MS Computer Science at UC San Diego, finishing December 2026. I work on agentic systems, LLM evaluation, and the unglamorous part of getting a model from a notebook to something people actually depend on.
Most of what I build lives in lab and company repositories rather than here, so this profile is a pointer more than a portfolio. The full writeups are at pranavsoma.me.
Agentic Systems Lead, Laboratory for Emerging Intelligence (Teradata-funded) Leading 6 engineers on an agentic workbench that runs staff and faculty workflows at UCSD, in pilot with 100+ users in the Jacobs School of Engineering. 60+ MCP servers covering the university's ~600 APIs, retrieved or generated by pipeline. A planner emits workflow steps and their dependencies; an executor harness runs them deterministically and degrades gracefully when a step fails. There is a refinement loop on top that splits every skill into a description tuned for retrieval and a body tuned for execution, optimized against latency, token spend, and skill-selection accuracy.
Smart Learning Hub Fine-tuned tutoring models (LoRA, SFT, RLHF) now running in 10 courses across UCSD, SDSU, and Cal Poly Pomona. The interesting part is the reliability layer: a reviewer model gates every response before a student sees it, and a meta-tutor rewrites the tutor's prompt from long-term signal, backed by a 5-model panel of judges. A controlled study over 600+ students measured 64% fewer TA overtime hours and 87% higher engagement.
M.S. thesis, computational redistricting Three congressional districting generators over Census 2020 data at 158K to 5.5M blocks per state: XGBoost with multilevel graph partitioning, a Sequential Monte Carlo sampler using spanning-tree balanced cuts, and greedy DP with Kernighan-Lin refinement. Plus the fairness metrics that grade them. Reproduces official district counts for AL, CA, and TX. Private while the thesis is in progress.
- ServiceNow, software engineer intern two summers running. Migrated the AI Voice Agents platform from AWS Lex V1 to V2 ahead of end-of-life (45% lower agent-response latency), then built the conversation-history analytics platform that shipped to 24+ enterprise instances handling 10K+ voice calls a day. Got a report page from 7 seconds to under 1 second over millions of records.
- Qualcomm Institute / Calit2, built AILA, a health assistant grounding diagnosis in a vector store of UMLS codes, with local Qwen models fine-tuned on the San Diego Supercomputer Center against a physician-authored eval set.
- Triton Software Engineering, engineering manager for a nonprofit fundraising platform (11 developers, raised $75K), and before that shipped a housing portal for LA County's sole provider of those services.
Python, C++, C, TypeScript, Java. PyTorch, Transformers, XGBoost, OpenCV. LLM fine-tuning, RAG and GraphRAG, MCP, evals. FastAPI, React, Docker, Kubernetes, GCP, AWS. Postgres, MongoDB, Parquet.