Building production AI systems and the infrastructure underneath them — multi-agent pipelines, LLM inference serving, and distributed systems.
- 🔭 I'm currently working on AI Agents, RAGs, LLM inference serving, and distributed systems infrastructure
- 🌱 I'm learning System Design, Concurrency, Backend Engineering, and AWS
- 💬 Ask me about AI Safety & Evals, Multi-Agent Systems, LLMs, RAG pipelines, LLM Inference Serving, or Backend/Distributed Systems
- 🎯 2026 Goal: Break into AI/ML engineering and ship a production agent system at scale
- ⚡ Fun fact: Big Tennis and Chess enthusiast — I'll happily watch a 6-hour live chess stream
Languages
AI / ML
Safety & Evals
Systems & Infra
Databases & Tools
| Project | Description | Stack | Link |
|---|---|---|---|
| ⚡ vLLM Inference Benchmarking Harness | Benchmarks LLM serving performance — throughput, latency percentiles, GPU memory, $/1M-token cost — for vLLM (fp16/AWQ/GPTQ) vs. a naive HuggingFace baseline; found and fixed a kernel-config bug causing a 4.2x throughput regression, recovering to 2187.9 tok/s at concurrency 32 | vLLM, Python, Prometheus, Grafana, Docker | Repo |
| 🗃️ lsmdb | LevelDB-style LSM-tree key-value store built from scratch in Go — WAL, per-SSTable Bloom filters, size-tiered compaction — benchmarked head-to-head against SQLite across 10 workloads | Go | Repo |
| 🧮 Mini Feature Store | Point-in-time correct ML feature store with Parquet offline storage, Redis online serving, and a YAML feature registry — built to prevent training-time data leakage in historical joins | Go, Redis, Parquet, PySpark | Repo |
| 📐 PEFT Reasoning Comparison | Controlled comparison of LoRA, DoRA, and IA3 against zero-shot/few-shot baselines on GSM8K math reasoning under a fixed compute budget, isolating the effect of adapter method from training scale | PyTorch, PEFT, Transformers | Repo |
| 🔐 SentinelMesh | Policy-enforced multi-agent orchestration with a custom Go Deep Prompt Inspection proxy — blocks prompt injection, PII leakage, and RBAC violations in real time | LangGraph, Go, ChromaDB, FastAPI, Docker | Repo |
| 🚨 Incident Response System | Autonomous incident-response pipeline with a human-in-the-loop approval gate — diagnoses production errors, generates & tests code patches, drafts customer replies, and opens a GitHub PR before any code ships | LangGraph, Groq, ChromaDB, Streamlit, GitHub API | Repo |
| 🏥 Clinic Voice Agent | Real-time phone voice agent for clinic appointment booking built on raw Twilio Media Streams — no hosted platform, full STT → LangGraph → TTS pipeline | Twilio, Deepgram, ElevenLabs, LangGraph, FastAPI | Repo |
| 🔒 RAG RBAC Chatbot | Enterprise knowledge-base chatbot with namespace-scoped retrieval, JWT auth, hybrid BM25 + vector search, PII redaction, and Llama Guard safety — RAGAS faithfulness 0.91 | LangChain, ChromaDB, FastAPI, Redis, Streamlit | Repo |
| ♟️ Chess Fundamentals RAG Tutor | Multi-modal RAG chess tutor over Capablanca's Chess Fundamentals — retrieves by text AND board-position (FEN) similarity, then a 4-agent LangGraph pipeline classifies position, checks a proposed move, and coaches with visible hedging on unverified content; 91.7% FEN-extraction accuracy, retrieval hit-rate 0.875, though Faithfulness (0.65) missed its 0.75 target — investigated and attributed to judge-model scoring limitations, not a prompt defect | LangGraph, Groq, Gemini Vision, ChromaDB, FastAPI, Streamlit | Repo |
| 🧬 Non-Linear Code Refactoring | Evaluated discrete diffusion decoding (LLaDA 1.5 + Path-Guided Unmasking) against autoregressive baselines on controlled code-refactoring benchmarks — reached 1.0000 coherence / 0.9986 AST similarity vs. a DeepSeek-Coder-6.7B AR baseline that completed 1 of 2 RefactorBench tasks | LLaDA 1.5, PyTorch, Transformers | Repo |
I'm always open to interesting conversations, collaboration, or just a friendly chat.
📫 Reach me at aditya.rallapalli0902@gmail.com or connect on LinkedIn.