CV

Contact Information

Name Adi Banerjee
Professional Title Software Engineer
Email abanerjee.email@gmail.com
Location Seattle, WA
Website https://adi-ua.github.io

Professional Summary

Software engineer at Amazon in Seattle building agentic systems for vendor operations: an MCP server backed by DuckDB over S3 Parquet, a graph-based documentation-drift detector on Bedrock, and a real-time sourcing-explainability agent over a 30M-document OpenSearch index.

Experience

  • 2025 -

    Software Development Engineer 2
    Amazon
    • Built agentic capabilities for dropship vendors to diagnose delivery performance drops and warehouse suspensions through Amazon’s vendor-facing AI assistant: an MCP server and skills backed by DuckDB over S3 Parquet, built spec-first with OpenSpec, evaluated with Langfuse traces and LLM judges, and authorized per user by task-based role.
    • Proposed and built a documentation-drift detector that models a wiki corpus as a Neptune graph and uses a Rust Lambda on Bedrock to find stale pages and generate reviewable edits. A topic-based diff algorithm feeds the model 200 tokens of graph topics per query instead of 50K tokens of page content, analyzing 30+ pages in under 5s from a 48ms cold start.
    • Designed the ingestion layer and co-built a sourcing-explainability agent that streams buying-decision changes from 3 upstream systems through Kinesis and Lambda into OpenSearch (30M documents): a 100ms lookup replaces multi-hop API resolution, with a 7.5s p95 time to first token against a 20s requirement and 85s end-to-end answers from a 198s p95 baseline.
    • Co-designed and led launch of embedded QuickSight dashboards in Vendor Central across 5 marketplaces (React, Java, CDK), retiring a 7-service legacy analytics stack so analysts publish new metrics without engineering work.
    • Amazon SCOT hackathons: Spark 2025 winner (multi-agent tech-debt profiler on LangGraph, 1000+ teams), OSS AI 2026 runner-up for Technical Excellence (the drift detector, 18,800+ lines across 3 packages shipped to beta in 2.5 days), Spark 2026 Evaluation & Verification Award.
  • 2024 - 2025

    Software Development Engineer
    Amazon
    • Led a suspension system overhaul with EventBridge batch processing and DynamoDB distributed rate limiting: execution from 2 days to 5 min, throughput from 1 to 200 requests/min.
    • Tech lead across 3 teams on a network optimization workflow affecting 230MM+ configs.
    • Migrated 26 vendor email use cases across 12 services off a 10-year-old Tier-1 notification service to SES and an internal platform, using 11 service-level feature flags on a graduated rollout, with a 0% error rate since launch.
  • 2023 - 2024

    Research Engineer
    University of Arizona
    • Built deep RL transformer models for math OCR with a 6% gain over SOTA, and rewrote the MathML parser in Rust (Nom).

Education

Publications

  • 2021
    Modular Procedural Generation for Voxel Maps
    AAAI Fall Symposium

    Pyarelal, A., Banerjee, A., Barnard, K. (2022). Modular Procedural Generation for Voxel Maps. In: Gurney, N., Sukthankar, G. (eds) Computational Theory of Mind for Human Machine Teams. AAAIFSS 2021. Lecture Notes in Computer Science, vol 13775.

Skills

Languages: Java, Python, Rust, Go, TypeScript, C++, SQL
AI & Agents: MCP, Bedrock, LangGraph, Strands, Langfuse, LLM-as-judge evals, Neptune, OpenSpec, Kiro, Claude Code, PyTorch
Infrastructure: AWS CDK, Lambda, DynamoDB, OpenSearch, Kinesis, EventBridge, S3, DuckDB, QuickSight, Spring Boot, FastAPI, React, Docker, Kubernetes

Projects

  • Whisper

    Shared daily safe-word service with Google OAuth, timezone-aware rotation, JNI-based Rust word generation, Bucket4j rate limiting, Prometheus metrics, and a Helm chart for Kubernetes deployment.

    • Java 21
    • Rust (JNI)
    • Spring Boot
    • React
    • Docker
    • Helm
  • dups

    Concurrent duplicate file finder with a two-pass hashing strategy (4KB partial hash for files over 1MB, full SHA-256 for confirmed candidates), channel-based worker pool, interactive per-group deletion, and cross-platform builds via goreleaser.

    • Go
    • Goroutines
    • SHA-256
  • Hindsight

    Keyless backtest visualizer for equity trading strategies with real-time portfolio charting, trade annotation, and drawdown analysis. Deployed on Vercel with GitHub Actions CI.

    • Python
    • FastAPI
    • React
    • Vercel
  • TaylorGPT

    Designed a 1.4M parameter character-level transformer decoder model to generate Taylor Swift lyrics from scratch using PyTorch. Built custom masking, multi-head attention, transformer blocks, and enabled multi-GPU training.

    • PyTorch, Lightning, Numpy
    • Custom transformer implementation
    • Multi-GPU training
  • Start Journey

    Designed a full-stack stable diffusion-based image generator with PyTorch, CUDA, HuggingFace, and FastAPI for the backend and React in TypeScript for the frontend. Enabled scalable image generation and user-friendly web interface.

    • Python, PyTorch, HuggingFace, CUDA, FastAPI, React
    • Full-stack ML deployment