Trisham Patil

About Me

My Journey

I did not start in computer science. I grew up in a practical, non-technical environment, studied mechanical engineering, and learned early how to work through constraints instead of waiting for ideal conditions. My first years were shaped by rigorous math, physical modeling, and systems thinking rather than software tooling. That foundation still defines how I approach AI today: as an engineering discipline that must be measurable, reliable, and useful under real operating conditions.

The turning point came during graduate work, where simulation and computational analysis pushed me into programming through Python and MATLAB. What began as scientific computing expanded into data workflows: ingestion, cleaning, transformation, validation, and feature construction. I moved from solving isolated equations to designing repeatable data pipelines, including time-series processing and schema decisions that could survive changing product requirements instead of breaking every quarter.

From there, I transitioned deliberately into model development and ML engineering. I trained classical ML and deep learning systems, then moved into transformer fine-tuning for NER, classification, and domain-adapted NLP tasks. I built dataset pipelines myself, including scraping, curation, weak-supervision loops, and LLM-assisted labeling to scale annotation quality. As model complexity increased, I started working directly with training trade-offs at the systems layer: sequence length vs batch size, memory limits on NVIDIA T4-class GPUs, fp16 mixed precision, and gradient accumulation to keep training stable and efficient.

Today I work end to end across AI systems, backend infrastructure, and deployment. On the data side, I design ETL/ELT workflows with orchestration patterns used in tools like Airflow and Dagster, and I work with warehouse and lake-oriented stacks such as BigQuery and Snowflake. On the modeling side, I build and evaluate multi-stage pipelines, including multi-task architectures that combine entity extraction and classification in one production flow. On the serving side, I ship these systems through FastAPI and Node.js microservices, queue-backed execution with Celery and RabbitMQ, and cloud deployments on AWS/GCP with Docker and CI/CD.

The systems I care most about are the ones that close the loop from raw data to production decisions: recruiter AI with RAG, multilingual CV parsing, domain-specific NLP for finance and geopolitical signals, and agentic orchestration where retrieval, inference, and API actions must work together. My long-term direction is clear: build full-stack AI systems where data engineering, model training, GPU-aware optimization, and backend reliability are treated as one integrated engineering problem, not separate silos.