Research & Development

Aug 9, 2026

Research & DevelopmentTechdiscussion

Is anyone actually shipping AI features, or are we all still prototyping?

Every team I talk to has an LLM prototype. Very few have one in front of paying users. The gap seems to come down to unglamorous things: • What happens when the model returns nonsense at 2am and nobody notices? • How do you test something whose output changes between runs? • Who pays when a single user pastes a 40,000-token document? I am not asking whether it is useful — clearly it can be. I am asking what it took to get from demo to production, and what you cut to get there. If you have shipped something real, I would genuinely like to hear what broke first.

Indian Tech CommunityManager

Jul 27, 2026

Research & DevelopmentAIGPUInfrastructure Cost

How are Indian startups actually handling GPU costs for AI features — renting, cloud credits, or local clusters?

GPU cost is becoming the biggest line item for anything touching inference at scale. Curious what smaller Indian teams are actually doing — spot instances, committed cloud credits, or has anyone gone as far as a small local cluster?

Indian Tech CommunityManager

Jul 12, 2026

Research & DevelopmentAIOn-device MLMobile

Experimenting with on-device small language models for a mobile app — early results

Tried replacing a cloud LLM call with an on-device small model (under 3B params) for a simple classification feature in our app. Latency dropped from ~800ms to under 100ms and we cut API costs entirely for that feature, at the cost of noticeably lower accuracy on edge cases.

Indian Tech CommunityManager

Jun 6, 2026

Research & DevelopmentWebAssemblyEmerging TechBackend

Is WebAssembly actually production-ready for backend workloads in 2026, or still mostly edge/plugin use cases?

Seeing more articles about WASM outside the browser but not many real production case studies for backend services. Anyone running WASM in production beyond edge functions?

Indian Tech CommunityManager

Apr 5, 2026

Research & DevelopmentAIRAGMachine Learning

Fine-tuning vs RAG for a domain-specific support bot — what we learned after trying both

Started with fine-tuning on our support tickets, results were inconsistent and expensive to iterate on. Switched to RAG over our docs and ticket history and both accuracy and iteration speed improved. Fine-tuning still made sense for tone/style, RAG for factual grounding — using both now.

Indian Tech CommunityManager

Mar 11, 2026

Research & DevelopmentInnovationTeam CultureR&D

We gave engineers one day a month for open research — most of it didn't 'ship' but the side effects were worth it

Skeptical going in, since a day a month across 15 engineers is real cost. Most projects didn't become features, but two turned into internal tools that saved real time, and engagement/retention feedback improved noticeably. Treating it as R&D insurance now, not a shipping pipeline.

Research & Development

Create Post

Is anyone actually shipping AI features, or are we all still prototyping?

How are Indian startups actually handling GPU costs for AI features — renting, cloud credits, or local clusters?

Experimenting with on-device small language models for a mobile app — early results

Is WebAssembly actually production-ready for backend workloads in 2026, or still mostly edge/plugin use cases?

Fine-tuning vs RAG for a domain-specific support bot — what we learned after trying both

We gave engineers one day a month for open research — most of it didn't 'ship' but the side effects were worth it