ML engineer — LLM inference and evaluation. I measure latency, determinism and accuracy on real hardware, and publish the harness alongside the numbers so the result can be checked rather than taken on faith.
MS in Computer Science · 2.5 years of ML engineering · open to ML/AI engineering roles
Portfolio · LinkedIn · vineethkodakandla@gmail.com
| Project | What it does | Live |
|---|---|---|
| bitwise-forensics | Asks whether an LLM serving engine returns the same tokens when requests share engine steps. 2 devices × 12 configurations × 40 prompts × 8 trials on OpenVINO GenAI, with bootstrap intervals on every rate. | results |
| llm-eval-observatory | A living dashboard that evaluates open-weight models nightly — capability drift, prompt-injection robustness, and LLM-as-judge agreement. GitHub Actions is the engine, Vercel the window; the dashboard is the artifact, not a screenshot of one. | dashboard |
| pathwise-ai | SD-WAN management platform: LSTM link-degradation forecasting, autonomous traffic steering, a digital-twin sandbox and intent-based networking, behind a multi-tenant React dashboard. | demo |
| ananta-lab | A governed memory-augmented agent: Ebbinghaus-decay episodic memory, a SHA-256 tamper-evident audit chain, confidence-banded model routing and a Reflexion RAG loop. Zero dependencies. | demo |
| edith-lab | A multi-agent orchestrator — Researcher / Analyst / Critic / Synthesizer — with a Reflexion loop and a Governor policy gate. Framework-free. | demo |
| RAG document intelligence | Production-grade RAG for enterprise document Q&A: semantic chunking, hybrid dense + BM25 retrieval with RRF fusion, FastAPI streaming, RAGAS evaluation. | — |
Serving & inference OpenVINO GenAI, continuous batching, KV-cache quantization, ONNX ML PyTorch, LSTM/sequence forecasting, scikit-learn, bootstrap & interval estimation LLM systems RAG (hybrid dense + BM25, RRF), pgvector, ChromaDB, RAGAS, LLM-as-judge, prompt-injection testing Backend Python, FastAPI, Postgres, Docker Frontend TypeScript, Next.js, React, three.js Infra GitHub Actions, Vercel, scheduled evaluation pipelines
Every project above links to a live deployment or to committed results. If a number appears in a README of mine, the code that produced it is in the same repository.