. A look at real production outages caused by partition keys, timestamp modes, and windowing contracts nobody consciously chose.

Slides (PDF)

How Startups Transform Legacy Organisations from the Inside

How a small acqui-hired startup team, dropped into a 4,000+ person legacy organisation, actually changed the way it worked from the inside — years of setbacks, small wins, and the slow work of building trust rather than pushing best practices.

Watch · Slides (PDF)

OWASP Top 10 for LLM Applications — live breaking and fixing

A live, hands-on session breaking and fixing LLM applications against the OWASP Top 10 for LLM risks — prompt injection, misinformation, excessive agency, supply chain, and more, each with a real demo and a real fix.

Slides (PDF) · Repo

Vector Search & Storage in PostgreSQL with pgvector

Inside pgvector: how embeddings get stored (and TOASTed), why B-trees can't index vectors, and how IVFFlat and HNSW trade off build time, recall, and speed — plus quantization and the production gotchas that show up past the demo.

Slides (PDF) · Repo

From Text to Vector: How High Dimensional Data Gets Stored, Searched & Retrieved

What actually happens after you call an embedding API — how text becomes vectors, how they're stored, and how semantic search finds the closest match. Starts with a brute-force scan across 50,000 vectors, then shows how IVFFlat and HNSW deliver 50–150x speedups, all in plain SQL you can run on your laptop.

Slides (PDF) · Repo

Storing high-dimensional data at scale

What breaks when you push vector search to 100M+ rows on Postgres, how IVF/HNSW and quantization help, and why DiskANN plus hybrid BM25/vector search mattered in production.

Watch · Repo

2025

Git internals beyond add, commit, push

A practical deep dive into Git internals, for using Git with more confidence than the everyday porcelain commands give you.

Slides (PDF)

Inside pgvector: how PostgreSQL stores and manages high-dimensional data

How pgvector represents and indexes high-dimensional embeddings inside PostgreSQL, and what that means for query performance at scale.

Watch · Slides (PDF) · Repo

2024

Unlocking Efficiency and Collaboration: How Backstage Solves Key Developer Challenges

How Backstage, the open-source developer portal, tackles knowledge discovery, code ownership, system interfaces, dependency mapping, tech radar, and service creation.

Slides

Effective data consumption from Kafka

Configuring consumers for reliability and speed — fetch size, session timeouts, rebalancing, consumer lag, and the tradeoffs between throughput, latency, and resource use.

2023

Why, what, and how of practically working with Apache Kafka®

What problems Kafka actually solves, when to reach for it, and a long walk through common use cases — the knobs to turn and the tradeoffs that come with them.

Recording

2022

GitLab Bengaluru — first in-person meetup

Spoke at GitLab's first in-person Bengaluru meetup, alongside a full house on a rainy Saturday morning.