| 🔬 2 Papers Under Review | ⚡ 75% Tail-Latency Cut (MOSAIC) | 🌍 120 Events/hr (DisasterGuard) | 🛠️ 5 End-to-End ML Systems Shipped |
|---|
| 🎯 Focus | Systems-aware ML infrastructure, edge AI, and clinical/high-stakes ML |
| 🎓 Education | B.Tech CSE (AI/ML), Lovely Professional University |
| 💼 Status | 🟢 Open to ML/AI Engineering & Systems internships, full-time roles |
| 🔬 Research | 2 papers in peer-reviewed submission (ICCNCom, Springer Nature) |
- 🔧 Contributing to MLflow (open source) — fixing a tags-propagation bug in
Model.log()→ModelVersion, with a regression test in progress - ⚙️ Building an automated Codeforces → GitHub sync pipeline (GitHub Actions + scraping-based submission capture)
- 📚 Sharpening DSA fundamentals (Graphs, DP, Greedy, Trees, Binary Search) for technical interviews
Most ML engineers go deep on systems or deep on models. I do both — and I ship.
In the last year I went from building applied ML pipelines to writing a hardware-profiled Linux userspace scheduler, a GNN-based clinical subtyping framework on MIMIC-IV, and a production real-time disaster alerting system. Every project here has real metrics, real hardware, or real users — not just notebooks.
Actively transitioning from applied ML → ML systems research, with a focus on efficient inference at the edge and high-stakes AI domains.
| Paper | Novel Contribution | Venue | Status |
|---|---|---|---|
| MOSAIC | First hardware-counter-profiled userspace scheduler evaluated on disaster-scenario edge workloads; online zero-shot workload classifier via nearest-centroid with EW updates | ICCNCom | 🔄 Accepted |
| T2D Subtyping | Progression-aware GNN subtyping of Type 2 Diabetes using DTW-attention alignment on longitudinal EHR; validated via Kaplan–Meier survival analysis | Springer Nature | 🔄 Under Review |
First hardware-profiled userspace scheduler designed for disaster-response edge nodes — benchmarked on real hardware, not simulation.
Python C Linux cgroups v2 Intel RAPL perf_event_open SQLite Docker Kubernetes
- 75% reduction in P99 tail latency (12,738ms → 3,178ms) via a hardware-counter-profiled 6×6 interference matrix used as an admission control primitive, measured with
perf_event_open()on real hardware - Zero task starvation (vs. 6.3% baseline) and 86% energy efficiency gain (631 → 1,171 tasks/Wh) via Intel RAPL feedback wired into cgroups v2 DVFS throttling
- Authored an online ML workload classifier (nearest-centroid + EW centroid updates) achieving perfect classification accuracy across 6 disaster-domain workload classes — zero training data required
- Reproducible benchmark harness: 5 schedulers × 7 metrics × 5 arrival patterns, 63 unit tests, 100% SSE-driven live telemetry dashboard, Docker/Kubernetes DaemonSet deployment
- 6-page workshop paper targeting USENIX HotEdge · Sep 2025 – Feb 2026
Moves T2D subtyping beyond static clustering — models how patients evolve over time, then validates that the subtypes actually predict different survival outcomes.
Python PyTorch Scikit-learn Pandas NumPy UMAP Lifelines MIMIC-IV
- Progression-aware subtyping framework on longitudinal MIMIC-IV EHR data, modeling patient trajectories across multiple temporal scales
- DTW-attention alignment + GNN-based phenotype fusion outperforms static K-Means with a silhouette score of 0.41
- Clinically validated subtypes via Kaplan–Meier survival analysis, revealing differential complication risk and treatment response patterns
- Aug 2025 – Jan 2026
Monitors 10 Indian cities for disaster risk in real time — from weather ingestion to shelter assignment in under 500ms.
Node.js React 18 Supabase PostgreSQL HuggingFace Leaflet.js SSE Vercel Render Tailwind CSS
- 120 weather events/hour ingested via OpenWeatherMap API with sub-500ms alert generation latency
- Event-driven pipeline: HuggingFace AI risk classification (LOW/MEDIUM/HIGH) + deterministic rule-based fallback + Haversine geodesic algorithm for automatic nearest-shelter assignment across 10 Indian states
- Production system on free-tier infrastructure (Vercel + Render + Supabase): SSE real-time map updates, Supabase Auth, Gmail SMTP alerts, GitHub Actions CI/CD, 100% uptime via UptimeRobot
- Feb 2026 – Mar 2026
📂 More Projects — click to expand
Goes beyond hand gestures — interprets the full non-manual signal of ISL including facial expressions, gaze, head pose, and shoulder dynamics.
Python MediaPipe OpenCV Graph Neural Networks Temporal Modeling
- Semantic graph fusion (SFG) framework fusing non-manual features — facial expressions, gaze direction, head pose, shoulder dynamics — for holistic ISL interpretation
- Lightweight and explainable: built on MediaPipe + OpenCV for real-time inference with no GPU dependency
- Temporal stabilization layer reduces jitter in continuous signing sequences
Detects outbreaks 7–14 days before traditional surveillance systems by fusing probabilistic signals rather than waiting for confirmed case counts.
n8n Bayesian Signal Fusion Probabilistic Modeling
- Bayesian probabilistic signal fusion with regional seasonal baselines, explainable alert reasoning, and cross-region spread modeling
Languages
ML / AI & Data
Systems & Infra
Backend & Databases
Frontend & Tools
|
|
|
Actively solving DSA & CP problems, with automated sync across platforms.
Focus areas: Graphs · Dynamic Programming · Greedy · Trees · Binary Search
