Reza Shabani

Reza Shabani helped build Replit’s data infrastructure and develop its AI team, whose work included Ghostwriter code completion and training code language models. As Replit’s Head of AI in a May 2023 interview, he explained the decision to move from adapting Salesforce’s CodeGen to training models from scratch—and why useful, fast completions mattered alongside benchmark scores.[1]

Shabani initially worked on Replit’s ability to query and process data at scale. He recalled that even identifying the most-forked project could time out before that infrastructure existed. Moving data into a form that could be processed in minutes supplied the groundwork for training models and fine-tuning them on Replit data. He then recruited and developed the AI/data team, which built Ghostwriter’s completion, code explanation, generation, transformation and chat features, while also supporting search and other large-scale data work.[1]

The original Ghostwriter completion system shipped in August 2022. Shabani described taking a smaller version of Salesforce’s open CodeGen model and applying distillation and other optimizations to reduce its footprint and make it fast enough to serve. Salesforce supplied the underlying model; Replit’s work centered on making it practical for its users.[1]

Customization exposed the limits of that approach. Replit users wrote web code such as JSX and inline CSS, and Shabani said experiments fine-tuning CodeGen did not produce the behavior the team wanted. The team therefore built infrastructure to train models from scratch. Replit’s training article, credited to its AI team, identifies customization, reduced dependency on external providers and cost efficiency as the main reasons for owning that capability.[1][5]

The large training run followed months of infrastructure work and smaller, single-language experiments. Shabani recalled wanting further systematic tests: data preparation could still improve, runs were expensive, and failures could be difficult to diagnose. Amjad pushed the team to train on all available data—“let's just yolo this”—while Michele Catasta proposed repeating the dataset over multiple epochs and supplied evidence addressing the risk of overfitting. Shabani credited that push with compressing work he expected to take two or three months into a week.[1]

The resulting replit-code-v1-3b was a 2.7-billion-parameter code-completion model. Replit reports that it trained on a 175-billion-token dataset repeated over three epochs, totaling 525 billion training tokens, using MosaicML and 256 A100-40GB GPUs. Its custom SentencePiece Unigram tokenizer had a code-specific vocabulary of 32,768 tokens. Shabani explained that a smaller vocabulary tailored to code helped training and inference speed.[6][1]

Shabani credited MosaicML with reducing the operational burden after Replit had already deployed its own training infrastructure. Access to GPUs across cloud providers, recovery from hardware failures and preconfigured training setups let the team spend less time troubleshooting utilization and crashes. Existing data pipelines also enabled rapid preparation of public Replit data for a separate fine-tuned variant, including deduplication and filtering personal information and harmful content.[1]

Shabani distinguished HumanEval from AmjadEval, the team’s name for Amjad’s practical testing. HumanEval asks a model to complete a Python function from its signature and description, then tests the result. Amjad’s judgment also considered contextual usefulness, response speed and whether suggestions made coding feel more productive. Shabani recalled a model with strong HumanEval scores whose practical behavior was poor; plausible-looking but incorrect completions were another reason a model could fail this test.[1]

One demonstration illustrated the difference: given comments describing possible process states, Replit’s model proposed readable constants, while CodeGen initialized a value to zero. Shabani also showed a counterexample in which Ghostwriter incorrectly assumed an argument needed to be a list. These examples made contextual behavior concrete without treating either model as consistently superior.[1]

The released base model’s card reports a self-measured HumanEval pass@1 of 0.219. Shabani separately described improvements after tuning on public Replit data, despite that material including beginner code. He suggested that instructional assignments and answers might help a model learn reasoning, but presented that explanation as a suspicion rather than an established mechanism.[6][1]

In May 2023, Shabani proposed moving from predicting text in one file to an autonomous agent operating the IDE. He wanted to train on sequences of development actions: creating files, making related changes across files, running code and responding to errors. Replit’s operational-transform data recorded editing sequences that could support this work. His proposal concerned a future model that could help build software through actions, rather than completion alone.[1]

Shabani expected some prompt-engineering workarounds to diminish as models improved. He compared them to middleware that would have translated a user’s search request into keywords for an early search engine. His prediction was qualified: communicating intent and context to more human-like models would still require some form of prompting.[1]

He also argued that rapid AI development made questions about automation urgent, including how to let people share its economic benefits and how AI could “help people as opposed to replace them.” His practical advice was to learn to use AI, understand its limitations and identify where it could improve productivity.[1]

Shabani shared the training approach through a Full Stack lecture published May 25, 2023. A later software contribution was Lexy 0.0.3, released July 6, 2024, whose package record credits him as author and lists his GitHub account, shabani1, as maintainer. The project describes itself as data pipelines for AI applications, including retrieval, agent context and structured extraction.[8][9]

Before Replit, Shabani’s work already involved programming and extracting information from large text datasets. His earlier economics research and quantitative-finance work provide context for those skills.[1][4]

Shabani completed an economics BA with highest honors and an MS in applied economics and finance at the University of California, Santa Cruz in 2006. He received a Paul & Daisy Soros Fellowship in 2007 to pursue an economics PhD at Berkeley, where he subsequently earned the degree.[7]

For his dissertation, Shabani recalled recording CNBC for ten hours a day, extracting closed captions into a second-by-second transcript and combining it with high-frequency trading quotes. He tagged company mentions and studied subsequent price and trading-volume responses. He dated this work to 2009 and said it required learning Python and other programming skills beyond the data tools he already knew.[1]

Shabani said he received no academic job offers and declined a BlackRock offer to pursue his own startup. After the startup failed, he returned to BlackRock two and a half years later. His work there included using NLP and machine learning to analyze material such as analyst reports, conference calls and earnings reports.[1]

One research strategy he described used non-negative matrix factorization on descriptions of funded startups to identify clusters of emerging technologies. The team then examined which publicly traded companies discussed those technologies, distinguishing companies that raised a trend themselves from those responding only after investors asked. Shabani said he did not think the strategy was ultimately traded.[1]