I take AI from demo to production: agents, automations, and the unglamorous debugging in between. Author of diffprompt, live on PyPI. Friends call me funny, ambitious, and delusional. They're right about all three.
Second year at IIT Bhilai, studying Data Science & AI. I got into this field the normal way: built things, broke things, then realised the actual job is keeping things alive in production. Now I build AI systems end to end: agents, automations, data pipelines, and the 2am debugging sessions nobody posts about.
By day I ship client-facing systems at Altagic. By night I maintain diffprompt, contribute to open source AI tooling, and queue Valorant with the boys. Everything I build gets stress-tested before it ships, because I'd rather break it than let a user do it.
I coordinate placements for 200+ students at CCPS, lead the AI/ML domain at OpenLake, and hold the strong opinion that DSA grind scores tell you almost nothing about whether someone can actually build.


One flagship library, a stack of deployed systems, and an internship under NDA. Every card reads like a case study: the problem, the build, what shipped.
Behavioral prompt regression testing. LLM-as-judge cascades score divergence between prompt versions, HDBSCAN clustering surfaces semantic drift. Catches the regressions that break production AI quietly.
Problem: multi-agent backends die quietly when one route fails. Build: LLM orchestration on Firebase with self-healing routing and per-user Firestore memory.
Problem: black-box trading models nobody can explain. Build: LightGBM + CatBoost ensemble with SHAP explainability, Dockerised with CI/CD, live signal scoring on Streamlit Cloud.
Problem: cold outreach at scale without sounding like a bot. Build: LangGraph pipeline that scrapes, filters by role criteria, drafts personalised emails, and waits for human approval in Gmail before anything sends.
What I can say: agentic automation systems with AI in the loop, built for reliability and scale, running in production. What I can't say: everything else. The NDA is doing its job.
Problem: product listicles take hours of manual copy and layout. Build: storefront-to-HTML generator covering copywriting, images and rendering end to end.
"You opened Instagram again. Enjoy your McDonald's application." Watches your active window; catch yourself doomscrolling and your study doc flies open while the terminal roasts you. No timers, no graphs, just consequences.
A model that predicts when Drake drops Icemannnnn. Yes, this is a real repo. Yes, it counts as data science. No, I will not be taking questions.
Compress any LLM conversation into one portable context prompt. Paste it into a fresh session and continue exactly where you left off. ChatGPT and Claude share links supported.
Turns exported connection data into growth velocity, company concentration and structural leverage metrics. The 1.5L impressions were not an accident.
Feature engineering on the FIFA dataset: parsing positions, traits and contracts into features that actually predict player market value. FIFA rage, productively channeled.
Maps a real classroom's friendships as a graph: hidden communities, centrality scores, and which students secretly hold the whole network together.
Where I'm contributing right now: AI tooling and infrastructure used by real teams. This list updates itself, because hardcoding your own PRs is embarrassing.
The taste section. Three lanes, alternating directions, hover to pause and read the commentary.









"DSA grind scores tell you almost nothing about whether someone can actually build and ship."
Exhibit A: everything on this page was built, not LeetCoded.
"Samosa with mayo and ketchup is genuinely good and the backlash is performative."
Try it before you report me.
Long-form writing on taking AI to production. Case studies, debugging stories, occasional takes.
The internal soundtrack. Three singles, zero streams, full conviction.