AI/ML engineer-in-training building inference systems from scratch. B.Tech CS (AI & ML), Jain University, Bengaluru. Graduating 2027.

I like understanding systems down to the metal instead of stitching APIs together. Most of what I build is local-first, privacy-first, zero external dependencies, from data prep through inference serving.

inferd: a from-scratch LLM inference server.

• QLoRA fine-tuning pipeline on Qwen3-27B for domain adaptation.
• Speculative decoding with rejection sampling. Draft model distilled via sequence-level KD against the fine-tuned target's own generations.
• Paged KV-cache, implemented in Triton.
• Continuous batching. Benchmarked throughput scaling ~6x from one to eight concurrent users.

prr: a self-hosted code-review bot. Pairs an Ollama model with ruff, mypy, and bandit. Results normalized through a typed Pydantic schema and posted as inline PR comments on GitHub.

Predictive Ghost-Text Daemon (ideation stage): always-on autocomplete via a resident Unix-socket inference process. Sub-100ms latency budget, prefix caching, shell-history personalization through a fine-tuning pipeline with secrets-scrubbing.

Python, Java, SQL, PyTorch, TensorFlow, scikit-learn, XGBoost, CUDA, Triton, MLX, LoRA/QLoRA, speculative decoding, paged KV-cache, continuous batching, RAG (FAISS/ChromaDB), FastAPI, Docker, Ollama, Jenkins, Git/GitHub Actions.

Theory foundation: constrained optimization, KKT conditions, multi-armed bandits.

AI/ML engineering internships with a path to full-time. Specifically roles where inference performance, model serving, or applied ML infra is the job itself.

• LinkedIn: likhith-v