Doppel: Mentorship & Professional Growth
A passion project I am building to help early-career professionals connect with mentors and peers, work toward their goals, and support one another's growth. Launched in September 2026 as an invite-first community.
TinyGPT: Pretraining, SFT & DPO
Built a training pipeline for a 15M-parameter TinyGPT: pretraining on tiny-shakespeare, supervised fine-tuning for single-sentence responses, and DPO with teacher-generated preference pairs. Implemented speaker-based dataset preparation, response-only SFT loss, and preference-pair construction. Evaluated checkpoints using fixed generation prompts alongside loss, reward margin, and preference accuracy.
Evidence-Driven Deep Research Agent
Research agent with an explicit evidence state, step-level reward signals, and a guided planner that uses those signals to decide what to search next. Runs an adversarial decomposer to generate counter-evidence sub-questions, a multi-stage retrieval pipeline with claim extraction and conflict detection, and a principled stopping criterion. Ablation across two queries: same compute envelope as baseline, full sub-question coverage vs. partial.
LabPilot: AI Copilot for R&D Experiment Optimization
Decision-support loop for R&D labs: surrogate model trained on historical experiment data recommends the next best experiment, with uncertainty quantification and adaptive bandit policies (UCB, LinUCB, greedy). An LLM reasoning layer explains each recommendation; a literature search layer adds source-backed justification. Full session loop: submit observed result, model adapts and recommends next step.
Emergency Guidance Agent: CPR Copilot
Voice-and-video CPR coaching assistant using browser camera and microphone, built on Google Gemini Live. The core design decision: a strict finite state machine controls step transitions across six states (intake → escalation → see_patient → start_compressions → continue_cpr → complete). The model provides real-time guidance within each state; the application enforces the safety sequence. The model cannot advance or skip steps.
Joint Knowledge Graph Embedding, Entity Typing, and Language Modeling
Supervised Contrastive Learning for Causal Event Classification
