working on integrating LLMs

Engineering across the board:

* Helped develop custom OCR extraction model to extract data from hundreds of financial and rent-roll documents in the commercial real estate industry

* Engineered ETL pipelines to collect, process, and analyze datasets from rent-roll documents and financial statements; streamlined data flow and ensured data quality

* Implemented a custom AI editor, similar to Notion, but exclusively for commercial real estate investors; responsible for full-stack implementation (front-end, backend, AI) to enable intelligent features such as (auto-completion, improving writing, “Ask AI”, etc.)

* Prototyped and experimented with new AI technologies and features, mostly with RAG techniques such as agents, reranking & filtering, fusion retrieval, for document search/retrieval

* Cross-collaborated with various engineers and designers, conducted extensive user research + analysis, helped put out multiple “fires” across product, led design decision related to website/product, and everything else that comes with an early-stage startup

I oversaw the design and execution of LLMs on Pelican's platform. I built the quiz product feature from scratch to launch, leveraging RAG to optimize user matches with 529 education investment funds. I managed both front-end and back-end components of the crowdfunding platform, including implementation of scaleable back-end/database system, JSON-based APIs, etc.

I experimented with a range of both open-source and closed-source Language Model (LLM) solutions, aiming to optimize the code review workflow. These models include GPT-3.5/4, Falcon, Llama, Mosaic's MPT, etc. I developed a series of features aimed at discerning subtle low-context code modifications. This encompassed tasks such as bug detection, formatting discrepancies, and enhancing overall code elegance and efficiency. Overall, I learned about the current capabilities/limitations of open-source and closed-source LLMs (unique strengths/weaknesses, output quality, reliability, etc).

Currently, electrode-based retinal implants are severely limited in the restoration of vision loss. These implants are highly-invasive, limited in resolution, and degrade in utility overtime. My lab is developing a novel, minimally-invasive retinal prosthesis to overcome such limitations and stimulate retinal neurons to restore vision in blindness. This system uses gold nanorods (AuNRs) and near-infrared (NIR) light to activate retinal neurons by causing temperature changes in the neuronal membranes. The lab is using a custom experimental setup to validate the photothermal activation of retinal neurons ex-vivo. An important aim in the ex-vivo validation is to determine the number of RGCs activated per unit time by analyzing the images taken by the camera. The imaging system captures ~1000 retinal neurons, and it is very challenging, if not impossible, to manually identify individual cells and obtain their time traces, especially when we have tens of retinal samples and data to analyze.

My work focused on implementing various image segmentation algorithms that automatically identify individual cells from a grayscale fluorescence image of a retinal explant and comparing their performance.