๐ B.Tech CSE (Data Science) @ VIT Vellore โ CGPA 8.47, 2023โ2027 ๐ผ Full Stack Intern @ Beryl Systems โ shipped 5 production Go REST APIs for ~500 daily users and cut p95 latency ~7x (850 ms โ 110 ms) ๐ Building backend and distributed systems, and applied ML / LLM systems ๐ฏ Ship production-grade software end to end, and measure it until the numbers hold up
Java cache server (~18K LOC, 6 Maven modules) with O(1) LRU, LFU and self-tuning ARC eviction.
- 3 pluggable concurrency modes (global RW lock, 16-way striping, lock-free reads), benchmarked with JMH
- Non-blocking Netty server speaking a custom protocol and RESP (works with redis-cli/Jedis): ~65K ops/s at p99 โ 0.2 ms
- Hit rates on a 256-entry cache, 1,000 Zipf keys: LRU 74.2%, LFU 81.0%, ARC 76.6%; LFU 61.7% vs. LRU 51.0% on scan-heavy loads
- Client-side consistent hashing (150 virtual nodes) moves only ~1/N of keys on node join/leave; append-only log + atomic snapshots
- 500+ JUnit tests, CI/CD, Docker image, 3-node compose cluster
Tech: Java 17, Netty, JMH, Docker, GitHub Actions
GitHub ยท docker pull himanshmunjal/jcache
Full-stack Go (Gin) shortener with a layered handler/service/repository design.
- Redis cache-aside redirects + goroutine click tracking (device, GeoIP, hashed IP) with no added redirect latency
- QR codes encode a hidden short URL, so every scan flows through the redirect pipeline; scan log and counter commit in one DB transaction
- Atomic sliding-window rate limiter as a Redis Lua script that fails open
- JWT auth with bcrypt and rotating refresh tokens stored only as SHA-256 hashes
Tech: Go, Gin, PostgreSQL, Redis, React, Tailwind CSS GitHub
Mixed-precision KV cache (FP16/INT8/INT4) that moves tokens between tiers by importance instead of dropping them. A drop-in transformers.Cache.
- ~40% KV memory saved at โค0.16% perplexity loss on Qwen2.5-1.5B and Phi-3-mini; up to 56% at 0.20% (1.94 GB โ 1.15 GB, 8K context)
- Per-channel key / per-token value INT4 quantization cut key reconstruction error 17.3%
- 3 importance scorers (H2O, VATP, KeyDiff); ~1.8x faster adaptive decoding (12 โ 22 tok/s)
- Leak-proof eval (perplexity, Needle-in-a-Haystack, LongBench); matched FP16 retrieval in 17 of 18 cells
Tech: Python, PyTorch, HuggingFace Transformers GitHub
Picks SMOTE, ADASYN, SMOTE-Tomek or SMOTE-ENN per minority-class cluster, feeding a confidence-weighted RF + XGBoost + LightGBM ensemble.
- Leakage-free study on 28 datasets, 9 variants, 3,780 CV folds (Friedman/Nemenyi, Holm-corrected Wilcoxon, Bayesian equivalence tests)
- Beats SMOTE-ENN (Bayesian P โฅ 0.96) but is practically equivalent to plain SMOTE; resampling mainly helps at โค75 minority samples
- Found and fixed 5 evaluation flaws that inflate new methods; IEEE-format paper draft
- Reproducible MLOps stack (DVC, MLflow, Prefect, FastAPI, Docker) with 130+ pytest tests in CI
Tech: Python, scikit-learn, XGBoost, LightGBM, MLflow, DVC, Docker GitHub
Hybrid RAG over source code for file/line-cited Q&A and blast-radius impact analysis.
- Tree-sitter AST parsing across 5 languages, dual-vector Qdrant index (semantic + node2vec call-graph) and cross-encoder reranking
- Diagnosed a retrieval failure by measuring embedding quality directly, swapped in a contrastively trained model and recalibrated the confidence gate
- Async Celery/Redis ingestion; audited on gin, gson, fastapi and react-use, fixing 20+ parsing and grounding bugs
Tech: Python, FastAPI, Qdrant, Tree-sitter, Celery, React, Docker GitHub
8-stage DuckDB SQL pipeline over 13M GitHub Archive events (1.14M PRs) modeling first-touch to second-PR for 17K contributors across 297 repos.
- 35% of first-time contributors return within 30 days (29% if the first PR merged vs. 17% if not)
- Fast first response is a repo-level effect, not a PR-level one: pooled Cox HR 1.69 drops to 0.99 within repos
- Power analysis for a repo-randomized experiment (~663 repos per arm to detect +5 pp)
Tech: Python, DuckDB, SQL, lifelines, statsmodels, FastAPI, Streamlit GitHub
Languages: Java Go Python SQL JavaScript R
AI / ML: PyTorch HuggingFace scikit-learn XGBoost LightGBM RAG Qdrant ChromaDB
Backend & Systems: Netty Gin FastAPI REST WebSockets JWT Consistent Hashing Multithreading
Data & MLOps: PostgreSQL Redis BigQuery DuckDB MLflow DVC Prefect Celery Docker GitHub Actions
Frontend: React Tailwind CSS
- ๐ผ Full Stack Intern, Beryl Systems โ 5 production REST APIs (Go, JWT, RBAC) for ~500 daily users; p95 latency 850 ms โ 110 ms; fixed 4 security vulnerabilities (OWASP)
- ๐ฏ Manager, General Enquiry Team, VIT โ led a 15-member team resolving 500+ student queries
- ๐งฉ 200+ problems solved on LeetCode, 50 Days Badge
- ๐ Finishing benchmarks and significance testing for a Spectral-Gated Dual-Attention BiGRU for human activity recognition
- ๐ Learning Apache Kafka, Airflow, and BigQuery for data engineering
๐ Portfolio: himansh-portfolio.vercel.app ๐ผ LinkedIn: linkedin.com/in/himansh-munjal ๐ง Email: munjalhimansh2211@gmail.com
I build systems that work under real load, from lock-free caches to retrieval-augmented pipelines, and I test them until the numbers hold up.


