Agentic AI • Backend • LLM Infrastructure — I build autonomous agent systems and the distributed backends that run them in production.
I'm an Agentic AI, Backend & AI Engineer working at the intersection of autonomous agent systems, distributed backends, LLM inference, and cloud-native infrastructure. I design and ship agentic systems -- multi-agent orchestration, tool use, planning, and memory -- backed by high-throughput, fault-tolerant services (Go, FastAPI, gRPC, Kafka) and the serving layers, pipelines, and GPU infrastructure that make them reliable at scale.
Currently working on:
- Agentic AI systems -- multi-agent orchestration, planning, tool use, reflection, and long-horizon task execution
- MCP (Model Context Protocol) servers and integrations for tool-augmented, context-aware agents
- Autonomous agent memory -- short-term, episodic (vector), and long-term (knowledge graph) context management
- Low-latency, fault-tolerant backends for LLM, GenAI, and agentic workloads
- LLM serving infrastructure -- vLLM, TensorRT-LLM, KV cache optimization
- GPU workload orchestration on Kubernetes with autoscaling and workload-aware scheduling
Agentic AI & LLMs
Backend & Distributed Systems
Cloud & Infrastructure
Data & Storage
Core Competencies — deeper breakdown of what I build
Agentic AI
- Agent architectures -- ReAct, plan-and-execute with dynamic replanning, reflection/self-critique loops, tree-of-thought
- Multi-agent orchestration -- supervisor/worker patterns, specialized sub-agents, inter-agent messaging, shared scratchpads
- Tool use -- dynamic tool registries, schema/function calling, permission-scoped and sandboxed execution
- Agent memory -- short-term buffers, episodic vector memory, long-term knowledge graphs, context compaction
- MCP (Model Context Protocol) -- server development, tool registries, context injection, Claude & GPT integrations
- Evaluation & safety -- trajectory/trace evaluation, guardrails, loop/cost budgeting, failure recovery
LLMs & GenAI
- RAG systems -- hybrid retrieval (dense + sparse), cross-encoder reranking, query decomposition, RAGAS evaluation
- Fine-tuning -- LoRA, QLoRA, RLHF, DPO, dataset curation, experiment tracking (MLflow, W&B)
- Inference optimization -- vLLM, TGI, TensorRT-LLM, KV cache tuning, continuous batching, speculative decoding, quantization (GPTQ, AWQ)
Distributed Systems & Backend
- System design -- microservices, event-driven architecture, CQRS, domain-driven design
- Scalability -- horizontal sharding, consistent hashing, load balancing, distributed caching
- Resilience -- circuit breakers, bulkhead isolation, rate limiting, exponential backoff, graceful degradation
- Observability -- distributed tracing (OpenTelemetry, Jaeger), metrics (Prometheus, Grafana), structured logging
| Project | What it does | Stack |
|---|---|---|
| Autonomous LLM Agents | Multi-agent system with ReAct + plan-and-execute, MCP tool access, multi-tier memory (buffer / vector / knowledge graph), and sandboxed Docker execution | LangGraph, MCP, GPT-4, Claude, Weaviate, Docker |
| Project | What it does | Stack |
|---|---|---|
| Enterprise RAG System | Hybrid retrieval + cross-encoder reranking, multi-LLM router with fallback, semantic caching (~40% cost cut), RAGAS eval & guardrails | FastAPI, LangChain, Weaviate, vLLM, Redis |
| AI Gateway Microservices | Go gateway for LLM traffic: multi-provider routing, semantic cache, token rate limiting, full OpenTelemetry + cost attribution | Go, gRPC, Redis, PostgreSQL, OTel |
| GPU Kubernetes Platform | Multi-tier GPU node pools, KServe/vLLM serving, DCGM monitoring, MIG partitioning, GitOps with autoscaling | Terraform, K8s, Helm, ArgoCD, KServe |
| LLM Fine-tuning Platform | LoRA/QLoRA fine-tuning, dataset pipeline, eval suite, MLflow/W&B tracking, auto-deploy to vLLM | PyTorch, PEFT, DeepSpeed, MLflow |
| LLM Engineering Fundamentals | Transformer from scratch in NumPy: BPE tokenizer, RoPE/ALiBi, multi-head attention, decoding strategies, 148+ tests | Python, NumPy |
| Project | Focus |
|---|---|
| MicroGPT Reward-Gated | Four-quadrant reinforcement -- intrinsic reward matrices inside a transformer with real correctness signals |
| MicroGPT 1-bit | 1-bit GPT (BitNet b1.58) in pure Python -- ternary weight quantization with zero dependencies |
Attention Variants in Minimal GPT — comparing 6 attention mechanisms
Systematic comparison of 6 attention mechanisms in a zero-dependency GPT (~12K parameters) across 3 rounds of experiments with identical hyperparameters, seeds, and data.
Variants tested: standard multi-head (baseline), gated value attention, relational scoring, salience weighting, context-augmented attention, difference attention.
Key findings:
- The gated-context variant consistently outperformed standard attention (lower loss, faster convergence)
- Salience weighting helped most on longer sequences but added ~15% overhead
- Difference attention underperformed on small models but may scale better (untested at larger scales)
- Daily Digest: ai-daily-digest -- auto-curated trending AI papers, repos & blogs
- AI Bookmarks: sanketny8.github.io/ai-bookmarks -- curated AI resources synced from browser
- Open to: agentic AI systems, system design discussions, backend architecture, open source collaboration, code reviews