I make AI systems with trust issues — one hangs up on scammers, one reads contracts assuming the worst, one refuses to give health advice without citing the WHO. CS undergrad at VIT Bhopal ('27, AI & ML).

The "OP" in the username is self-reported and has never been independently verified. The 007 is classified.

ScamCall Guardian — flags scam calls in English, Hindi, Hinglish & Tamil while you're still on the call, not in a post-mortem. Whisper → regex rules → fine-tuned MuRIL → LLaMA 3.3, fused by a deterministic risk scorer. 93.9% accuracy, and a Redis semantic cache that cut LLM latency by 94% because scammers don't wait for cold starts.

FinePrint AI — an agentic contract analyzer (LangGraph, two-lane state machine) grounded in 42 Indian statutes. It reads the terms and conditions. Voluntarily. We don't fully understand it either.

VisionRAG — cross-modal search where a sentence can find an image and an image can find a document. 92.6% Recall@5 in under 140ms — and a published benchmark where CLIP's text encoder scored 59.3%. It knows what it did.

Thiran (live) — a Django online judge with sandboxed multi-language execution, AST-level static analysis, kernel resource limits, and a documented threat model. 500+ users trusted it to reject their code fairly, at 1.9 seconds per rejection.

Jeevo — a multilingual WhatsApp health assistant doing RAG over 23+ WHO / ICMR / NIH / CDC guideline docs, because health advice without sources is just confident vibes.

• AI intern @ amasQIS.ai — computer vision & predictive maintenance on industrial data; taught models to spot machine failures before the machines fully committed to them (+17% model performance, 92% detection accuracy)
• SDE intern @ AlgoUniversity (YC-backed) — where Thiran was built, shipped, and stress-tested by 500+ users with strong opinions about verdicts
• top 1.4% of 40,000 at AlgoUniversity's accelerator camp — a fact I have now mentioned exactly once

• year 4 of 4 at VIT Bhopal, CGPA 8.86 and holding
• building things, benchmarking things, occasionally letting the things win
• grinding DSA (150+ problems down, self-respect intact)

Python since forever, PyTorch when it matters, and an unreasonable amount of time with LangGraph, Transformers, FAISS, Redis, FastAPI, and Docker. Also Java, C++, SQL, Postgres, and enough AWS to be dangerous. Prometheus + Grafana on everything, because if it isn't monitored, it isn't real.

is JP actually OP? — results may vary. see pinned repos and judge for yourself.

why do all your projects speak three or four languages? — because scams, contracts, and health emergencies don't wait for translation.

are you hireable? — extremely.

s.jayapradeep7@gmail.com · LinkedIn

I reply faster than my scam detector flags a "your parcel is stuck at customs" call.