Pratyush Singh

I ship products where being wrong is expensive.

Trading desks at Thomson Reuters. Carrier networks at Nokia. Ultrasound through FDA review at Philips. A 450,000-unit-a-day fulfillment center at Amazon. And AI agents I build that aren't allowed near the money. MBA at USC Marshall, graduating May 2027 and looking for full-time product roles.

What are you hiring for? The page reorders to match.

Five years shipping at the seam of strategy and engineering, then an MBA. I read the codebase, write the PRD, and build the prototype myself when that is the fastest way to learn.

I design AI features around one question: what is the model allowed to decide? Everything I build ships with evals, a guardrail table and a launch gate, and I've put an LLM classifier into production at Reuters.

I find the problem the dashboard can't show you. At Reuters that meant sitting with traders until we saw why nobody used our product, then fixing it.

Computer science degree, three years of hardware-in-the-loop testing, and I write production-grade prototypes. Engineers don't have to translate for me.

I turn messy operations into a metric people trust, then use it to decide where the money goes. At Amazon, one normalized metric reordered a whole repair budget.

I run programs across hardware, software and partners where a slip has a customer attached: carrier rollouts, an FDA submission, and AI launches gated on evidence rather than dates.

Sending this to someone? Share the standalone version: AI product manager portfolioproduct manager portfoliotechnical PM portfolioproduct operations portfoliotechnical program manager portfolio.

  • 18% to 84%daily usage of a Reuters trader dashboard, after watching traders use it in London and New York
  • 94%accuracy on a GPT-4o tagging system I shipped, replacing 80% of manual tagging across 50,000 items a day
  • −40%customer late shipments at an Amazon fulfillment center, from a funding case I built
  • 94%accuracy on a GPT-4o few-shot classifier in production at Reuters, 50,000 items a day
  • 0wrong or unauthorized refunds across 19 eval tickets in Support Desk, my multi-agent build
  • 30 of 30questions answered correctly every run by Sonnet, against 29 of 30 for Haiku at 1/20th the cost
  • 18% to 84%daily usage after I found the real adoption barrier by sitting with traders
  • −40%scan-to-diagnosis time at Philips, from 12 usability gaps found by observing 20+ clinicians
  • 3 launchesowned end to end at Reuters with 24 engineers, widening coverage 30%
  • 150+technical requirements per release cycle across eight hardware models at Nokia
  • 8 to 45 msp95 latency spike on a streaming API that I helped trace and restore within 72 hours
  • 171automated tests behind StrideAI, a FastAPI service with CI and regression gates that I built
  • 3.4×fault-rate gap between two "identical" Amazon departments that raw counts had hidden
  • 4 h to 15 minper dataset after I automated HL7/DICOM validation at Philips; three other teams adopted it
  • ~200 hof editorial time saved every month by the Reuters tagging system
  • 99.9%first-time success across 20+ carrier rollouts at Nokia, worth $4.5M in partner sales
  • On timeFDA submission for the Philips ultrasound platform
  • 5 gatesinternal, beta, 20%, 40%, 100%: the evidence-gated rollout I designed for StrideAI

Selected work

Three independent builds. Each has working code, the documents a team would need, and the numbers from its own test runs.

Support Desk

Agents that resolve support tickets end to end, refunds included. The model can only propose an action; plain code decides whether it happens.

For AI PM: safety metrics are release gates. Any wrong action, false promise or leaked card number above zero blocks a deploy.

For product: auto-resolution is a target band (50 to 65%), not a number to maximize, because pushing it higher means escalating less.

For technical PM: idempotent executor, defense in depth (the executor re-checks limits itself), bounded revision loop, per-step traces.

For ops: an approval queue for refunds over the limit, an audit log, and CSV export. The limit starts at $50 and rises after a clean week.

For programs: a launch plan that starts in shadow mode, comparing every decision with a human's before anything is sent.

  • Multi-agent
  • Evals
  • Release gates
  • Policy as code
  • Live demo
Read the case studyTry the demo
Release gates (offline evals)
Eval tickets passing19 / 19
Wrong or unauthorized actions0
Promises that didn't happen0
Must-escalate reaching a person100%
Resolved without a person63%

Offline runs use a stand-in model with injected faults. They test the safety net, not live-model quality.

StrideAI

An adaptive running coach, run as a full program: from charter and RAID log to a tested FastAPI service and an iOS companion. Rules make the call; the LLM only explains it.

For TPM: a RAID log with triggers and owners, five ADRs, an executive status that asks for one decision, and a rollout where percentage follows evidence, not the calendar.

For AI PM: real-world validation changed the hypothesis. Matching the athlete's morning decision turned out to be the wrong success metric.

For product: an MVP scope with explicit trade-offs, and a product change (tracking accumulated recovery debt) driven by what validation showed.

For technical PM: wearable ingestion, SQLite persistence, Docker, CI, a groundedness checker and a fallback when the LLM fails, covered by 171 tests.

For ops: acceptance, override and outcome-coverage metrics feed a regression gate and a small operational dashboard.

  • RAID
  • ADRs
  • Rollout gates
  • FastAPI
  • iOS HealthKit
Read the case study
Program at a glance
Automated tests171
Architecture decision records5
Tracked risks, with triggers7
Critical-path dependencies5
Baseline safety violations0

Second Avenue

A data product that warns a consignment retailer when contracts are falling behind target, while there's still time to act. I pre-registered the test, and it said my first bet didn't hold.

For product ops: the flag was judged on Ops cost as well as accuracy. About 23 flags a week at 10 minutes each was three hours of review time, mostly false alarms.

For AI PM: a plain-language question layer with six governed tools and no free SQL, evaluated on 30 questions across two models with cost per run.

For product: discovery, sizing, a PRD, then a prototype that tested the riskiest assumption before asking engineering for anything.

For technical PM: BigQuery views, a seeded data generator, data-quality checks and a backlog synced to GitHub Issues.

For programs: six epics traced to the revised PRD and laid out as Now, Next, Later, with the backlog synced to GitHub Issues.

  • Experiment design
  • SQL
  • LLM eval with cost
  • PRD
  • Roadmap
Read the case study
What the pre-registered test showed
False alarms, first bet80%
False alarms, after the fix36%
Warning before contract ends30 days, vs 0 today
Question layer, Sonnet30/30 at $1.08
Question layer, Haiku29/30 at $0.05

Fictional company, simulated data of 1,968 contracts.

Experience

Most recent first.

Lines that matter most for the role you picked are highlighted.

  1. AmazonJun to Aug 2026
    Etna, Ohio

    Pathways Senior Operations Manager Intern, Worldwide Operations

    • Introduced the site's first volume-normalized jam metric across 12,900 jams and 2.9M packages. It exposed a 3.4× fault-rate gap between two identical departments that raw counts had hidden, and now anchors the site's mechanical review forum.
    • Cut customer late shipments 40% at a 450,000-unit-a-day fulfillment center by building the funding case that ranked conveyor repairs first, then presented the customer-impact readout to senior leadership.

  2. Thomson ReutersNov 2023 to Jul 2025
    Bangalore, India

    Lead Product Quality Engineer (product lead), Reuters Technology

    • Shipped a GPT-4o few-shot classification system that replaced 80% of manual editorial tagging across 50,000 daily items at 94% accuracy, saving about 200 hours a month. Won leadership adoption with an ROI business case.
    • Directed product vision, roadmap and technical requirements for three launches with 24 engineers, widening European and Asian coverage 30% and contributing to 12% market-share growth.
    • Found an adoption barrier the usage data couldn't show by sitting with traders in London and New York: the dashboard covered their Bloomberg terminals. Daily usage went from 18% to 84%.
    • When a streaming API's p95 latency jumped from 8 ms to 45 ms, worked with the platform team to isolate the bottleneck and restored the SLA for 12 high-frequency trading clients within 72 hours.

  3. NokiaSep 2020 to Oct 2023
    Bangalore, India

    Product Quality Engineer, hardware-software integration (Nuage Networks SD-WAN)

    • Shaped three roadmap priorities by benchmarking Nuage VSP against Cisco Viptela, VMware VeloCloud and Fortinet across eight enterprise use cases, contributing to $2.3M in competitive displacement wins.
    • Ran a product development program across eight hardware models, delivering 150+ technical requirements per cycle and enabling 20+ carrier rollouts at 99.9% first-time success, worth $4.5M in partner sales.

  4. PhilipsJan to Jun 2020
    Bangalore, India

    Product Engineer, embedded systems (ultrasound)

    • Redesigned the ultrasound scan workflow by observing 20+ cardiologists and radiologists, turning 12 usability gaps into shipped design changes and cutting scan-to-diagnosis time 40% across 300+ machines.
    • Lifted customer satisfaction 15 points across five partner hospital networks by translating clinical workflows into technical specifications for the platform's FDA submission.
    • Automated metadata validation across 200+ fields against HL7 and DICOM standards for clinical software at 200+ hospital sites, taking each dataset from 4 hours to 15 minutes. Three other teams adopted it.

Writing and smaller work

How I think on paper: working-backwards docs, roadmap trade-offs, case answers, and teardowns.

Portrait of Pratyush Singh

About

I came to product through a computer science degree, three years of hardware-in-the-loop quality work at Nokia, and a product seat at Reuters where I learned what it costs when a trading desk's SLA slips. This summer I worked on fulfillment operations at Amazon. Now I'm at USC Marshall as a Dean's Merit Scholar, sharpening the strategy and finance side.

I work best where the brief is a little unclear and the engineering is a little hard. Outside work I founded Trektroops, which grew from a 12-person group to 570+ paying members, held together by a leader-rating system that kept quality up across 50+ trips at once.

I'll happily talk cricket analytics, sim racing, personal finance or espresso.

Education

  • MBA, USC Marshall, May 2027. STEM designated. Dean's Merit Scholarship, GMAT 760.
  • B.E. Computer Science, R.V. College of Engineering, 2020. First class with distinction.

At Marshall

  • AVP, Marshall Product Club
  • AVP Marketing, South Asian Business Association (sponsorship up 4×, $8K to $32K)
  • Finalist, Marshall Tech Strategy Case Competition, 2nd of 18

Certifications

Hiring for a product role in 2027?

pratyush.singh.2027@marshall.usc.edu
LinkedIn Résumé (PDF)