ML Engineer and Software Development Engineer
M.Tech Computer Science, Information Security & Privacy — SVNIT Surat · GATE CS 2025, 93.8 percentile
I work where machine learning meets systems engineering. Training a model is the easy half; the interesting problems live in the pipeline around it — what happens when the classifier is wrong, what it costs, and who can explain the decision afterwards.
Portfolio · LinkedIn · LeetCode · md.humayun7777777@gmail.com
IPS-ML / ShieldNet — my main machine learning project, and the work I would most like to be asked about.
An inline intrusion prevention system is not a classifier with a nicer name. It executes its own predictions: a false positive does not raise an alert for somebody to dismiss, it drops a legitimate connection. So the whole thing is evaluated as a control system — recall and false alarm rate as the headline metrics, three outcomes instead of two, and a SHAP attribution written to an append-only audit log beside every enforcement action.
Trained on 2,497,980 audited CICIDS2017 flows, evaluated on 749,394 held out. Macro-F1 0.8814 at a 0.00135 false alarm rate, 0.432 ms median inference. A model that predicts "benign" for everything scores 82.96% accuracy on the same data and detects nothing, which is why accuracy is not the headline anywhere in that repository.
Hybrid AI-Based Autonomous Drone Swarm Navigation — MeitY-sponsored research at SVNIT Surat. Built BattleWorld, a multi-agent simulator wrapped as a Gymnasium environment with a uniform 12-dimensional state interface, so seven commander architectures could be benchmarked on identical input. The hybrid LLM+PPO controller reached 67.9% action accuracy against 60.7% for the best classical baseline. Paper under review at IEEE GSCON 2027.
| Project | What it demonstrates | Evidence |
|---|---|---|
| ips | Explainable ML in an enforcement path | 2.5M flows · macro-F1 0.8814 · FAR 0.00135 |
| Smart-Home-Automation | A safety-critical embedded system, hardened, verified and empirically evaluated | Dual-core ESP32 · HMAC auth + replay protection · 2 TLA+ models (TLC-checked) · 6k-trial adversarial campaign · 3M-iteration fuzzing · paper draft |
| job-agent | Data pipelines built to degrade, not fail | 11 sources · 2,158 records · 793 → 357 after dedup |
| ddos-5g-sdn-project | Detection and mitigation inside an SDN controller | 54.3 → 31 Gbps measured under flood |
| vae-aevb-reimplementation | Reading a paper closely enough to reproduce it | SSIM 0.678 · PSNR 23.40 dB |
| medical-chatbot | Grounded retrieval that refuses when it should | 1,152-page corpus · local inference |
| Backend-System-design | Class-level design judgement | Spotify and Zomato, modelled from UML |
Every claim in those repositories is traceable to something I can regenerate.
Where a result is weak I say so: the limitations sections are written before the
results, not after, and the deviation register in ips lists all thirteen
places the implementation departed from the plan, with impact ratings.
That habit is the thing I would want a reviewer to notice.
Languages C++ · Python · SQL
Embedded and IoT ESP32 · FreeRTOS · Arduino · MQTT · Home Assistant
ML TensorFlow · Keras · scikit-learn · XGBoost · SHAP · LoRA · PPO
LLM and retrieval LangChain · Pinecone · Ollama · RAG
Infrastructure Docker · AWS (EC2, S3) · GitHub Actions · Linux
Networking SDN · OpenFlow · Mininet · Scapy
Sharpening C++ and system design for SDE interviews, and reading toward cross-dataset validation for the intrusion prevention work — generalisation is an objective there, not yet a result.
Outside coursework I run NITIAN RAHI, an engineering education community of around 52,000 across YouTube, Facebook and Instagram, where I explain ML and systems topics to students. Teaching a thing is the fastest way to find out whether you understand it.