Topu Kumar Mondol

I build backend and AI systems in Python — agentic pipelines, retrieval, and the infrastructure that keeps them running.

B.Sc. in Electrical & Computer Engineering, Rajshahi University of Engineering & Technology

Graduating October 2028 · CGPA 3.51 / 4.00 · Rajshahi, Bangladesh

Projects

OpsIQ — AI-powered ops intelligence platform

Python · FastAPI · LangGraph · LangChain · FAISS · PostgreSQL (async SQLAlchemy) · 2026

  • Verified root-cause classification accuracy by replaying three documented production incidents — GitLab 2017, Cloudflare 2019, AWS 2020 — through the pipeline, asserting both category matches and domain-specific terms drawn from the published postmortems.
  • Built a 5-node LangGraph DAG where log analysis and timeline extraction run in parallel and join at root-cause inference, over three temperature-tuned LLM instances (0.7 chat, 0.3 RAG, 0.1 postmortem) isolating generative, grounded and deterministic tasks.
  • Replaced per-request session lookups with stateless HMAC-SHA256 tokens carrying a version field, so bumping one database column invalidates every issued token without the server storing any of them.
  • Made restarts transparent by persisting FAISS stores to disk under per-user, per-session paths and mirroring conversation state to PostgreSQL, restoring summary buffers and recent messages on reconnect.
  • Automated the release path with a GitHub Actions pipeline that runs the 123-test suite on every push and deploys only when tests pass, replacing manual redeploys.

ContentPlatform — personalized recommendation engine

FastAPI · PostgreSQL · SQLAlchemy 2.0 · JWT + OTP auth · APScheduler · Docker · Pytest · Locust · 2025

  • Sustained a 99% success rate at 100 concurrent users (~800ms p95) by load testing with Locust from 10 to 500 users, and root-caused the failure at 500 as database connection pool exhaustion.
  • Stopped any one category from dominating the feed by capping it at 3 of the top 5 slots per page, using a slot-based ranker with softmax sampling at temperature 0.7 and 10% random injection.
  • Kept recommendations following current behaviour rather than old activity by decaying interaction weights (view 1, like 3, save 5) on a 30-day constant, recomputed for every user in hourly batches.
  • Delivered a full 20 posts per page with zero repeats across the entire feed by building all pages in one pass against a shared seen-set, seeded so pagination stays stable.
  • Cut admin block latency from up to 7 days to a single request by re-checking user state in the database on every authenticated request instead of trusting the JWT.

Technical skills

Languages
Python, C++, C, SQL
AI and ML
LangChain, LangGraph, RAG, retrieval systems, vector databases (FAISS), embeddings, LLM integration
Backend and databases
FastAPI, SQLAlchemy 2.0, PostgreSQL, REST APIs, JWT auth, APScheduler
Cloud and infrastructure
Docker, CI/CD with GitHub Actions, Hugging Face Spaces, Neon managed PostgreSQL, containerized deployment
Tools
Git, GitHub, Pytest, Locust, Pydantic

Certifications

Machine Learning Specialization

DeepLearning.AI / Andrew Ng, Coursera · December 2025

CS50x: Introduction to Computer Science

HarvardX, edX · October 2024

CS50P: Introduction to Programming with Python

HarvardX, edX · 2024

Algorithms

Solved 400+ data structures and algorithms problems across LeetCode, GeeksforGeeks and Codeforces, spanning arrays, trees, graphs, binary search and dynamic programming.

Open to Summer 2027 software engineering internships. Reach me at topukumar538@gmail.com.