Siddhant Karki

I’m an AI Engineer who builds production systems end to end—from data pipelines and model training to APIs, evals, and deployment. I work below the abstraction layer when needed, with hands-on depth in tokenizer internals, model adaptation, distributed processing, and retrieval augmented generation (RAG). I turn ambiguous AI problems into shipped systems, from voice pipelines and analytics to knowledge infrastructure and language model research.

In my current role at CallD.AI, I own production systems across live-call infrastructure and AI post-call analysis, from ingestion through storage, processing, APIs, and dashboards. I redesigned speaker re-labeling with turn-index maps to cut LLM token usage by 75% and improve processing speed 3x, then reduced cost further via an OpenAI-to-Bedrock migration and Nova Micro prompt caching. I also enforced schema-validated structured outputs with Pydantic to eliminate malformed responses, improved distributed reliability using PostgreSQL row locking with concurrent workers and jittered retries, and built an analytics platform with materialized-view-backed reporting and performance-focused query/index work. My project work extends this mindset—hybrid retrieval with metadata-first MCP toolchains, fully local ONNX inference, automated wiki synthesis with verification status, tokenizer cost remediation, and fine-tuned ASR with cross-domain evaluation.