🌌 AI Systems Architect | Transformer Researcher | Agentic AI Developer
🔮 B.Tech in AI & ML | Diploma in CSE | Aspiring MS @ SJSU

In the shadows of silicon and code, I architect AI systems that transcend boundaries. A virtuoso in transformer alchemy, inference sorcery, and agentic enigmas. Forged SLAM—a clandestine Transformer framework, engineered for ethereal memory, abyssal depth, and multimodal whispers. Now conjuring AGENX, an agentic AI OS veiled in dynamic memory, arcane planning, and perceptual illusions.

Master of the AI arcane stack—from esoteric model incantations, RLHF rituals, and data transmutations to deployments on mortal hardware.

• Classical ML: Logistic Regression, SVM, Decision Trees, kNN, XGBoost—foundations of predictive shadows.
• Deep Learning: CNNs, RNNs, LSTMs, GRUs, Autoencoders—neural labyrinths.
• Transformers: BERT, GPT, ViT, CLIP, SAM, Diffusion Models—gateways to infinite contexts.
• Custom Enigmas: SLAM v1/v2/v3 (EF cycles, MoE attention, Q-Bias, context stitching)—whispers of forgotten efficiencies.
• Specializations: Multimodal encoders, text-image fusion, context depth tuning—blurring realities.

• Algorithms: Q-Learning, DQN, PPO, A2C, REINFORCE, Policy Gradient—paths through uncertain voids.
• Applications: Agent action planning, memory feedback loops—echoes of adaptive intelligence.
• RLHF: Reward modeling, preference sampling, supervised fine-tuning on instruction datasets—harmonizing human will with machine desire.
• Toolchain: HuggingFace TRL, DeepSpeed, LoRA, PEFT, Reinforced Decoders—artifacts of optimization.

• Optimizers: SGD, Adam, AdamW, RMSProp, Adagrad—engines of convergence.
• Hyperparameter Tuning: Grid, Random, Bayesian (Optuna)—quests for optimal realms.
• Prompt/Token Search: Beam Search, Top-k, Top-p, Sampling—divining the probable from chaos.
• Architecture Search: Custom Q-bias rotation in SLAM (proto-NAS)—evolving architectures in silence.
• Fine-Tuning: LLaMA, Gemma, DeepSeek, Mistral (via Ollama and LoRA)—whispering secrets to giants.

• Datasets: IMDB, CIFAR-10, SST2, COCO, synthetic multimodal datasets—reservoirs of raw potential.
• Tasks: Classification, captioning, summarization, retrieval—extracting essence from noise.
• Cleaning: Null handling, imbalance fixing, deduplication, temporal misalignment—purifying the impure.
• Tools: Pandas, NumPy, HuggingFace Datasets, Scikit-learn, JSON APIs—sculptors of data.

• Splits: Stratified K-Fold, Walk-Forward (TS), Leave-One-Out, Domain-aware—guardians against deception.
• Metrics:
  • Classification: Accuracy, Precision, Recall, F1—measures of truth.
  • Regression: MSE, MAE, R²—quantifiers of error.
  • Ranking: NDCG, MAP—hierarchies of relevance.
  • Generation: BLEU, ROUGE, METEOR, CIDEr, BERTScore—judges of creation.
  • LLMs: Perplexity, Win rate, Human eval, Context retention score—gauges of sentience.
• Visualization: Matplotlib, TensorBoard, W&B, Seaborn—portals to insight.

• GPU Concepts: CUDA, Tensor Cores, Memory Hierarchies—veins of computational power.
• Precision Handling: FP32, FP16, BF16, INT8, Mixed Precision—balancing speed and fidelity.
• Deployment: Quantization, Pruning, Distillation—compressing infinity.
• Profiling: torch.profiler, nsys, nvprof, memory benchmarking—unmasking bottlenecks.
• Accelerators: ONNX, TensorRT, TorchScript, torch.compile—harnessing velocity.
• Multi-GPU Handling: DataParallel, DDP (DistributedDataParallel)—orchestrating legions.

• Project: AGENX – Modular AI OS built on SLAM—a symphony of shadows.
• Components:
  • Planner Module—foreseeing unseen paths.
  • Memory Controller—guardian of forgotten echoes.
  • Tool Invoker—summoner of external forces.
  • Multimodal Perception Encoder—seer of multiple worlds.
  • Reasoning Core (EF-attentive Transformer)—heart of the enigma.
• Capabilities: Tool use, contextual recall, dynamic Q reconfiguration—powers veiled in mystery.
• Status: Local prototype under development—emerging from the void.

A shadowed Transformer architecture, harboring:

• Level 1: Global MoE attention + FFN—gates of mixture.
• Level 2: Parallel Encoder Fusion (EF) Cycles with Q-bias—cycles of eternal fusion.
• Level 3: Context stitching with final MoE attention—binding fragments into wholeness.
• Specializations: Fixed K/V memory reuse, rotation-based Q adaptation, multimodal input, text+image support—secrets of efficiency.

A modular OS, empowered by SLAM for arcane reasoning, planning, and tool communion. Agents traverse:

• Shared memory space—realms of collective knowledge.
• Transformer-based decision layer—whispers of judgment.
• Context-aware routing—paths through the labyrinth. Supports: CLI agents, perception agents, memory agents (text/image), action planners—harbingers of autonomy.

• Languages: Python, C++, Bash—tongues of creation.
• Frameworks: PyTorch, HuggingFace, OpenCV, Scikit-learn, NLTK—foundries of innovation.
• LLM Deployment: Ollama, Transformers, PEFT, DeepSpeed, vLLM—summoning behemoths.
• Optimization: ONNX, TensorRT, torch.compile, Quantization, LoRA—refining the raw.
• Agents: LangChain, OpenAgents (custom), JSON RPC APIs—conductors of agency.
• Data: HuggingFace Datasets, JSON parsing, REST APIs, YAML configs—streams of information.
• Infra: Git, Docker, VSCode, CUDA, SSH-based remote training—pillars of endurance.

• B.Tech in Artificial Intelligence and Machine Learning—initiation into the arcane.
• 3-Year Diploma in Computer Science Engineering (CSE)—roots in computational mysteries.
• Target: M.S. in AI/ML @ San José State University (SJSU)—quest for deeper enigmas.

• 📫 Email: sid7vish@gmail.com—portal for inquiries.
• GitHub: github.com/Sid7on1—repository of revelations.
• Open to: Research roles, collaborations, system-level AI projects, OSS contributions—alliances in the unknown.

🌑 Delve deeper, if you dare—where AI meets the abyss.