Projects.deploy()
The trifecta: Efficiency, Responsibility, and Hard Engineering.
CodeBase — AI-Powered Codebase Intelligence
Semantic search over any GitHub repo — understand any codebase in seconds.
Indexes any GitHub repository into a FAISS vector store — ask natural-language questions and get grounded, code-aware answers. Multi-repo comparison with side-by-side architectural analysis. Embeddings in FAISS, metadata in AWS DynamoDB, raw files in S3. Groq LLaMA-3.3-70B drives inference for sub-second response times.
SkinIQ
A dermatologist in your pocket — from model training to production deployment.
Fine-tuned EfficientNet-B3 on HAM10000 (10k dermoscopy images, 7 skin condition classes) with weighted cross-entropy loss and WeightedRandomSampler to handle 58:1 class imbalance. Deployed a Dockerized FastAPI inference server on Railway with model weights hosted on HuggingFace Hub. Full product with scan history, personalized skincare routines, progress visualization, and routine calendar tracking.
Prompt Optima
Reframing AI performance as a cost problem.
Team project with @Rushil1234 at HackHarvard 2024. My contribution: the compression scheduler that decides what to strip vs. preserve based on a downstream task signal, and the Gemini API integration. Middleware that removes low-entropy tokens before the API call hits a paid endpoint — 60–95% token reduction depending on input redundancy.
Explainable AI Research
Improving algorithmic recourse beyond standard proximity-based CFEs.
Validated the Acceptability & Weighted Proximity (AWP) model to improve algorithmic recourse beyond standard proximity-based Counterfactual Explanations. Engineered evaluation pipelines modeling user-specific acceptability thresholds and aligned feature-change costs with human feasibility constraints.
Codebook Compliance Audit — LLM Behavioral Diagnostics
Do LLMs apply codebook definitions, or just pattern-match on label names?
Built a counterfactual behavioral diagnostic to test whether LLMs genuinely apply codebook definitions when classifying political text, or pattern-match on label names. Novel three-way decomposition attributes each prediction to definition compliance, label-name compliance, or surface-feature matching. Applied to three datasets (CCC, BFRS, Manifestos) across 12 experimental conditions. Key finding: RAG improves accuracy but worsens definition compliance monotonically. Compliance-penalty fine-tuning achieves statistically significant compliance improvement (p<0.001).
Distributed JBOD System
High-concurrency storage built from scratch.
Architected a distributed storage system supporting network-mounted volumes with strict data integrity. Custom socket protocols and mutex synchronization for maximum throughput.
Mosaic-Ward
Privacy-first patient monitoring — no raw video stored.
Real-time patient monitoring system using MediaPipe pose/gesture detection and Gemini multimodal LLM for clinical assessment. WebSocket streaming delivers sub-100ms latency with zero raw video retention — HIPAA-conscious by design. Abstracts patient data to skeletal landmarks before any inference call.
Healthcare Claims Adjudication
Hybrid rules-engine + LLM system for claims processing.
A production-ready healthcare claims adjudication system combining deterministic rules with LLM-based reasoning for ambiguous cases. Features RAG-powered retrieval from historical claims, human-in-the-loop review UI, and comprehensive audit logging with decision explanations.
pytorch/ignite — CharacterErrorRate
Implemented CharacterErrorRate metric for ignite.metrics.nlp — merged in v0.5.2.
Contributed CharacterErrorRate (CER) to the pytorch/ignite metrics library — a standard NLP evaluation metric for ASR and OCR systems. 235 lines of implementation across source + tests, 15 unit tests covering edge cases (empty strings, Unicode, batch aggregation), and full Sphinx documentation matching library standards. Merged as PR #3785 and shipped in the v0.5.2 release of pytorch/ignite (5.1K+ GitHub stars).
sentence-transformers — Docs PR #3856
Documentation contribution to sentence-transformers (18.9K⭐) — under review.
Documentation improvement PR to UKPLab/sentence-transformers — the standard Python library for state-of-the-art sentence, text, and image embeddings. PR #3856 adds clarity for practitioners working with fine-tuning workflows and embedding-based retrieval pipelines.
More on GitHub: SF Urban Analytics · Nittany AI ML Portfolio · AI Secretary Agent · and more
View All Repositories