 Established an autonomous LangGraph-based assistant that dynamically invoked tools like vector search and NL2SQL

based on user input; achieved over 95% resolution rate for internal queries across critical workflows

 Fine-tuned proprietary LLMs on 50K+ lines of internal application code, enabling contextual code suggestions and

reducing average developer debugging time by approximately 30%

 Introduced a flexible text-to-task pipeline to translate natural language into structured business rules; allowed non-tech

users to create and deploy 300+ logic rules, cutting manual setup time by 90%

 Integrated a robust NL2SQL model into the assistant to support self-serve querying of enterprise databases with 100+

tables; achieved 90%+ accuracy on production-level reporting tasks

 Deployed and maintained a self-hosted Langfuse stack for real-time monitoring of token usage, latency, and prompt

performance, enabling quicker issue diagnosis and iterative model tuning

 Automated nightly regression tests using Langfuse APIs and benchmark queries; ensured assistant stability across updates

and eliminated silent prompt failures in production

 Authored and iteratively refined all production prompts using techniques like chain-of-thought reasoning and role-based

scaffolding, consistently maintaining 94%+ task accuracy

 Integrated the assistant into internal systems via FastAPI services and Dockerized deployments, making it accessible to

over 100 users across engineering, product, and support functions

Built a secure GenAI assistant for internal analysts using LangChain, FAISS, and OpenAI APIs, enabling real-time access

to financial filings and compliance docs; reduced manual search effort by 45%

 Fine-tuned Whisper models on proprietary earnings calls and advisory meetings, improving transcription accuracy to

92%+, including multi-speaker and noisy environments

 Designed and deployed modular services using Flask, containerized via Docker, exposed via REST APIs, and orchestrated

with Kubernetes for scalable delivery of LLM features across internal teams

 Integrated assistant UI with internal tools using TypeScript and lightweight frontends, enabling 24/7 query access to

portfolio data and legal references without disrupting existing workflows

 Created a prompt evaluation and monitoring suite using Langfuse and regression datasets to track latency, accuracy, and

failure patterns; maintained >93% prompt reliability during peak trading hours.

 Containerized modular Python scripts to extract and preprocess large datasets from CSVs, APIs, and SQL databases;

implemented data cleaning, type casting, and null handling routines used across internal tools

 Automated recurring reporting and file transformation tasks using Pandas, NumPy, and scheduling with cron; reduced

manual effort by 80% in monthly operational workflows

 Maintained internal dashboards using Flask and Jinja2, exposing backend data summaries and metrics for business teams

through lightweight web interfaces

 Conceptualized RESTful APIs for internal microservices to enable real-time access to user activity logs, transactional data,

and historical exports; reduced latency and improved inter-service communication

 Wrote Python utilities to parse unstructured log files and JSON payloads, identifying system errors and user behavior

patterns; enabled better monitoring and debugging across environments

 Collaborated with the data team to create reusable data transformation pipelines, implementing file validations, schema

checks, and performance-optimized batch processing using multiprocessing and memory profiling