class ManasaDeshagouni:
def __init__(self):
self.education = "MS Computer Science @ San Jose State University 🎓"
self.previously = "Software Engineer @ UnitedHealth Group (2+ years)"
self.research = ["Zero-shot malware detection 🔬", "Cross-domain activity recognition 📱"]
self.battle_tested = "ML inference at p95 < 100ms serving 50K+ daily transactions"
self.trophy_wall = ["🏆 ICPC Asia Regional — Top 10 / 136 teams",
"📜 Best Paper Award, ICCDS 2024",
"✍️ 2x First-Author Publications"]
def current_obsession(self):
return "Teaching machines to recognize threats they've never seen before"- 🔭 Building ML systems that survive production — not just notebooks
- 🛡️ Researching zero-shot malware family identification — 19K+ binaries, 98% open-set AUROC
- ⚙️ I like my ML with an SLA: ONNX Runtime, p95 under 100ms, shadow-mode A/B rollouts
- 🎮 I build things that are fun too — real-time quiz battles, voice lie-detectors, AI agents
- 💬 Ask me about PyTorch, self-supervised learning, FastAPI, React, or competitive programming
|
A malware detective that identifies families it has never seen before. Neural fingerprints from 19K+ binary visualizations, bridging Windows PE ↔ Android. 72% Retrieval@3 · 98% open-set AUROC — served through an interactive analyst dashboard. |
A competitive coding arena where quizzes become live multiplayer battles 🎮 — real-time WebSocket rooms, live leaderboards, timed challenges, and an in-browser Monaco code editor with execution. p95 ≤ 180ms under load. |
|
A 609K-parameter model that recognizes human activity from a phone in your pocket — across 7 datasets it was never trained on. Shift-aware, label-free calibration hits 95.9% macro-F1 on zero-shot targets. |
A voice-based behavioral classifier that reads between the lines — pauses, disfluencies, pitch, energy, sentiment — from short, noisy audio clips. 89.4% accuracy · 93.5% precision with robust validation. |
|
An AI agent that swallows your digital chaos — PDFs, notes, receipts, code snippets — into a 100K+ document FAISS vector index and gives back semantic search, summaries, and context-aware answers. |
Built at HackSJ 2025 in 24 hours (team of 3) — an AI reviewer that learns from your team's past PR comments via RAG and suggests context-aware improvements, with a human-in-the-loop approval step. |
| 🥇 | ICPC Asia Regional Contest — Ranked Top 10 of 136 teams, solving 7 problems 📊 Standings |
| 📜 | Best Paper Award — Brain Tumor Detection using Machine Learning, First Author, ICCDS 2024 |
| ✍️ | Published Research — Deep Learning Techniques for Detection of Deepfakes, First Author, IJSRSET, ICSCR 2023 |
