Hi, I'm Ajay Sharma —

I work where
AI systems meet product.

2.9+ years across AI evaluation and product operations — improving model quality, solving user conversion problems, and shipping AI products grounded in real workflows.

See My Work → Download Resume
0
Frontier Lab Programs
0
Feature Adoption Growth
0
Reasoning Failure Axes
0
Trial-to-Paid Conversion
Dual-Pass
Factuality Method
0
Observed Conversion Lift
0
AI Products Shipped
AI-NATIVE PROJECTS

Products built from problem framing to working prototype.

Each project started with a user problem, not a technology choice.

Case Study Live Prototype

Signal - Returns Discovery Tool

problem

  • Operational PMs spend 7-8 hours per week manually extracting and synthesizing signals across 5 disconnected data sources, with no existing tool that performs ranked, cross-source synthesis at scale.

user segment

  • PMs owning returns, fulfillment, or CX verticals at product-led companies who need ranked problem signals, not raw insight dumps.

solution

  • Built a live AI pipeline that ingests 5 source types, returns data, CRM transcripts, reviews, fraud signals, and NPS, processes each through a dedicated RAG layer, and synthesizes a ranked top-5 problem list with confidence scores and evidence citations.
  • Chose separate RAG namespaces per source over a unified index to prevent signal contamination across qualitative and quantitative data types, a deliberate tradeoff that increases infra cost but preserves retrieval integrity.

impact & metrics

  • Targets a reduction in PM discovery time from 7-8 hours per week to 2 hours, a 75% time savings, validated against a manual extraction baseline.
  • Currently processing 5,700+ live records; PM signal adoption target is 60% vs. 0% baseline, measured by PMs tagging the tool as the source of their problem definition.
View Prototype →
Signal Landing Page
Signal ★
Case Study AI Simulator

RoleBridge - Career Transition Interview Simulator

problem

  • Career-transition candidates lose offers in the interview room, not at the resume screen — their answers collapse under follow-up pressure because they have never practiced explaining their experience in the target role's language.

user segment

  • Professionals actively switching roles or industries who have the experience to get shortlisted but lack a way to stress-test their interview narrative before the real conversation.

solution

  • Built an AI-driven interview simulator that parses the user's actual resume and target job description to generate contextually grounded, role-specific follow-up pressure — not generic practice questions.
  • Chose to focus the session on a single user-selected resume section rather than a broad interview, forcing deep articulation over surface coverage — the key design tradeoff that separates drill-down from generic coaching.

impact & metrics

  • Product delivers a structured evaluation report to the user's inbox scoring evidence quality, ownership clarity, and role-language fit — giving candidates a concrete pre-interview diagnostic.
  • Target: session completion rate as primary success metric; report open rate as the secondary signal of perceived value.
View Prototype →
RoleBridge Landing Page
RoleBridge ★
Live Product AI Workflow

Resumatch — AI Resume Tailoring Engine

problem

  • Job seekers applying to multiple roles spend 20–30 minutes manually tailoring their resume per application, a bottleneck that makes high-volume search unsustainable.

user segment

  • Professionals applying at volume who need role-specific, recruiter-ready positioning without rewriting their accomplishments from scratch each time.

solution

  • Built a web app that accepts a base resume and a job description, runs a 9-stage AI pipeline with 4 independent rewrite passes across Summary, Experience, Projects, and Skills, and outputs a downloadable PDF in approximately 1 minute.
  • Chose section-isolated AI rewrites over full-document generation to prevent hallucination — factual data including titles, tenures, and company names are locked and never touched by the model.

impact & metrics

  • Reduces a 20–30 minute manual editing cycle to a 1-minute automated run, targeting over 95% time reduction per application.
  • 4 targeted AI decisions across a 9-stage pipeline — each section rewritten with dedicated context, not a single monolithic prompt.
View Prototype →
Resumatch Landing Page
Resumatch ★
Case Study Live Project

Swiggy Instamart — Smart Cart Retention Feature

problem

  • Quick-commerce cart pages serve static, category-based recommendations that ignore real-time cart composition, missing high-intent upsell moments for shoppers mid-session.

user segment

  • Mobile grocery shoppers building mixed carts on Swiggy Instamart who are receptive to contextual product discovery at the point of checkout.

solution

  • Designed and prototyped a dual-engine cart recommendation system combining a deterministic rule layer for reliability with a live LLM (Gemini 2.0 Flash) layer for contextual bundling (e.g., cart containing "Chips" surfaces a "Movie Night" recommendation cluster).
  • Chose a server-side API architecture over client-side AI calls to eliminate API key exposure — a deliberate security-first tradeoff that also reduced attack surface without impacting recommendation latency.

impact & metrics

  • Identified and remediated a high-severity API key exposure vulnerability present in the original client-side implementation.
  • Identified redundant AI calls as a latency and cost risk during rapid cart toggling and scoped a caching solution as a next-sprint requirement.
View Prototype →
Smart Cart ★
Swiggy Instamart Landing Page
EXPERIENCE

Where I've Built Product Judgment

June 2024 – April 2026

Team Lead - AI Evaluation & Quality Systems

Turing

RLHF · Benchmark Design · LLM Evaluation · Quality Systems
  • Diagnosed broken production rubrics scoring near-zero (0.0–0.2) by isolating conflicting constraints and vague referents; enforced an 8-section verification checklist that restored benchmark scoring reliability across a frontier LLM engagement.
  • Designed adversarial multi-turn conversations using a structured 3-phase methodology - context seeding, topic steering, implicit recall - to systematically surface model failure modes across 4 reasoning axes (RVE, SC, IM, IR).
  • Identified systematic false positives in an internal LLM evaluator by reconciling automated flags against human QA outcomes; shared structured edge-case examples that helped refine evaluation alignment and usability.
  • Promoted from IC to dedicated QA based on evaluation output quality; shaped multimodal training data reliability across OCR, diagram reasoning, and chart interpretation within an RLHF pipeline for a major platform.
  • Led teams of 3–14 across sustained engagements and high-churn bootcamps; maintained annotation quality standards, disseminated guideline revisions, and delivered performance-based staffing recommendations to leadership.
April 2023 – April 2024

Associate Product Operations Manager

Preplaced

Problem Discovery · Marketplace Trust · Funnel Optimization
  • Diagnosed trial-quality degradation and supply-side capacity overload as dual conversion blockers via Mixpanel data and 20–25 mentor interviews; contributed the pre-booking filter concept that became a customizable screening and intake-cap feature.
  • Surfaced a post-trial context-decay problem - where buyer conviction faded within 1–2 days despite positive sessions - and contributed to a structured post-session plan; adoption scaled from 3.6% to ~60% in six months, associated with a ~21% observed uplift in trial-to-paid conversion.
  • Translated recurring mentor no-show complaints into a rule-based accountability framework with explicit thresholds, reducing no-show rates by 10% and strengthening marketplace trust for first-time users.
  • Operated close to users and funnel data in a founder-led pre-seed startup; repeatedly pulled into founder discussions because of direct signal from calls, objections, post-trial outcomes, and conversion patterns.
WHO I AM
Ajay Sharma

Two adjacent tracks. One clear direction.

Preplaced gave me close exposure to user problems, conversion funnels, and marketplace operations in a founder-led startup. Turing put me inside AI evaluation systems - diagnosing rubric failures, designing adversarial test cases, and learning what makes quality measurement trustworthy or noisy. Together, those tracks built the foundation I bring to product work: user insight, systems thinking, and ground-truth knowledge of how AI breaks in practice. AI PM is the convergence, not a pivot.

Connect on LinkedIn →

Shipped AI products. Operated inside RLHF systems. Ready for the product side.

Open to PM, APM, Product Analyst, and Founder’s Office roles where product sense and AI depth both matter.

Email Me → Download Resume → GitHub → Connect on LinkedIn →