AVAILABLE · NEW GRAD 2026 · REMOTE / RELOCATE

Building production-grade
intelligence systems.

I'm Veeraj Sai, an AI/ML engineer working at the seam of forecasting, LLM / RAG tooling, and multimodal fairness research. Currently shipping wind-power scheduling for 30+ plants at Greenko and a bias-aware hate-detection dataset for my B.Tech thesis at NIT Jalandhar.

−49% Scheduling MAE 18K Multimodal pairs 30+ Wind plants live KDE Store published NIT Jalandhar · IT '26
View projects → Download résumé ↓ Contact →
GitHub ↗ LinkedIn ↗ Kaggle ↗ veerajsai.me
Portrait of Veeraj Sai Subrahmanyam
● REC · 24.000fps SUBJ · V.SAI NIT-JLD / 2026 ID 0x07A4
Now Forecasting
Research Multimodal · Fair
Latency 15-min · Δt
01 / About

Engineer first, researcher second, builder always.

Final-year IT student at Dr. B. R. Ambedkar NIT Jalandhar, graduating May 2026. My work centers on the unglamorous half of ML: pipelines, evaluation, drift, fairness, and ops, without losing the depth of research.

I write models that run in production (15-min wind scheduling for 30+ plants), and I write papers that sit on real datasets (18K image-text pairs, 8 protected identity axes). I've shipped one tool to the KDE Store, mentored juniors at IOTA Club, and grinded 200+ LeetCode problems on the side.

I'm currently looking for AI/ML, Data Science, NLP, or LLM/RAG engineering roles, globally, remote or on-site.

PythonPyTorchLightGBMvLLM OllamaHuggingFaceStreamlitLinux
/ 01

Forecasting systems

Time-series pipelines that survive contact with the real grid. Ridge + GBM ensembles, 15-minute intervals, drift-aware retraining.

Production
/ 02

LLM & RAG tools

Local Ollama RAG, vLLM serving, Qwen/HateBERT inference, prompt-engineered counterfactual rewriting at 12K-sample scale.

Applied
/ 03

Multimodal AI research

CLIP + EfficientNet fusion, gradient-reversal adversarial debiasing, fairness eval across protected identity categories.

Research
/ 04

EEG / BCI pipelines

Motor-imagery classification on BCI-IV-2a with CNN-LSTM, BiGRU-CNN, and graph-conv variants. Full signal-to-decision pipeline.

Signals
0%
Scheduling MAE reduction · 7.28 → 3.70 MWh
0K
Multimodal image-text pairs curated
0+
Wind plants on live scheduling
0+
LeetCode problems solved
0
Protected identity axes evaluated
0m
Forecast interval · operational granularity
02 / Selected Work

Five projects, five different problems.

Each one shipped, measured, and documented. Not a notebook left to rot.

/01
AI/ML Engineer Intern · Greenko Group

Wind-Power Scheduling at Grid Scale

Built and shipped a 15-minute interval forecasting + scheduling system for 30+ wind plants. Replaced operator-tuned heuristics with a Ridge + LightGBM ensemble, paired with a Streamlit ops dashboard and a local Ollama-based RAG chatbot for site engineers to query weather, schedules, and historical anomalies.

PythonLightGBMRidgePolarsStreamlitOllamaRAG
−49%
Scheduling MAE
3.70
MWh final · ↓ 7.28
30+
Plants live
Case study → Code (private) →
FORECAST · 15-MIN
MAE 3.70 MWhridge + lgbm
/02
B.Tech Thesis · Multimodal Fairness Research

COUNTER-HATE — Counterfactual Multimodal Bias

A framework for bias-aware hate-speech detection over image-text pairs. 18K samples, 6K human-annotated, 12K LLM-guided counterfactual rewrites, synthetic Z-Image-Turbo visuals, and adversarial debiasing via gradient reversal. Evaluated across 8 protected identity axes.

PyTorchQwen2.5-7BHateBERTCLIP ViT-B/32EfficientNet-B0GRLvLLM
0.884
F1 (best)
0.149
FPR · debiased
18K
Multimodal pairs
Spotlight ↓ Repo → Paper draft →
PIPELINE · FUSION + DEBIAS
F1 0.884FPR 0.149
/03
Open-Source · KDE Plasma 6 Widget

CryptoStats · Linux Desktop Crypto Ticker

A lightweight, real-time cryptocurrency widget for KDE Plasma 6, written in QML + curl. Published on the official KDE Store. Built for engineers who want a glanceable ticker without a heavy Electron tab eating their RAM.

QMLJavaScriptcurlKDE Plasma 6Linux
KDE
Store published
Δt
Real-time updates
KDE Store → GitHub →
PLASMOID · LIVE
5 marketsplasma · v6
/04
Coursework + Research · BCI

EEG / BCI · Motor Imagery Classification

Brain-computer-interface pipelines for the BCI Competition IV-2a dataset. Benchmarked CNN-LSTM, BiGRU-CNN, and GCN-based architectures with signal-level preprocessing, channel-wise normalization, and subject-aware cross-validation.

PyTorchSciPyNumPyMNEGCNBiGRU
4-class
Motor imagery
22ch
EEG · 250Hz
Repo →
CH · MOTOR IMAGERY
BCI IV-2acnn-lstm · gcn
/05
Energy Research · Microgrid Optimization

Solar Irradiance + Standalone Microgrid

Solar irradiance forecasting for Jalandhar using NASA POWER hourly data, with ML horizons at 1h, 6h, and 24h. Coupled the forecaster to a standalone PV + wind + battery microgrid optimization to size and dispatch under realistic uncertainty.

Pythonscikit-learnLightGBMNASA POWERPyomo
1h/6h/24h
Forecast horizons
PV+W+BES
Microgrid mix
Repo → Whitepaper →
IRRADIANCE · 24H
NASA POWERpv · wind · bes
03 / Trajectory

A short, steep timeline.

Jun 2025 → Aug 2025 · Hyderabad / Remote

AI / ML Engineer Intern

Greenko Group · Renewable Energy
Owned the model + ops loop for short-horizon wind power scheduling.
  • Designed Ridge + LightGBM ensemble; cut scheduling MAE from 7.28 → 3.70 MWh (−49%).
  • Shipped a Streamlit operator dashboard consumed by 30+ wind plants.
  • Built a local Ollama RAG chatbot over plant manuals + weather logs for on-call engineers.
Jan 2025 → Present · NIT Jalandhar

B.Tech Thesis · COUNTER-HATE

Multimodal Fairness · NLP + Vision
End-to-end framework for bias-aware multimodal hate detection. 18K image-text pairs, 12K LLM-guided counterfactual rewrites, Z-Image-Turbo synthetic images, adversarial debiasing via gradient reversal. Best F1 0.884 / FPR 0.149.
Aug 2024 → Present · NIT Jalandhar

Technical Lead · IOTA Club

Student community of 200+
Ran ML and Linux workshops, mentored juniors through their first end-to-end ML projects, led the GDSC HackMol 6.0 organizing track.
2023 → 2024 · Side work

CryptoStats · KDE Store

QML · Linux desktop
Built and published a real-time cryptocurrency widget for KDE Plasma 6. First time navigating the full open-source release loop: packaging, review, store listing.
2022 → 2026 · Jalandhar

B.Tech, Information Technology

Dr. B. R. Ambedkar NIT Jalandhar · Graduating May 2026
JEE Main rank 15,043 · concentration in ML, signals, and systems.
04 / Stack

A small galaxy of tools I actually ship with.

Hover any node. These aren't just installed, they're battle-tested in projects above.

PyTorch LightGBM Python Qwen 2.5 vLLM CLIP HateBERT Ollama HF Transformers Streamlit Polars RAG scikit-learn SciPy GRL Linux Bash C / C++

Languages

  • Python
  • C / C++
  • SQL
  • Bash · QML

ML / DL

  • PyTorch
  • TensorFlow
  • LightGBM
  • scikit-learn

NLP / LLMs

  • HF Transformers
  • vLLM · Ollama
  • RAG pipelines
  • Prompt engineering

Data & Viz

  • NumPy · Pandas
  • Polars
  • SciPy · Matplotlib
  • Streamlit

Tools / Systems

  • Git · Jupyter
  • Linux
  • Lightning AI
  • Kaggle

Domains

  • Time-series · Energy
  • NLP · Fairness
  • EEG / BCI signals
  • Production ML ops
05 / Research Spotlight

A closer look at COUNTER-HATE.

Pre-print · 2026
B.Tech Thesis

Multimodal Counterfactual Augmentation for Bias-Aware Hate Speech Detection

Hate-detection models systematically underperform on protected identity language, and the bias compounds when text and image are fused. COUNTER-HATE attacks this with three coupled moves.

(1) A curated multimodal corpus of 18,000 image-text pairs with 6,000 human-annotated samples spanning 8 protected identity categories. (2) A Qwen2.5-7B-Instruct counterfactual rewriter that generates 12K minimally-edited paired samples flipping the protected attribute, augmented with Z-Image-Turbo synthetic visuals. (3) A multimodal classifier fusing CLIP ViT-B/32, EfficientNet-B0, HateBERT, and TF-IDF, trained adversarially with a Gradient Reversal Layer for protected-attribute invariance.

0.884
F1 · best config
0.149
FPR · debiased
12,000
LLM counterfactual rewrites
8
Protected identity axes
Stage 1 · Data
Image-Text Pairs
→
Qwen2.5-7B
→
Counterfactual
Annotators ×6K
→
8 protected axes
→
Z-Image-Turbo
Stage 2 · Encode
CLIP ViT-B/32
EfficientNet-B0
HateBERT
TF-IDF
Stage 3 · Fuse + Debias
Multimodal Fusion
→
Gradient Reversal
→
Fair Classifier
06 / Why Hire Me

Fresher on paper, operator in practice.

Most ML grads stop at notebooks. I've already shipped models against real grid SLAs, written research-grade evaluation harnesses, and pushed open source through real release processes.

AI / ML Engineer
Production pipelines
  • Live forecasting at 15-min cadence, 30+ plants
  • Ridge + GBM ensembles, retraining loops
  • Streamlit ops dashboards for non-ML users
Data Scientist
Modeling + evaluation
  • Polars/Pandas at scale, careful CV design
  • MAE / FPR / F1 reporting that matches ops reality
  • Stakeholder-readable visual reporting
NLP / LLM Engineer
Applied LLMs · RAG
  • Local Ollama RAG over enterprise docs
  • vLLM serving + Qwen 2.5 prompt pipelines
  • 12K counterfactual rewrites with quality QA
Research Engineer
Multimodal · fairness
  • Adversarial debiasing with gradient reversal
  • Multimodal fusion across 4 backbones
  • Reproducible eval across 8 identity axes
ML Ops / Platform
Linux · open source
  • Published Plasma 6 plasmoid on KDE Store
  • Comfortable in tmux, systemd, GPU box
  • Mentored 20+ juniors through Linux + ML
Signals / BCI
EEG · time-series
  • BCI-IV-2a, 4-class motor imagery
  • CNN-LSTM, BiGRU-CNN, GCN benchmarks
  • Subject-aware CV and signal preprocessing
07 / Let's talk
Currently open

Have a hard problem in ML, NLP, or forecasting?
I'd love to hear about it.

Best reached over email. I reply within a day. Open to AI/ML, Data Science, NLP/LLM, and Research Engineer roles. Remote, hybrid, or on-site, anywhere.

Email Veeraj → Résumé · PDF ↓ LinkedIn ↗ GitHub ↗
veeraj.sai@gmail.com Kaggle · veeraj16 Jalandhar / Hyderabad · India UTC +5:30