SAYANTANMANDAL
I design and ship systems — from serverless cloud pipelines to peer-to-peer networking in Rust.
Systems thinker, hands on keyboard.
Final-year Computer Science engineer who treats every project as a system to be reasoned about — not just code to be written.
I work end to end: serverless cloud architectures on AWS, multi-tenant SaaS backends, real-time interfaces in React, and low-level systems programming in Rust. A year of internship experience across an enterprise (Honeywell) and two AI-focused teams (OneSol AI Labs, Elogix) taught me how production systems actually behave — and break.
I care about the parts users never see: clean boundaries, predictable failure modes, and code that the next engineer can read at 2 a.m. I learn by building — deliberately choosing unfamiliar problems and shipping them end to end.
- Distributed & cloud-native systems
- Developer tooling & systems programming
- Applied LLM / retrieval systems
A broad, deliberate toolkit.
Nine languages, picked per problem — not per trend. Grouped by where they live in a system.
Languages
01Backend & APIs
02Frontend
03Data & Stores
04Cloud & DevOps
05Applied AI
06Where I've shipped.
Honeywell
- Re-architected services for cloud portability — moving from Azure to a cloud-agnostic design running across GCP, Alibaba Cloud, and on-prem Windows services so the platform stays available if any single provider fails.
- Built an AI log-analysis agent that pulls logs from ELK and AKS, clones the relevant repo, and localizes the root cause (config, database, or code) — automatically proposing a fix and opening a pull request for code issues.
- Standardized structured logging with Serilog into ELK/Kibana and shipped features behind LaunchDarkly flags for safe rollout and instant kill-switches.
- Cut GitHub Copilot LLM token usage ~80% across internal AI agents via a knowledge-graph context-retrieval and caching layer, and automated a provisioning workflow that cut turnaround time ~40%.
OneSol AI Labs
- Deployed a high-throughput AI inference platform on AWS EC2 serving concurrent GPU-accelerated generation at sub-2s p95 latency, tuning async I/O and memory allocation for hardware utilization.
- Built a Milvus vector-search pipeline using approximate nearest-neighbor indexing, cutting semantic search latency ~65% on a 50K-item corpus over brute-force lookups.
- Implemented async job queuing with shared-memory IPC for ~3× throughput under concurrent load, with Dockerized CI/CD that eliminated manual release steps.
Elogix Software
- Built three production AI systems — an AI resume evaluator, a document-grounded QA chatbot, and an interview-assessment platform — on FastAPI + LangChain + GPT-4 with FAISS semantic search.
- Added multi-format document parsing (PyMuPDF, python-docx) and React front-ends, deployed via Docker and GitHub Actions CI/CD with clean separation of concerns.
Things I've built.
A slice of 30+ shipped projects — chosen to show range across cloud, SaaS, systems, and applied AI.
SentinelGrid
Serverless real-time disaster-monitoring platform
Ingests global natural-event data from NASA's EONET API on a scheduled AWS CloudWatch pipeline into DynamoDB. Fully serverless (Lambda, DynamoDB, API Gateway) with least-privilege IAM and a React dashboard rendering live geospatial visualizations.
Shopify Insights Platform
Multi-tenant SaaS analytics platform
Multi-tenant analytics with strict per-tenant isolation: ingests Shopify customers, orders, and products via the Admin API and HMAC-verified webhooks. Redis caching, a Redis/Bull job queue with exponential-backoff retries, JWT auth, AES-256 credential encryption, and property-based tests on a TypeScript/Express + Next.js stack.
MyVPN
Serverless peer-to-peer VPN in Rust
A peer-to-peer VPN with NAT hole-punching over the iroh QUIC transport and end-to-end TLS 1.3 encryption (Ed25519 keys). Captures full-tunnel traffic through a Wintun virtual adapter with automatic LAN discovery — no central server.
vul-detector
LLM-assisted source-code vulnerability scanner
Clones a repository and runs Tree-sitter AST plus pattern analysis across six languages, mapping findings to CVE/CWE identifiers and generating PDF/JSON reports with LLM-suggested mitigations through a FastAPI backend and a React dashboard.
flux
Native Windows media player, 120+ formats
A production-grade desktop media player built on Electron + Rust with a bundled mpv/FFmpeg core: hardware-accelerated playback of 120+ formats, a 15-band equalizer, HDR tone mapping, and yt-dlp streaming from 1000+ sites.
BitDown
High-performance BitTorrent client
A BitTorrent client built with Tauri + Rust on the librqbit engine (DHT/PEX/UPnP), with adaptive HTTP streaming preview during download, TMDB metadata auto-fetch, and an RSS feed manager — wrapped in a glassmorphic React UI.
Proof of work.
optimized a CNN inference pipeline for ~60% faster execution
B.Tech CSE · Vellore Institute of Technology (VIT-AP)