Krish

Foundations now. LLM products soon. | ML • Python • Data stack | Building in public for SaaS founders

[Krish @Born_TechK]

Most telecom companies lose customers before they know it's happening.

I built an ML model that predicts churn before it costs revenue.
Here's the business case + what I found

🧵 1/4

[Krish @Born_TechK]

Built an extractive QA system using deepset/roberta-base-squad2

It reads your paragraph → finds the exact answer span → gives a confidence score.

Added it as Tab 2 to yesterday's summarizer app.

One app. Two NLP tools. Live on Streamlit.

🔗Demo Link below

[Krish @Born_TechK]

🔗Live Demo - https://t.co/7p4yhqYxGq

GITHUB repo - https://t.co/UKNEIdIbTo

[Krish @Born_TechK]

Built an extractive QA system using deepset/roberta-base-squad2

It reads your paragraph → finds the exact answer span → gives a confidence score.

Added it as Tab 2 to yesterday's summarizer app.

One app. Two NLP tools. Live on Streamlit.

🔗Demo Link below

[Krish @Born_TechK]

Fine-tuning👇

Cost: high
Latency: lowest, no retrieval
Data: needs labeled training data (usually 100s–1000s of examples)
Updates: expensive, any change means retraining
Use when: the task is consistent/stable (not changing data), you need very specific output format

[Krish @Born_TechK]

WHEN TO USE:

Prompt Engineering VS RAG (Retrieval-Augmented Generation) VS Fine-tuning 👇

[Krish @Born_TechK]

RAG (Retrieval-Augmented Generation)👇
Cost: medium, embeddings + vector DB + LLM API
Latency: medium
Data: ideal for private/frequently-updated knowledge
Updates: high frequency, just re-embed new docs
Limit: requires the answer to actually exist somewhere in your documents

[Krish @Born_TechK]

🔗 DEMO LINK - https://t.co/7p4yhqYxGq

[Krish @Born_TechK]

Built an extractive QA system using deepset/roberta-base-squad2

It reads your paragraph → finds the exact answer span → gives a confidence score.

Added it as Tab 2 to yesterday's summarizer app.
One app. Two NLP tools. Live on Streamlit.

🔗Demo Link below👇

[Krish @Born_TechK]

Building in public → https://t.co/Yhj5N2p5Sj

Live Demo → https://t.co/dpuUDBQ4n9

[Krish @Born_TechK]

Text summarization app live - BART model, Streamlit, deployed.

Input: any long document
Output: concise summary + compression %

The same tech that can auto-summarize your customer tickets.

[LIVE DEMO BELOW👇]

[Krish @Born_TechK]

@dk119819 @X Hey! Let's connect

[Krish @Born_TechK]

@sreenandhanpp Hey! Let's connect

[Krish @Born_TechK]

GITHUB REPO -> https://t.co/cbVuFCBETI

[Krish @Born_TechK]

Your support team reads 500 tickets/day.
T5 reads each one in 0.3s and outputs a 2-sentence summary.
Built a demo today. Same model also translates.
One architecture. Any text task.

Link Below 👇

[Krish @Born_TechK]

LIVE DEMO -> https://t.co/kD8gkVCaHy

GITHUB REPO -> https://t.co/giwGug06ZC

[Krish @Born_TechK]

My fine-tuned BERT called a positive SaaS review NEGATIVE with 57.8% confidence.
That 57.8% is the model telling you it's lost, not certain.
Here's the domain mismatch problem 🧵

[Krish @Born_TechK]

Built and deployed a working demo.
Fine-tuned BERT on IMDb → wrapped in Streamlit → live URL.
The next step is fine-tuning on actual SaaS feedback data.
Demo + GitHub → [link in reply ↓]

[Krish @Born_TechK]

SaaS founders spend hours manually tagging customer feedback.
"Feature request", "bug report", "churn risk", "positive signal".
A domain-fine-tuned model does this in real-time.
No labeling. No spreadsheets. No lag.