pa
Resume (PDF, opens in a new tab)

Pratyush Agarwal

I like building things people actually use.
Three years in production data systems, now building software with that same rigor.
Proof, not promises, one commit at a time.

Currently building Mnemo

Resume (PDF, opens in a new tab)See what I'm buildinghi@pratyushagarwal.com (opens email)

Now

A dated feed of what I ship — commits, writes, and milestones — click each item to know more.

223 contributions in the last year

LessMore

  1. Reading someone else's documentation well is its own skill, separate from writing the code. The single most important sentence in the whole integration, that access is granted from the signed webhook and never the redirect, was buried a click deeper than the main guide. That is the story, not the Stripe plumbing itself.

    View Post (opens in a new tab)

  2. Built the same way as everything it lists: a design doc, a dependency-ordered plan, one commit per verified task. The site is not a writeup about the pivot, it is the pivot, running in public.

    View Repo (opens in a new tab)

  3. The real lesson was not any single bug. An AI coding agent will happily produce code that runs, looks fine, and is quietly wrong. Planning first and checking both behavior and structure is what caught a login timing leak, a broken migration, and a wrong uniqueness rule.

    View Post (opens in a new tab)

  4. The CRUD core is done and clean-slate verified, the floor the rest stands on. Auth comes next, then a semantic graph, then an agent layer on top. Nine milestones, each with its own design doc and learning write-up.

    View Repo (opens in a new tab)

  5. Public from commit one changes how you build. Every choice gets a written reason before the code, because the reasoning is part of the repo. The discipline is the point, not a side effect of it.

    View Repo (opens in a new tab)

fraud reduction at Apna within one month
~20%
BigQuery warehouse spend cut at Apna
~$2,500/mo
manual reporting saved at Kohler
~10 hrs/wk
Claude Builder Club Hackathon (UIUC)
2nd place
published (Springer, IEEE, Elsevier)
3 papers

Projects

Selected builds — filter by tag, click to expand.

  • Problem

    Notes, tasks, and links live in disconnected tools that never show how anything interlinks, and none of them actually know you.

    Approach

    Two services over an HTTP contract: a NestJS API for notes, tasks, auth, and kanban, and a FastAPI AI layer, orchestrated with Docker Compose. Strict layering: thin routers, service logic, repository DB access. Nine stacked milestones from plain CRUD to pgvector semantic search, GraphRAG over Neo4j, and an MCP-exposed agent.

    Outcome

    In active development, documented milestone by milestone with design docs, plans, and learning write-ups in the repo.

  • Problem

    Public-company research is fragmented across EDGAR, market data sites, and filing types, and generic AI chat answers without citing filings.

    Approach

    One workspace per ticker: filings, fundamentals, charts, governance, and TickerChat, a Claude-powered assistant grounded in official SEC documents, with PDF export. Next.js frontend on Vercel, FastAPI ingestion service on Render.

    Outcome

    Placed 2nd at the hackathon. Live demo online.

  • Problem

    Raw city-scale transit data is useless without curation, and dashboards alone don't answer ad-hoc questions.

    Approach

    PySpark pipeline over AWS S3 producing curated datasets; XGBoost, Random Forest, and LSTM models for fare and demand forecasting; models served via a FastAPI MCP server and a conversational agent built with Strands Agents on AWS Bedrock.

    Outcome

    End-to-end: raw data to models to a chat interface for real-time mobility insights.

  • Problem

    Repetitive multi-service workflows burn hours and break silently when APIs rate-limit or duplicate work.

    Approach

    Node.js pipeline using the Claude API for evaluation and generation, Google Sheets and Drive for state, Apollo for enrichment. Built around production concerns: per-task model routing, safe re-runs that skip completed work, rate limiting, deduplication, OAuth per provider, and a human approval gate at every stage.

    Outcome

    Powers a real job search end to end, but the design generalizes to any service-orchestration workflow.

Publications

Peer-reviewed papers — click to expand for the abstract.

  • Read paper (opens in a new tab)

    This research introduces a task scheduling algorithm inspired by Artificial Neural Networks (ANN), designed to deliver superior power efficiency and scalability compared to existing methodologies. By leveraging the predictive capabilities of ANN, the algorithm achieves constant-time predictions, even as task loads scale significantly.

    Experimental evaluations using real log files demonstrate performance across metrics such as average start time, finish time, execution time, resource utilization, task failure rates, and power consumption. The model's robustness is proven in both underloaded and overloaded data center conditions, ensuring efficient task scheduling while maintaining system stability under dynamic workloads.

  • Read paper (opens in a new tab)

    This review examines the critical challenges facing fog and edge computing architectures, which play a pivotal role in managing the explosion of sensor-driven data. The study highlights pressing issues such as security vulnerabilities, network latency, and inefficient resource utilization, providing an in-depth analysis of existing solutions.

    The paper explores advanced approaches, including nature-inspired algorithms and metaheuristic optimization techniques, as promising solutions to enhance the performance of fog and cloud computing systems, serving as a guide for researchers and practitioners in next-generation computing environments.

  • Read paper (opens in a new tab)

    This paper investigates the application of the Harmony Search Algorithm (HS) for task scheduling in cloud environments, offering a comparative analysis against Genetic Algorithms (GA) and Moth Search Algorithms (MSA). Simulations conducted using CloudSim and MetaheuristicOpt packages reveal HS as a highly efficient and cost-effective solution for dynamic task scheduling.

    The study evaluates performance across key metrics, demonstrating HS's superiority in optimizing execution time, resource allocation, and operational costs, setting a foundation for further exploration in metaheuristic-based task scheduling methodologies.

Experience

Roles and internships — click to expand for detail.

  • Key responsibilities

    • Automated multiple workflows with Office Scripts (TypeScript), VBA, SharePoint, and Excel macros to replace manual data cleaning and reporting from lab experiments, with early results indicating 15 hours/week in researcher time savings.
    • Prototyped a Python model that predicts extrusion experiment outcomes from input parameters, helping researchers target promising parameter ranges and cutting failed runs; early results point to a 20% reduction in trial-and-error.
    • Engineering an automated data ingestion pipeline that cleans and loads lab experiment data into Power BI, with early results showing a 30% cut in manual data prep time and fewer recurring reporting errors.
    • Maintaining 4 production Power BI dashboards and building 5+ new ones for technical and non-technical stakeholders, giving research leadership a single source of truth for project and portfolio decisions.
    • Validating data quality and consistency across project teams to protect data integrity for reporting and analysis.

  • Key responsibilities

    • Built and shipped production data services integrating 8 source systems, writing PySpark and orchestration code that powered company-wide analytics and AI automation initiatives.
    • Optimized pipeline performance through partitioning and incremental loads, cutting data ingestion time by 40% and query execution time by 25% on high-volume tables.
    • Designed ETL pipelines with Azure Data Factory and Databricks, ingesting SharePoint and on-prem SQL Server data into Parquet-based raw and curated layers for downstream systems.
    • Integrated on-prem operational databases with Azure cloud platforms, building the data flow behind company-wide analytics and upcoming AI automation work.
    • Wrote reusable, parameterized Python and PySpark transformation code so new data sources onboard through configuration rather than rewrites, reducing duplicated logic and pipeline errors.
    • Brought Git-based version control and code review to the team's data workflows, improving governance and release consistency for pipeline code.
    • Standardized relational data models from on-prem SQL Server schemas, authoring ER diagrams and a data dictionary so future schema changes wouldn't break existing downstream analytics.
    • Built and maintained 6 Power BI dashboards on modeled data, tracking manufacturing KPIs and saving plant leadership 10 hours of manual reporting each week.
    • Analyzed Ignition form submissions across plants, identifying schema inconsistencies and shipping structural fixes to protect long-term data model integrity.

  • Key responsibilities

    • Led cross-functional coordination for a quantum satellite-themed game, aligning sponsors, programmers, design, and art to ship the project in 2 months and around $2,000 under budget.
    • Prioritized the development backlog by weighing core gameplay value, implementation complexity, and system dependencies, so critical features shipped first and bottlenecks cleared fast.
    • Ran an Agile workflow tracking milestones, tasks, bugs, and dependencies, keeping the team on schedule through resource constraints and shifting timelines.
    • Managed maintenance and optimization of a game with 20,000+ downloads, resolving bugs, improving stability, and supporting adaptation for a broader global audience.
    • Worked closely with senior production leadership to validate planning decisions, refine execution, and keep stakeholders aligned throughout development.

  • Key responsibilities

    • Built a user verification and data ingestion service using SQL, Python, Postman, and Government of India APIs, cutting processing time for Operations agents by 50% and improving onboarding for 1,200+ users weekly.
    • Saved $2,500/month in BigQuery warehouse spend and cut query execution time by 20% through partition-filtering and query scope-limiting on high-cost workloads.
    • Partnered with Trust and Safety leadership and operations managers to consolidate fragmented reactive and proactive investigation workstreams into a single unified system, reducing investigation overlap by 35% and saving 20+ agent hours weekly.
    • Prototyped a query-based job pause system for recruiters flagged with fraud complaints, validated it internally, and handed it to engineering for a platform-wide rollout that improved the team's north star metric by 5%.
    • Built a proactive fraud screening workflow in Retool with engineers, flagging new recruiters whose documents (PAN, GST, Aadhaar, DL) or devices matched previously blocked accounts before they could post jobs.
    • Migrated all Trust and Safety queries and dashboards across 3 database moves (PostgreSQL to BigQuery, partitioned BigQuery, then Presto), keeping agent workflows running through each switch.
    • Built a reusable library of parameterized SQL query functions, letting ops managers look up one or many recruiters on demand and speeding up ad-hoc requests.
    • Debugged pipeline issues with the data engineering team, including dropped rows during ingestion and a mis-mapped event tag, then backfilled the data to restore accurate metrics for leadership.
    • Integrated third-party government identity APIs and designed user verification workflows, reducing sign-up drop-offs by 10% and improving onboarding conversion.
    • Built a SQL-based fraud detection workflow in Metabase and Retool that cut fraudulent recruiter activity by 20% within a month, strengthening platform security.
    • Rebuilt the team's core metrics with 15 new queries and backfilled 6 months of data, showing which workstreams drove results and where agent effort went.

Certifications

Credentials I've earned, and one I'm still working toward.

  • Courses

    • Foundation: Introduction to LangChain - Python

      Completed · Verify (opens in a new tab)

    • Foundation: Introduction to Deep Agents

      In Progress

    • Foundation: Building Reliable Agents

      Not Started

    • Foundation: Monitoring Production Agents

      Not Started

    • Foundation: Introduction to LangSmith Deployment

      Not Started

Writing

Notes on building in public — Substack and LinkedIn.

Follow along on Substack (opens in a new tab)

About

Pratyush Agarwal

I spent three years working with data across Kohler, Apna, and Google, building pipelines, automation, and systems that teams relied on every day. Somewhere along the way I realized the parts I loved most were the software parts: designing APIs, shipping products, watching people use what I built. That's the work I want to keep doing.

I'm early in my software engineering career, but I have real production experience, I learn fast, and I'm building the proof in the open: every project public, every milestone logged, every lesson written up. If you want to see how I work, it's all here.

Education

  • University of Illinois Urbana-Champaign

    2024–2026

    M.S. in Information Management

  • Manipal University Jaipur

    2018–2022

    B.Tech in Computer and Communication Engineering