Shubhangi Gupta
Open Source & AI Ecosystem Builder | Product & DevRel | Community of 35K+ | Inclusive Tech Advocate 🏳️🌈
Delhi, India
Actions
Shubhangi is a Product and DevRel leader building at the intersection of open source, AI, and inclusive tech communities.
She leads a 35,000+ member developer network and drives initiatives through Claude Code, Google Developer Groups Noida and Women Techmakers Delhi. A GitHub Campus Expert and speaker at multiple tech events like DevFest and MLH Hackcon, she champions open collaboration, equitable representation, and community-led innovation in AI and infrastructure.
Area of Expertise
Topics
Open Source Accelerating AI: From Research Labs to Real-World Builders
AI didn’t accelerate because models got bigger.
It accelerated because they got open.
In 2019, building a production-grade image detection system meant years of research and heavy computing. In 2022, Stability AI released Stable Diffusion openly, and image generation ran on consumer GPUs within weeks. The shift wasn’t just architectural, from CNN pipelines to diffusion models, it was cultural: open weights, public checkpoints, shared benchmarks.
This talk covers:
• The architectural inflexion point, how diffusion models, open releases, and tooling like Hugging Face Transformers and Diffusers turned AI research into reusable infrastructure.
• The ecosystem effect, how LangChain, model hubs, and shared datasets compressed experimentation from years to days.
• The OpenCLAW moment: how open coding agents accelerated development while exposing real security and governance risks.
Core takeaway:
Openness didn’t just speed AI up —> it redistributed power over who builds and shapes its infrastructure.
Open Source, Open Board: I Built a Robot to Play Board Games With Me
A few years into adult life, I ran out of people to play board games with on a Tuesday night. So I built one. Not an app, a physical robot that lives on the table, picks up the dice, moves the pieces, and reacts when it wins or loses.
This talk walks through the open-source stack I built on the LeRobot SO-101 arm: a perception module that turns an overhead camera into a structured game state, a per-game engine (rule-based for simple games, Stockfish for chess), an imitation-learned motor module trained from teleop, and an affect layer that gives the robot personality. It plays card games, Snakes and Ladders, and Ludo; chess is in progress.
I'll cover three things:
1. The engineering (what runs on-device, latency budget for vision plus arm control, which open models survived contact with a real board)
2. The affect layer (how to make the robot's emotions feel natural)
3. The failures (dropped pieces, misunderstood moves, detecting false moves)
You'll leave with a reference architecture, a soldering iron's worth of inspiration, and a sharper sense of why embodied affect is the under-explored layer in physical AI.
Stage Fright to Speak Right
This session will help you get rid of the fear of public speaking and help you build your next talk!
Your Search Is Broken — And You Chose That
Most engineering teams pick their search method based on what they already know. They reach for BM25 because it ships with OpenSearch. They bolt on an LLM because the PM asked for "AI search." Nobody benchmarked it.
Three years later, the RAG pipeline is confidently returning wrong answers — while a graph-based approach returns exact, explainable results in 340ms.
This talk is a decision framework, not a feature tour.
We run one real query, "Find me an engineer, ex-Googler, now running a Series B fintech" — through five methods: BM25, semantic, hybrid, RAG, and GraphRAG. Each gets closer. Each exposes a new failure mode. The audience watches hybrid search outperform BM25 by 14% on relevance (AWS/OpenSearch benchmark, 2024), sees RAG hallucinate a funding stage despite correct retrieval (5–15% real-world rate, K2view 2024), and watches GraphRAG traverse a knowledge graph to return a verified, fully-explained answer.
The talk closes with a decision matrix => query complexity × data structure × latency — that maps directly to OpenSearch implementation paths available today in 2.10+.
Every claim is sourced. Slides, demo app, and decision matrix.
Beyond transcription: deploying Whisper for neurodivergent-aware voice inference
Most voice interfaces flatten speech into text and discard everything else; they were never designed to "listen" to human intent.
With 15-20% of the global population neurodivergent, this is an accessibility failure at infrastructure scale. This session presents a voice-first research prototype (not a diagnostic tool) that uses PyTorch-based Whisper to extract speech pattern markers: pacing, hesitation, repetition, and adapt responses based on those signals.
The speaker walks through containerising the Whisper inference pipeline, deploying it on Kubernetes with GPU acceleration, and designing adaptive systems that preserve privacy by keeping voice data local. Attendees will see a live demo comparing standard and ADHD speech patterns, learn how to build voice interfaces users can tailor to their own communication style, and leave with a forkable repo containing Dockerfiles, K8S manifests, and the full adaptive response pipeline.
Shubhangi Gupta
Open Source & AI Ecosystem Builder | Product & DevRel | Community of 35K+ | Inclusive Tech Advocate 🏳️🌈
Delhi, India
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top