Yash J.

Institutional‑grade evaluation and trading platform for investment newsletters. Engineer #1. • Architected & built the entire end-to-end platform spanning email ingestion, LLM-based trade strategy extraction & trade simulation. • Designed an event-driven, concurrent system capable of handling peak throughput of 1000+ rps with ~3 seconds P99 latency. • Developed a structured extraction pipeline using LLMs (tool-calling, few-shot prompting, structured outputs) to transform unstructured newsletter text into executable trading strategies stored in a SQL database. • Implemented automated evaluation using LLM-as-a-judge to measure extraction quality, achieving an F1 score of ~0.8 & ~90% correctness on real-world newsletter data. • Built market-data ingestion & monitoring infrastructure using Databento for historical & near real-time pricing to support strategy validation & paper-trading execution.

Social Media Livestream analysis for Enterprise • Created an LLM powered livestream content analyzer using Gemini 2.0 Flash for enterprise usecases such as business strategy, investment & trading signal extraction, threat & security detection, etc. • Implemented caching using Redis that reduced database load by 40% & cut key API endpoint latency to ~500ms. • Developed a content & volume-based statistical anomaly detection system for livestreams using Z-scores. • Engineered high-throughput distributed systems to process 100+ concurrent live video streams. • Implemented semantic search using Google's Text Embedding API and PostgreSQL with pgvector. • Improved discovery & visualisation by overhauling UI/UX, incorporating topic/location filters & interactive layouts.

Engineering & Product Teams Oversight Platform for Management •Engineered Prism, an engineering visibility platform for CTOs & managers with GitHub PR analysis & codebase RAG using Qdrant vector database. •Implemented Linear/Jira workflow automation with Temporal orchestration, built using FastAPI backend & Next.js frontend. •Deployed Prism across Railway & Vercel infrastructure for scalable multi-environment hosting. •Developed Copilot by customizing bolt.diy framework into a specialized PM tool for automated PRD generation/editing & intelligent auto-ticketing workflows. •Implemented Cloudflare edge deployment architecture that supported hundreds of user signups within 3 days of launch.

Alexa Trust & Privacy • Owned & delivered accessibility improvements for the Alexa Privacy Dashboard (APD), improved accessibility by ~75%. • Solely designed & implemented logic to prevent Alexa-enabled Astro devices used in commercial environments from storing business-sensitive audio data. • Identified & eliminated redundant downstream service calls in APD’s data retrieval flow, fetched audio directly from storage & reduced customer-facing latency by ~15%. • Developed a mass-deletion feature enabling users to delete multiple Alexa interaction records at once, implemented frontend & backend changes & collaborated with downstream service teams to extend API support. • Took on on-call responsibilities ahead of schedule & mentored new hires on APD architecture & operational best practices.

• Taught Responsible Data Science (DSUA-202), alongside Prof. Julia Stoyanovich, to a class of 160 undergraduate students. • Assisted in developing the curriculum as well as assignments and lab work. • Held 6/12 lectures in addition to office hours.

• Built a pipeline to process 2 TB of data and generate recommendations based on cosine similarity. • Achieved metrics in the high 80s & 70s: Accuracy@10 of 0.88, Recall@10 of 0.86 & Precision@10 of 0.78.

Machine Learning R&D Lab • Designed & implemented a novel sound-based malfunction detection system for industrial equipment in noisy environments using unsupervised learning techniques. • Performed extensive signal preprocessing, including MFCC extraction & spectrogram generation, to convert raw audio into machine-processable input features. • Achieved an AUC of 0.81 using an LSTM-based autoencoder on a highly imbalanced dataset, reducing false alarms in real-world test scenarios.

Legal & legislation Analysis Platform. • Architected & built the company’s core NLP & ML platform, supporting analysis of legal & legislative documents (bills, acts, decrees, court decisions, constitutions, etc.) • Led early product & technical roadmap definition in close collaboration with the CEO. • Built & managed a team of 3 engineers, establishing development, evaluation, deployment & release processes. • Owned the full ML lifecycle end-to-end: data collection, feature engineering, model training, evaluation & production deployment. • Developed LSTM & GRU-based models for document classification, named entity recognition, summarization of long-form legislative texts, etc. • Implemented custom tokenization, layout-aware preprocessing & sequence batching pipelines to handle scanned & digital documents across multiple formats as well as languages. • Applied transfer learning across related legislative tasks, reducing model training time by ~25% & infrastructure costs by ~15%. • Delivered product demos to non-technical clients & directly supported onboarding.

Built a model using random forest to identify the people most likely to donate (F-1 score of 0.94), facilitating targeted fundraising and a 30% increase in the amount raised compared to the previous year.

Coursework: Intelligent Systems, Big Data, Computer Simulation and Modeling, Advance Database Systems, Software Engineering, Software Project Management, Distributed Systems.