Software Development Engineer · Full-Stack · Distributed Systems · Applied ML
PDEU (B.Tech CS, CGPA 9.29) + IIT Madras (B.S. Data Science) · Graduating 2027 · Ahmedabad, India
I build software that ships to production. Full-stack web apps, ML pipelines, distributed systems on Kubernetes, whichever part the problem needs. Not laptop demos, not tutorials. Real services with race-condition guards, GitOps pipelines, and the kind of incident I've debugged at 1 AM because Redis went down.
Currently shipping Cortex (distributed task platform on Kubernetes with Argo CD), Pitlane Live (F1 analytics on Render + Supabase), and a RAG system over 589 pages of FIA regulations. GATE 2026 qualified in CSE. Looking for SDE / ML systems internships where the work is hard and the production constraints are real.
ML & GenAI: PyTorch · scikit-learn · LightGBM · SHAP · LangChain · FAISS · HuggingFace · Sentence Transformers · Groq
React 19 · Node.js/Express 5 · Python worker · MongoDB · Redis · Docker · Kubernetes · Argo CD · GitHub Actions
A production-grade distributed task processing system. Tasks submitted from the React frontend are queued in Redis (lPush), picked up by a Python worker via blocking brpop, and tracked from pending → running → success/failed in real time. Every component containerised, every deployment declared as Kubernetes manifests, every push to main triggers a full build-push-deploy pipeline that Argo CD syncs to the cluster automatically.
The hard part: GitOps-driven CI/CD. GitHub Actions builds Docker images, pushes to Docker Hub tagged with commit SHA, then updates the companion infra repo's manifests via sed and commits. Argo CD watches that repo with auto-sync, prune, and self-heal enabled. Cluster drift gets corrected automatically.
Flask · Vue 3 · PostgreSQL (Supabase) · FastF1 · Gunicorn/gthread · Render
Full-stack F1 platform with race replays, live standings, and user predictions. Originally deployed on Azure VM with Gunicorn + gevent + Nginx; migrated to Render + Supabase when Azure student credits ran out. Switched gevent to gthread workers to fix an SSL monkey-patching recursion error on Render's runtime. Pre-built race JSONs committed to the repo to keep the 512MB server within memory bounds.
LangChain · FAISS · BGE-small · Cross-Encoder Reranker · Groq LLaMA 3.1 8B · Streamlit
Production RAG over 589 pages of FIA 2026 F1 Regulations. Two-stage retrieval: BGE bi-encoder for recall (top-20 via FAISS), cross-encoder reranking for precision (top-5). 80% Section Match@1 at ~1,000ms retrieval latency, CPU-only, zero cost. Three-layer hallucination mitigation: prompt grounding, temperature=0, and a numeric verification heuristic.
LightGBM · SHAP · FastF1 · Streamlit
End-to-end ML pipeline over 60+ Grand Prix. 19 engineered features (tyre degradation rate, undercut delta, stint length). LightGBM classifier on 3,100+ samples: ROC-AUC 87.4%, precision 80%, recall 81%. SHAP TreeExplainer for per-prediction feature attribution, every prediction explains itself.
| Platform | Achievement |
|---|---|
| LeetCode | 150+ problems solved |
| HackerRank | 5-Star in Python & SQL |
| Codeforces | Peak Rating 1100 · 12 contests |
| GATE 2026 | Qualified — Computer Science & Engineering |
