Anushka Shandilya

SWE @PhonePe | AI/ML

[Anushka Shandilya @Anushkaaa_sh]

Migrated my Sanskrit tutor's TTS from Sarvam to ElevenLabs.

Quick note: I compared free tiers. Sarvam gave me Hinglish voices for free, but ElevenLabs gates its Indian voices behind a paid plan, so I tested it on English voices. Real Hinglish quality on ElevenLabs will only show up once you upgrade.

I stuck with ElevenLabs for four reasons:

1) Three voice controls instead of one. Sarvam gives you pace. ElevenLabs gives you speed, stability, and style. That means a question can have a rising, expressive contour while a correction stays calm and steady. Real modulation per sentence, not just per turn.

2) previous_text for boundary continuity. Each new synthesis call takes the previous chunk's text as context, so the intonation carries across chunks instead of restarting cold. Sarvam had no equivalent and you could hear the reset on longer replies.

3) Simpler API. Sarvam is a stateful websocket with a configure step and completion events. ElevenLabs is client.convert() returning an async iterator of bytes. Half the integration code disappeared.

4) Pitch variation across a sentence. Sarvam's contour is flatter, which reads as news-reader. ElevenLabs varies more, which reads as conversational. Matters for a tutor.

Where Sarvam is still the harder pick to beat: pronunciation of Hindi-first words, cost at scale, and latency for Indian users.

For a Hinglish tutor, expressiveness won. For pure Hindi at scale, my answer would be different.

[Anushka Shandilya @Anushkaaa_sh]

I built an app for my mom who is a Sanskrit teacher. Used sarvam ai bulbul model and its genuinely amazing.

Posted this on instagram.

[Anushka Shandilya @Anushkaaa_sh]

I come from a non cs background and i recently landed software engineer role at a good company.

A lot of my relatives humiliated me many times coz I was good at studies, yet I ended up pursuing a non-CS branch of btech(thats what i was getting at a govt college).

A lot of people at my uni used to make fun of me coz I was always trying, hustling and working towards something. My routine was very different from everyone else’s. I didn’t really make friends and honestly, I don’t have many even today.

And today, I’m in a place that I once only hoped I could reach. I’m not saying this to brag but I know that whatever I have achieved, I’ve earned it.

The whole point of telling you this is simple: keep working towards whatever you want in life. Put your blinders on and keep going. People will talk, judge you, laugh at you and sometimes even make you doubt yourself.

Btw,
This is a photo of my parents. They were so happy and proud at that moment and honestly, that feeling means more to me than anything else.

[Anushka Shandilya @Anushkaaa_sh]

I come from a non cs background and i recently landed software engineer role at a good company.

A lot of my relatives humiliated me many times coz I was good at studies, yet I ended up pursuing a non-CS branch of btech(thats what i was getting at a govt college).

A lot of people at my uni used to make fun of me coz I was always trying, hustling and working towards something. My routine was very different from everyone else’s. I didn’t really make friends and honestly, I don’t have many even today.

And today, I’m in a place that I once only hoped I could reach. I’m not saying this to brag but I know that whatever I have achieved, I’ve earned it.

The whole point of telling you this is simple: keep working towards whatever you want in life. Put your blinders on and keep going. People will talk, judge you, laugh at you and sometimes even make you doubt yourself.

Btw,
This is a photo of my parents. They were so happy and proud at that moment and honestly, that feeling means more to me than anything else.

[Anushka Shandilya @Anushkaaa_sh]

Bought cute office desk stuff and now I'm shy to use it because everyone else's desk is so empty 😂 help

Just look at this pigi pen🫠

[Anushka Shandilya @Anushkaaa_sh]

Bought cute office desk stuff and now I'm shy to use it because everyone else's desk is so empty 😂 help

Just look at this pigi pen🫠

[Anushka Shandilya @Anushkaaa_sh]

@VishalYada89628 😂😂😂

[Anushka Shandilya @Anushkaaa_sh]

Career update: joined PhonePe as a SWE.

[Anushka Shandilya @Anushkaaa_sh]

@xcendxuv 😂😂😂

[Anushka Shandilya @Anushkaaa_sh]

@cph2691 @PhonePe Then you should work on yourself more...i am a fresher...i can make mistakes🙂

[Anushka Shandilya @Anushkaaa_sh]

@zero_bug_dev 😂😂😂

[Anushka Shandilya @Anushkaaa_sh]

@mansistwt Made my day🩷

[Anushka Shandilya @Anushkaaa_sh]

@SurajHardekar Yeaa! Their choice

[Anushka Shandilya @Anushkaaa_sh]

@shaham7362828 Haa

[Anushka Shandilya @Anushkaaa_sh]

@itsvikky07 Kuch bhi

[Anushka Shandilya @Anushkaaa_sh]

Career update: joined PhonePe as a SWE.

[Anushka Shandilya @Anushkaaa_sh]

Developers don't struggle to write code.
They struggle to understand code they didn't write.

So I built reflexRAG.

A RAG tool for GitHub that remembers more than the code:

→ commits
→ PRs
→ issues
→ discussions
→ decisions

It also checks its own context for stale or closed issues, contradictions, and version mismatches.

And with Contributor Mode, it helps you go from:

“Where do I even start?” → “I can contribute.”

Built from scratch.
Introducing reflexRAG.

[Anushka Shandilya @Anushkaaa_sh]

Developers don't struggle to write code.
They struggle to understand code they didn't write.

So I built reflexRAG.

A RAG tool for GitHub that remembers more than the code:

→ commits
→ PRs
→ issues
→ discussions
→ decisions

It also checks its own context for stale or closed issues, contradictions, and version mismatches.

And with Contributor Mode, it helps you go from:

“Where do I even start?” → “I can contribute.”

Built from scratch.
Introducing reflexRAG.

[Anushka Shandilya @Anushkaaa_sh]

Writing a backend, almost done and this is episode 5 of me building in public.

What am i building?
RAG platform for github by retrieving context not just from codebase but also from prs, issues and readme.

Why am i building this?
One day, I spent half of my day working on an issue that was already closed but github copilot did not warned.

What have i done today?
Finished writing every backend services - models, ingestion pipeline, LangGraph graph, critic reranker, cache worker and ofc main app. Now i'm just left with writing files that connects frontend.

My motive: most of the agents optimize for maximum task completion rather than maximising useful work.

Next episode will be much more detailed and i"ll discuss all the trajectories. I"ll put some links of previous episodes and i am open to ai engineer roles as well, dms are open.

[Anushka Shandilya @Anushkaaa_sh]

I am an ai engineer but i still want my younger brother to become that electronics nerd who build cute little things like "tamagotchi"😂.

He took robotics & ai and i think, both of us are going to learn the intersection of two together :)

btw i hated electronics lol

[Anushka Shandilya @Anushkaaa_sh]

I am an ai engineer but i still want my younger brother to become that electronics nerd who build cute little things like "tamagotchi"😂.

He took robotics & ai and i think, both of us are going to learn the intersection of two together :)

btw i hated electronics lol

[Anushka Shandilya @Anushkaaa_sh]

A friend reached out to me today asking if I knew any good websites for PPT templates.

He had to make a presentation on the sales and marketing of Mumbai Dabbawalas, but didn't have much time, so he used Claude to build the whole thing.

He showed me the PPT.
Honestly, I didn't like it much.

So I opened ChatGPT(obv i can't waste my claude tokens😭) and made the first slide with him. The difference was pretty noticeable.

It made me think: how many people have access to AI but still struggle to get the output they want?

And that feels like a product problem.

Maybe the future isn't about teaching everyone prompt engineering.

Maybe AI should get better at understanding what people mean, not just what they type.

If you’re building with AI, this is something worth thinking about: the next big AI product might not need better models. It might need a better interface between humans and those models.

[Anushka Shandilya @Anushkaaa_sh]

A friend reached out to me today asking if I knew any good websites for PPT templates.

He had to make a presentation on the sales and marketing of Mumbai Dabbawalas, but didn't have much time, so he used Claude to build the whole thing.

He showed me the PPT.
Honestly, I didn't like it much.

So I opened ChatGPT(obv i can't waste my claude tokens😭) and made the first slide with him. The difference was pretty noticeable.

It made me think: how many people have access to AI but still struggle to get the output they want?

And that feels like a product problem.

Maybe the future isn't about teaching everyone prompt engineering.

Maybe AI should get better at understanding what people mean, not just what they type.

If you’re building with AI, this is something worth thinking about: the next big AI product might not need better models. It might need a better interface between humans and those models.