Work

• Building automations for the growth office's reporting workflows to cut manual work and help the team extract more value from existing reports.
• Contributing to the firm's in-house ERP build, focused on finance data migration.

• Leading a full redesign and feature buildout of the PQRS Research website in Next.js, including a dual onboarding system with distinct registration and intake flows for mentors and mentees.

• Contributed to dataset design for construct validity (outdoor imagery, no visible labels or standard objects, scale-invariant object range) and served as primary collector for the 459-image benchmark, personally performing all instrument measurements and building a 3-tier blur-degradation pipeline (OpenCV).
• Ran the initial evaluation experiments across the VLM suite (Gemini 3.1 Pro, Gemini-Robotics-ER, GPT-5.4), then contributed to experimental design on batch structure and run repetition as the evaluation scaled to the remaining models.
• Sole author of the model error analysis: reviewed the 10 worst-performing responses per category per batch across GPT-5.4, Gemini variants, and Gemini-Robotics-ER, identifying a misidentification-vs-miscalibration failure dichotomy in how the models get physical estimation wrong.
• Findings directly shaped the prompt-templating intervention: comparing open-weight models' (Gemma 4, Qwen 3.5) worst responses against Gemini 3.1 Pro's reasoning to write structured system prompts that measurably improved Gemma 4 on angle and stability and Qwen 3.5 on angle. Also helped shape a tool-augmentation intervention supplying segmentation and depth context to aid spatial reasoning.

Conversational VR agents leave users waiting through speech recognition, LLM inference and speech synthesis. I built a Unity agent on Gemini 2.5 Flash with Google Cloud STT/TTS and used it as the testbed for a within-subjects study (N=18) on how the agent should behave while the user waits.

• Engineered a real-time VR conversational agent integrating Gemini 2.5 Flash with Google Cloud STT/TTS, averaging ~2.0s round-trip response latency through parallel API calls, caching, and audio preprocessing.
• Designed a within-subjects study (N=18) comparing three latency feedback mechanisms (embodied gestures, visual indicators, verbal fillers) against a no-feedback control, measured with the Godspeed questionnaire (animacy, anthropomorphism, likeability) and custom distraction and wait-perception items.
• Found that embodied gestures made the agent significantly more lifelike and human-like than visual loading indicators (paired tests, Holm-corrected, p<0.05) and significantly less distracting (Wilcoxon signed-rank, p<0.01).

Computational social science study (with Ananya Singla) of how people discuss pre-eclampsia on Reddit from 2013 to 2025 - from 'mom groups' to medical triage - using sentiment analysis, temporal trends, a pre- vs post-COVID-19 comparison, and topic modelling.

• Collected posts and comments from about 20 pregnancy and health subreddits with PRAW, weighting each subreddit by relevance and scoring records against keyword lists suggested independently by three LLMs.
• Built a cleaning and scoring pipeline (deduplication, contraction and emoji expansion, VADER sentiment, COVID-period tagging) with local JSON or Firebase storage and a one-command CLI that regenerates every figure and an HTML report.
• Analysed yearly sentiment and volume trends, pre- vs post-COVID differences (t-test), and themes via TF-IDF and 5-topic LDA.

• Architected a FastAPI backend with AST-based parsing for 15+ languages, enabling automated documentation generation across polyglot codebases at 50,000+ LOC/day.
• Cut AI embedding costs by 85% by building a hash-based system that fingerprints each function's AST and skips re-embedding unchanged code, eliminating redundant Azure OpenAI API calls.
• Built a production RAG pipeline using LangChain and Azure OpenAI, backed by Cosmos DB with IVF vector indexing for semantic code retrieval, serving 17+ enterprise repositories; migrated from FAISS to support persistent, production-scale storage.
• Shipped 12 production REST API endpoints with typed Pydantic schemas, powering real-time documentation preview in the client-facing web interface.

Designed and shipped a full-stack site for a CA firm with a careers page and a private admin dashboard where firm partners review and manage candidate applications.

• Delivered a full-stack site with a public careers page and a Firebase-backed admin dashboard, giving firm partners a private portal to review and manage job applications.
• Secured admin access via Google OAuth with an email allowlist (Firebase Auth), enforcing partner-only access without a separate auth backend.
• Built a GitHub-editable content layer so the client can update site copy post-handoff without developer involvement.

A command-line tool to search local files by meaning rather than keywords: documents become Sentence Transformers embeddings in a FAISS index, queried in natural language.

• Pluggable embedding models with a separate index per model, so models can be compared side by side; uses FAISS GPU when available and always saves a portable CPU index.
• Handles text, source code and .docx locally; PDFs and images are optionally converted to text with the Gemini API, and only those file types are ever uploaded.
• Tagging with suggestions from YAKE keyword extraction plus semantic similarity to existing tags, and rule-based 'smart folders' that combine directories, file types, tags, keywords and a semantic query.

Started as a course project (Web Development using AI, 2025): a resume tool that tailors your resume to a job and generates interview prep from it. Rebuilt in 2026 as Folio, one app where the resume you build is the same document every interview tool practises against: mock interviews with a scored debrief, job-description matching, STAR questions and server-graded DSA practice.

• Built a full-stack AI resume tool with real-time PDF preview and ATS-optimized generation using Next.js SSR and the Gemini 2.5 Flash streaming API.
• Integrated Gemini to generate personalized interview questions by cross-referencing uploaded resumes against a structured DSA/STAR prep database, surfacing role-specific preparation gaps.
• Implemented Firebase Authentication with Google OAuth for secure user management and session-based access control across the application.

• Designed a modular LLM task-planning system for a food-preparation robotic arm, decoupling high-level intent logic from low-level actuation to enable parallel development across firmware and AI teams.

Course project (with Jia Bhargava and Tanay Srinivasa) segmenting kelp canopy in Landsat tiles from DrivenData's Kelp Wanted competition. A three-stage pipeline where each stage passes on only what it can't settle: a small CNN drops tiles corrupted by Landsat 7 scan-line striping, a hard-voting ensemble of ResNet-18/34/50 and Inception-v3 decides whether a tile contains kelp at all, and a U-Net with an EfficientNet-B3 encoder segments the rest on remote-sensing indices (MNDWI) instead of raw bands.

• Achieved a Top 100 finish (top 17%) in a DrivenData satellite imagery competition using a U-Net with EfficientNet-B3 backbone for kelp forest segmentation in noisy real-world imagery.
• Built a two-stage filter before segmentation: a CNN streak detector to drop corrupted satellite tiles (F1 0.952), then a ResNet/Inception-v3 voting ensemble to skip tiles without kelp.

• Developed a preprocessing pipeline (contrast adjustment, noise reduction, edge detection) to improve zero-shot Segment Anything Model (SAM) segmentation on agricultural crop-weed imagery, without modifying the underlying model.
• Improved batch processing throughput for 10,000+ image datasets by implementing vectorized NumPy operations and Python multiprocessing.