Projects.

Selected work,and side projects.

Each card is an architecture blueprint of what I built.

Selected Work

6 entries
  1. Jan 2024 – Present/Fractal.ai/Senior Software Engineer

    Batch GenAI framework at enterprise scale

    Architected a batch GenAI framework that processes 50K+ daily call and chat transcripts, extracting churn drivers and risk signals via agentic workflows and surfacing results in a generative chat UI for natural-language search and triage. Cut batch runtime from 6h to 2h and trimmed LLM cost materially through prompt compression, tiered routing, and aggressive caching.

    • Python
    • LangGraph
    • AWS Bedrock
    • Postgres
    • Redis
    • Step Functions
    Fractal.ai
  2. 2024 – Present/Fractal.ai/Senior Software Engineer

    Enterprise RAG platform over private knowledge bases

    Built a production RAG platform over a large corpus of private enterprise docs: hybrid retrieval, reranking, and guarded generation behind a chat surface used by internal teams for research and triage. Owned the evaluation harness — retrieval metrics, groundedness checks, and regression suites — so new models and chunking strategies could ship without quietly regressing answer quality.

    • Python
    • Bedrock
    • OpenSearch
    • pgvector
    • LangGraph
    • FastAPI
    Fractal.ai
  3. 2022 – 2023/GlobalMonitor/Senior Software Engineer

    Conversational analytics + a headless CMS

    Engineered a unified analytics platform that consolidated research and analytics data, and shipped an LLM-powered text-to-SQL tool so non-technical users could query large datasets in natural language. Led a team of 4 on a scalable headless CMS with real-time updates and complex data workflows, which supported the publication of thousands of research artifacts.

    • React
    • TypeScript
    • Next.js
    • Node.js
    • GraphQL
    • LLM / text-to-SQL
  4. 2021 – 2022/Array/Frontend Engineer

    Sub-200ms interactions on a data-heavy app

    Profiled and rebuilt the hottest surfaces of a credit-focused dashboard: virtualized large tables, reworked data fetching to RSC-style streams, and rewrote the chart layer to avoid layout thrash. Shipped consistent sub-200ms interactions on long lists and brought Core Web Vitals into the green across the board.

    • React
    • Next.js
    • React Query
    • D3
    • Sentry
    Array
  5. 2019 – 2021/Amazon/Software Engineer

    Computer-vision inspection at fulfillment scale

    Built ML infrastructure for a product-inspection system running on millions of packages a day: training pipelines, model serving, and the deployment harness that gated new CV models into production. Owned the eval loop and on-call rotation for the serving layer.

    • Python
    • PyTorch
    • AWS SageMaker
    • Kinesis
    • DynamoDB
    Amazon Science
  6. 2018 – 2019/Decision Theater · ASU/Research Software Engineer

    Real-time sentiment analysis on streaming social data

    Designed a streaming sentiment pipeline for policy researchers: Kafka topics, Spark Structured Streaming jobs, and a visualization surface that let analysts slice volume and sentiment by topic in near real time. Shipped the end-to-end system and the dashboards used in briefings.

    • Kafka
    • Spark
    • Scala
    • Elasticsearch
    • D3
    Decision Theater · ASU

Side Projects

GitHub

  • sowmith.dev

    active

    This site. Next.js App Router, MDX blog, rough-notation hero, and a minimal design system I iterate on in the open.

    • Next.js
    • TypeScript
    • Tailwind
    • MDX
    Live

  • design.md

    active

    A format spec for describing a visual identity to coding agents — a DESIGN.md that an LLM can actually act on.

    • Markdown
    • Spec
    • AI Agents
    Code

  • pubsub-prototype

    active

    Minimal connection-state pub/sub in Python + PySide6 (Qt) — subscribe / unsubscribe, event delivery, and on-demand status polling.

    • Python
    • PySide6
    • Qt
    Code

  • meridian

    active

    Governed AI chat with agentic data pipelines underneath. Built for the OpenAI Codex Hackathon.

    • TypeScript
    • OpenAI
    • Next.js
    Code

  • notes mcp

    active

    Private MCP server for saving dev notes and daily logs straight from my editor.

    • TypeScript
    • MCP

  • obsidian vault agent

    active

    Agent that treats my Obsidian vault as a second brain — query, summarize, and cross-link across years of notes.

    • TypeScript
    • Obsidian
    • LLM

  • open-jobs finder

    wip

    Pipelines listings through Serper search and a small LLM pass to surface real open roles from the noise.

    • Python
    • Serper
    • LLM

  • n8n automations

    active

    A small set of n8n workflows that glue APIs, AI calls, and personal tools together without writing glue code.

    • n8n
    • LLM
    • Webhooks
Read the blog