Bitcamp 2026 · UMD Hackathon
Co-built a multi-agent hiring platform evaluating candidates across resumes, code, and live interviews — generating evidence-based match scores and skill-gap analysis, delivered through a recruiter-facing interface.
24H BUILD
34+ DEVPOST LIKES
APR 2026
Stack — Gemini 2.5 Flash · LLaMA 3.3 (Groq) · GitHub API · Whisper · Streamlit · MongoDB · bcrypt
Live captures — running the app locally from the repo, employer and candidate portals.
A theory-driven framework using LLM scoring and segmentation to evaluate urban images against Christopher Alexander's 15 properties of wholeness, applying regression analysis to identify interpretable predictors of human aesthetic ratings.
100 IMAGES SCORED
15 PROPERTIES
JAN 2026
Stack — Claude (LLM scoring) · SAM · Regression analysis
Claude mean score vs. human rating — r=0.61 (urban), r=0.69 (landscape), from the actual capstone analysis.
A multi-agent GenAI system handling appointment scheduling, patient triage, and doctor-ready summary generation, with automated email notifications end-to-end.
3 AGENT ROLES
SMTP AUTOMATION
SEP 2025
Stack — watsonx LLMs · LangChain · FastAPI · SMTP
Live capture — the booking assistant running locally, mid-triage.
Personal Project · Revived 2026
A DenseNet121 transfer-learning classifier fine-tuned on the HAM10000 dermoscopic dataset. Built a few years ago, then revived: fixed a checkpoint-path bug that had the deployed app crashing on load, rebuilt the lost front end from old screenshots, and — while tracing the training notebook — found and corrected a real label bug where melanoma images had been mapped to the wrong class name.
~89% VAL ACCURACY
10,015 TRAINING IMAGES
7 LESION CLASSES
Stack — PyTorch · DenseNet121 · Flask · HAM10000
Live captures — the rebuilt site running locally, model wired end-to-end through the Test & Result panel.
Personal Project · Revived 2026
An SVC + TF-IDF classifier for sentiment in Roman Urdu (Urdu written in Latin script), trained on ~20K labeled comments. The trained model files were never committed to the repo — retrained them from the included dataset, fixed a pandas incompatibility and a broken requirements.txt, and rebuilt the front end.
~67% TEST ACCURACY
~20K LABELED COMMENTS
3-CLASS SVC
Stack — scikit-learn · TF-IDF · NLTK · Flask
Live captures — real predictions from the retrained model, positive and negative examples.
Personal Project · Live on AWS
A from-scratch Multinomial Naive Bayes classifier served serverlessly — API Gateway routes to a Lambda that scores YouTube comment sentiment. Didn't need reviving: still live and correct years later.
API GATEWAY → LAMBDA
NAIVE BAYES FROM SCRATCH
STILL LIVE
Stack — AWS Lambda · API Gateway · Naive Bayes (from scratch)
$ curl -X POST .../predict -d '{"comment": "This video is amazing! Loved the editing and pacing."}'
{"predicted_sentiment": "positive"}$ curl -X POST .../predict -d '{"comment": "This was a complete waste of time, terrible content."}'
{"predicted_sentiment": "negative"}
Real requests against the live endpoint, run just now.
LLM-era recommender
Two-stage retrieval and ranking (FAISS → XGBoost / two-tower) with an LLM re-rank and explanation layer — offline evals on NDCG/recall, a simulated A/B harness, and drift monitoring.
Multi-agent system in LangGraph
Specialized agents with tool use, state management, an eval suite, and tracing — built with production trappings, not a demo.
