Sahaj Saliya

B.Tech in Information & Communication Technology (ICT) · CGPA 8.5 / 10 · Gandhinagar, Gujarat

Internship · Gurugram, Haryana · Hybrid

• Worked on 6G research and development: network architectures, AI agents and communication systems.
• Built reusable AI components (an LLM Gateway, a RAG Runtime, an AI Memory Layer and a Knowledge Base) for scalable agent execution, context management and workflow orchestration.
• Applied generative AI, LLM, RAG and agent-based methods to decision-making, context-aware retrieval and workflow automation for 6G network architectures.

Freelance · Remote

• Built 2 autonomous trading-signal systems with agentic AI and RAG that read market news, extract sentiment and emit structured signals in real time.
• Cut false-positive trade signals by 37% with filtering rules and confidence scoring.
• Built a CrewAI research pipeline on GPT-4, the Bloomberg API, Financial Modeling Prep and Alpha Vantage for automated earnings analysis.

0.992 AUROC, lesion vs normal (CrohnIPI). Reading a capsule-endoscopy study means checking tens of thousands of small-bowel frames by eye. An ensemble of 10 EfficientNet-B2 models (5 folds × 2 seeds) at the native 320 px finds lesions in 7 classes, grades severity in 4 tiers and draws a heat-map over each lesion.

PyTorch · EfficientNet-B2 · ONNX Runtime · FastAPI · Next.js · Docker · github.com/SahajIVVIX-1/Crohn-Disease-Early-Prediction

0.81 RAGAS faithfulness. Enterprise questions mix document lookups, database queries and actions such as sending email. A LangGraph Supervisor → Worker → Validator graph routes each query to Corrective RAG, NL-to-SQL, a human-in-the-loop email tool or chat.

LangGraph · LangChain · FastAPI · Qdrant · BM25 · Redis · Next.js · RAGAS · github.com/SahajIVVIX-1/Multi-Agent-RAG

0.967 PPO grader score on hard-clean. Cleaning a messy table takes many small, ordered decisions. An OpenEnv environment over pandas tables with reproducible (task, seed) episodes and a hidden clean ground truth.

OpenEnv · Stable-Baselines3 PPO · Llama 3.1 8B · Pydantic · FastAPI · Docker · HF Spaces · github.com/SahajIVVIX-1/open-env-nuclei

9 models benchmarked, explained with SHAP + LIME. A model that predicts which students will struggle is only useful if teachers can see why it made each prediction. Pandera-validated data and a ColumnTransformer pipeline feed a benchmark of 9 models: linear, tree ensembles, XGBoost, CatBoost, LightGBM, a PyTorch MLP, a PyTorch LSTM and FLAML AutoML.

scikit-learn · XGBoost · CatBoost · LightGBM · PyTorch · SHAP · LIME · MLflow · Docker · github.com/SahajIVVIX-1/student-performance-platform

5 stability fixes over the 2015 paper, each documented. The 2015 DCGAN is famously unstable, and most re-implementations show only a grid of blurry samples, never what the networks learned inside. Keeps Radford et al.'s all-convolutional architecture but swaps in an LSGAN loss, Spectral Normalization, a Xavier/Kaiming initialisation, discriminator dropout and label smoothing.

PyTorch · LSGAN · Spectral Norm · Matplotlib · Jupyter · github.com/SahajIVVIX-1/Modern-DCGAN-Reproducibility