ML researcher who also ships. B.Tech in AI & Data Science (Class of 2028, CGPA 9.2) at Amrita Vishwa Vidyapeetham. I do applied ML research — explainable AI, multimodal fusion, graph learning — and I build the systems that serve models in production: APIs, drift detection, monitoring, all the way down to a TCP/IP stack in Rust.
🔭 Currently: Building tcp stack from scratch 🌍 Available: remote work on full US Eastern business hours (my 6 PM–2 AM IST = 8:30 AM–4:30 PM ET). 📫 Reach me: daasaradhimannava@gmail.com · LinkedIn
My research spans applied ML across domains — the common thread is models you can trust and explain.
| Work | Area | Status |
|---|---|---|
| Semantic intrusion detection — heterogeneous GNN over network flows + knowledge-graph reasoning + LLM-generated explanations | Graph ML · XAI · Security | |
| Multimodal semiconductor yield prediction — late fusion of 591 sensor features + 4096 wafer-map vision features | Multimodal · Manufacturing | |
| Explainable AI methods | XAI | |
📌 My contribution on every listed paper: lead author — problem formulation, method design, and all experiments.
MiniFlow → MiniFlow-Serving — a deep-learning framework, then a serving layer for it
The full lifecycle in one project: I wrote a reverse-mode autograd engine from scratch (no PyTorch/TF — tensors, computational graph, backprop), trained models on it, then built the layer that serves them.
- Serving: FastAPI
/predict, sticky A/B testing with a two-proportion z-test, PSI-based drift detection running in the background - Observability: Prometheus metrics → Grafana dashboards, Docker Compose for the whole stack
- Engineering: 24 tests, 89% coverage; p99 latency 97 ms
- Why it matters: most students train models; almost none have built both the framework and the monitoring that catches it degrading in production
LLM from Scratch — a transformer LM, no shortcuts
Tokenizer → multi-head attention → positional encodings → training loop → sampling, all hand-built to own the internals. I can whiteboard attention, KV-caching, and why RoPE works — because I implemented them.
S-XG-NID — intrusion detection that explains itself
Heterogeneous graph neural network over network flows (CICIDS dataset) with a knowledge-graph reasoning stage; every detection ships with an LLM-generated human-readable explanation of why it was flagged. Bridges my security, graph-ML, and XAI work.
ARP, IPv4, ICMP, and TCP — congestion control, segment reassembly, RTT estimation — running in Linux userspace over a TUN device. Built to understand what every model.predict() HTTP call actually rides on.
- FabMind — the research codebase behind the yield-prediction paper, with explainability tooling
- GlassBox-Attack — adversarial attack/defense experiments; robustness evaluation of NN classifiers
- StudyMetrics + Battleship — two Android apps in Kotlin; StudyMetrics is built as 18 CI-gated Gradle modules
| ML / DL | PyTorch · NumPy · scikit-learn · transformers & attention internals · GNNs · adversarial ML · explainability (GradCAM/SHAP) |
| Serving / MLOps | FastAPI · Docker & Compose · Prometheus · Grafana · PSI drift detection · A/B test design · CI (GitHub Actions) |
| Languages | Python · Rust · Kotlin · C++ · SQL · Bash |
| Systems | Linux · TCP/IP (implemented one) · Android |
- Research internships (Summer 2027) — XAI, multimodal learning, graph ML, LLM systems
- Remote part-time / contract ML engineering on US hours — model serving, evaluation harnesses, inference infra
- Collaboration on open-source LLM inference (vLLM ecosystem) and XAI benchmarking
Fastest way to evaluate me: open MiniFlow-Serving, run docker compose up, and hit /predict.