Parv Khatri

Parv Khatri created Task Tracker so a developer could plan work with several coding agents and review their changes on one board. He is a senior software developer at SAP in the San Francisco Bay Area, where he works on cloud applications and the infrastructure that serves language models. Earlier, at Tata Consultancy Services, Parv led a team that automated SAP system administration across more than 300 servers. He completed his studies at New York University in 2021. Career and education

A shared task board for people and agents

Parv built Task Tracker because each additional agent gave him another stream of updates to follow. The board keeps tasks in a local SQLite database. Through the Model Context Protocol, agents such as Claude Code can read the same list as the developer, break a task into subtasks and log progress. By default, proposed edits go to an approval inbox. If an agent proposes a new title and the developer renames the task before reviewing it, the board remembers the title the agent saw. It flags the stale suggestion and leaves it unchecked, so the developer can approve the other edits and keep the newer title.

Training and serving language models

In 2026, Parv began writing the LLM Fine-Tuning Guide for developers and operations engineers new to model training. He starts with the problem a developer needs to solve. If a model lacks facts, he recommends retrieving them. If its answers still drift from the required style or format after careful prompting, fine-tuning may help. For a developer working with a consumer GPU, he explains how to train a small adapter on a quantized model, evaluate the result and deploy it.

Parv and Harshul Jain helped build AdmitPerf to test the rules a self-hosted model service uses to admit, defer or refuse requests before they take up GPU memory. The tool records the engine measurements behind each decision, so an operator can see whether a rule had a chance to act. Requests can pile up in the queue while the model's key-value cache is nearly empty. In that situation, a rule that waits for the cache to fill will keep admitting work. Parv also argues for counting every incoming request when judging a policy. A service that refuses 95% of requests and serves the rest on time has met its latency target for only 5% of callers' requests. Parv and Harshul on admission control

When retrieved sources disagree

In October 2026, Parv, Harshul Jain, Mehak Kapoor and Tanya Sah studied how an assistant can distinguish an outdated source from a question with several valid answers. On the CONFLICTS benchmark, they tested eleven decision models that assign probabilities to a fixed set of explanations in a single pass. Breaking the classification into a series of yes-or-no questions reduced accuracy for every model. The probabilities proved more useful for deciding when to ask a stronger model: when each classifier sent its least confident quarter of cases to Gemini 3.8 Flash, eight of the eleven combinations were statistically tied with Gemini alone at 25–37% of its billed cost. The team released the code, prompts and model outputs alongside the paper.