Sales copilot
DealCloser
Listens to a sales call while it's happening, transcribes it with Azure Speech, and suggests what to say next: how to handle an objection, what the next step should be.
.NET MAUI · C# · Azure Speech
Hi, I'm Saravana 👋
I build AI tools and apps, and I like getting them out the door.
I've been writing software for 14+ years, and for the last 8 or so it's mostly been AI: search over messy documents, LLM workflows, and the apps around them. I also consult for Digitalgrub and build a lot of their projects with the team.
Usually reaching for: Python, LlamaIndex, FastAPI, Postgres with pgvector, Next.js, Flutter, and whatever cloud the project lives on.
A bit about me
I like problems where there's a pile of text and someone needs a straight answer out of it: clinical notes, contracts, support docs, product catalogues. A lot of my work has been in healthcare, but there's plenty of sales tools, e-commerce, education and side projects in there too.
Day to day that's mostly Python (LlamaIndex, FastAPI, Postgres with pgvector) on AWS or Azure, with Next.js or Flutter on the front when a project needs one.
I'm AWS certified (Solutions Architect, Developer and Cloud Practitioner) and Snowflake SnowPro Core certified, if that's the kind of thing you look for.
Stuff I'm good at
- Getting reliable answers out of documents: RAG on pgvector or Redis, with hybrid search and filters.
- Wiring LLMs into real workflows, with evals and caching so you know it works and it stays cheap.
- Turning a Streamlit prototype into something people use every day.
- FastAPI services, Docker and CI, deployed on AWS, Azure, Cloudflare or a plain Hetzner box.
What I can help with
If you're stuck on one of these, I can probably help.
Answers from your documents
You've got PDFs, notes or a database full of text and you want answers you can trust. I build search and RAG on LlamaIndex with Postgres/pgvector or Redis.
LLMs in your workflow
Prompts, tools and guardrails, plus evals and caching so it stays cheap and doesn't quietly get worse.
APIs and apps around it
FastAPI services, Streamlit prototypes that grow into real apps, Docker and CI so shipping is boring.
Hosting that doesn't fall over
AWS, Azure, Cloudflare or a plain Hetzner box. I'll also make your WordPress site fast again.
A few things I've built
Some are open source, some are products. The full list is on the timeline.
Desktop app
ScreenPager
Do something in your browser once and it writes the tutorial for you: steps, screenshots, a PDF, even a Playwright script that replays the whole thing.
Electron · Playwright · PostgreSQL
Healthcare
Clinical summarization
Turns unstructured clinical notes into a clean, structured summary, with a chat for follow-up questions. A person reviews everything before it's used.
Python · Streamlit · BiomedBERT · Vector search
Healthcare
Appeal letters
Writing medical appeal letters is the same work over and over. This reads the clinical documents and images, drafts a consistent letter, and someone checks it before it goes out.
AWS Comprehend Medical · Bedrock · OpenAI · S3
Experiment
A 3D avatar that talks
FastAPI sends back speech plus mouth shapes from Azure's text-to-speech, and a Three.js face lip-syncs along. Works with your own GLB avatar too.
Python · FastAPI · React · Three.js · Azure TTS
Browser extension
AI Smart Web Reader
Cleans up any web page so you can actually read it, then lets you ask quick questions about what's on it.
TypeScript · Plasmo · Browser extension
There's more on GitHub: LlamaDBIndexer, Azure AI Search scripts, PDF semantic search, and a pile of agent experiments.
Lately
Something new goes out most weeks. Here's the latest.
A closer look: clinical summarization
Reading unstructured clinical notes by hand is slow, and everyone summarizes them a little differently.
So I built a pipeline that pulls the key facts out, applies some rules, and writes a standard summary. A person always reviews it before it's used. Under the hood it's Python, BiomedBERT, Streamlit and Postgres.
Early signs are good: reviews go faster, need fewer edits, and come out more consistent. I'm still measuring it properly, tracking review time per case, how many fields get filled, how many edits each summary needs, and how often clinicians agree with it.
I also make videos
Walkthroughs of things I'm building. A couple of playlists: