Projects

Research projects

• Persona-Induced Response Modulation

  Oct – Dec 2025

  An activation-steering pipeline that induces confident personas without fine-tuning, reducing expressed uncertainty by 95%. Ablated on Llama 3.1 8B and Gemma 2 2B.

  • Activation steering
  • Interpretability
  • LLMs

• Arabic TTS Fine-Tuning

  Sep – Dec 2025

  Supervised fine-tuning of NeuTTS-Air on a proprietary Arabic speech dataset, reaching 30% WER. Done in Prof. Bhiksha Raj's MLSP lab at CMU.

  • Speech
  • Text-to-speech
  • Fine-tuning

Agents & retrieval

• FoodieSpot restaurant-booking agent

  2025

  A Llama 3.1 8B agent with tool use that recommends restaurants through retrieval and books tables in a reservations database.

  • LLM agents
  • Tool use
  • RAG
  Code

• RAG with citations for SEC 10-K filings

  2024

  Retrieval-augmented question answering over financial filings that cites its sources, built with FAISS, ChromaDB, and multiple LLM backends.

  • RAG
  • Citations
  • Finance
  Code

Hackathons

Four first-place finishes in 2023.

• Hackmitten · MIT Thandavapura

  1st place · 2023

  A Portuguese pronunciation scorer using acoustic-waveform similarity and dynamic time warping, with no speech-to-text.

• Prayog · VVCE Mysuru

  1st place · 2023

  A real-time yoga pose estimator with joint-angle feedback and a ChatGPT assistant for corrections and guidance.

  Code

• SwarmIT · HPE & GSSIETW

  1st place · 2023

  A hospital scheduler pairing CNN skin-disease diagnosis with traffic-aware allocation for patient routing.

  Code

• Hackfest · T-Hub & ICE Hyderabad

  1st place · 2023

  A developer-tool suite: code-to-flowchart conversion, complexity metrics, and malicious-commit detection.

  Code