Harsh Gurawaliya

Built an agentic VC investment analyst that turns a one-line company description into an evidence-cited, source-traceable investment memo through a structured seven-stage dialectic: decompose, multi-source evidence, pro/con generation, adversarial critique, refine, quantitative scoring, and recommendation.

Engineered against LLM overconfidence with a mandatory devil's-advocate critique that stress-tests every argument, then rolls signals into a 0-100 composite score with an explicit confidence level.

Designed pluggable, declarative strategy profiles for dimensions, signal weights, rubrics, thresholds, connectors, and guardrails, allowing the same deal to be scored against multiple fund theses using signals from news, GitHub, hiring, web traffic, and funding sources.

Founded and built a patient-data platform that records doctor-patient consultations, transcribes them in real time with Whisper, and translates Tamil and Hindi into English.

Generated LLM-driven medical summaries from reports, transcripts, and lab data, with trend detection, similar-patient diagnosis assistance, role-based authentication, doctor review, follow-up scheduling, and a patient waiting-list queue.

Validated the product with six senior doctors in India, receiving positive testimonials on its potential to improve hospital workflows and patient outcomes.

Fine-tuned LLaMA 3 with QLoRA on a large patient-doctor conversation dataset for medical Q&A and triage scenarios.

Merged adapters into the base model, quantized it to GGUF, and deployed it locally in LM Studio for secure offline domain-specific conversational AI.

Trained a UNet with dropout and skip connections on 2,146 brain MRI scans using a custom Dice and Focal loss strategy to handle class imbalance and improve boundary precision.

Achieved 98% global accuracy and 98% mean IoU on the held-out test set.

Trained a PPO agent to land an orbital-class rocket in simulation, shaping rewards for stable touchdowns and tuning the policy network for consistent autonomous landings.

Built a complete MLOps loop including data ingestion, preprocessing, Scikit-learn training, a Flask and React app, Docker, GitHub Actions CI/CD, and AWS deployment for automated, versioned inference.