Built on fifteen years of product, trust, and behavioral work — and two years building with LLMs full-time.
I'm Simrat Bath. For fifteen years I've worked on product, trust, and behavioral systems — at Niantic on Pokémon GO, and as the lead designer on A2i (Assessment to Instruction) at Learning Ovations, then on Scholastic's literacy platform onboarding after the acquisition. Two years ago I started building with LLMs full-time. Project Heer, my 2025 personal research on AI persona deployment, ran two Instagram channels with custom-trained models and documented how AI disclosure shifts audience behavior, how gender dynamics emerge in AI engagement, and where solo creators face safety gaps platforms don't address. I also advise a stealth-stage AI startup whose recent pilot ran with 82 families and produced a 74% confidence lift and 28% week-10 retention. Now I run NS Studio LLC, an AI product advisory for F500 teams, and I'm researching how those teams design oversight for AI systems they don't fully control. I help leaders make three decisions: which AI products to build, which to kill, and how to ship the rest. I write about it under "Kill or Ship."
years in product, design, research, and behavioral systems.
building with LLMs full-time — evals, RAG, agentic flows, HITL.
countries where teachers use A2i, the literacy assessment platform I led design on.
families in a stealth-stage AI startup pilot I advised — 74% confidence lift, 28% week-10 retention.
I build AI-enabled products the way strong teams actually need them built: define the workflow, identify failure modes, specify guardrails, align cross-functional partners, and ship against measurable outcomes—not hype.
My career has always centered on how people interpret systems, where they lose trust, and what they do when products fail to match real behavior. That is the core of AI product management: turning ambiguous human workflows into systems that are useful, safe, evaluable, and ready for production.
I make roadmap calls when the signal is messy: stop full automation when false positives are too costly, choose offline-first when field conditions demand it, and prioritize workarounds when users have already written the spec with their behavior.
I define release gates around quality, safety, reviewer agreement, escalation reasons, override rates, and real user outcomes, not vanity demos or vague claims that the model is “better.”
I know where humans belong in the workflow: high-risk decisions, low-confidence model output, edge cases, cultural context, and irreversible actions. The goal is not automation at all costs. The goal is accountable scale.
I can move from discovery to PRD to prototype to launch because I have sat in the research, design, engineering, executive, startup, and operations seats. I do not need translation between functions. I am the translation layer.
These projects show the pattern: find the real user behavior, define the system, make the tradeoff, ship the intervention, and measure whether it actually changed the outcome.
A self-funded study of what happens when you actually deploy an AI persona in public. I trained custom image models with DALL-E, Gemini, Flux, ElevenLabs, and Replicate, ran two Instagram channels (a fashion-brand marketing track and a college-life persona track), and documented the trust, disclosure, and safety dynamics that show up the moment AI content meets real audiences. Predates FLUX.2, Nano Banana, and Midjourney's web product — every output required substantially more prompt engineering than the same work would today.
At Learning Ovations, I led product and UX thinking for A2i, an Assessment-to-Instruction system that turns student reading data into specific instructional recommendations, grouping decisions, and lesson plans for teachers. After Scholastic acquired Learning Ovations, I continued that work, focusing on how educators onboard, interpret recommendations, and turn assessment signal into classroom action.
Players with large friend lists were already hacking the product with symbols in names to identify close friends. I treated that behavior as a product requirement, not a curiosity, then designed a zero-learning-curve feature around an interaction players already understood.
I built an end-to-end AI content pipeline for independent creators from research to revenue: demand analysis, workflow architecture, LLM-assisted creation, human review gates, marketplace publishing, and performance iteration.
My style is direct, evidence-led, and outcome-obsessed. I am warm with people and ruthless with weak product logic.
When users invent parallel systems, the roadmap is already speaking. The PM’s job is to listen before the behavior calcifies into churn.
The best AI products know when to automate, when to route, when to explain, and when a human must own the final decision.
I build measurement into the spec because post-launch learning should not depend on dashboards someone remembered to add later.
The arc: language and meaning, product design, scaled consumer UX, founder-level ownership, and now an AI advisory built on the parts of product work that don't change when the model changes.
An AI product advisory for F500 teams. Audits, sprints, and embedded advisory work on which AI products to build, which to kill, and how to ship the rest. Advisory clients include a stealth-stage AI startup whose recent pilot ran with 82 families and produced a 74% confidence lift and 28% week-10 retention.
Two years of hands-on work with LLMs, evals, RAG, agentic workflows, and trust/disclosure research. Project Heer (2025) deployed two Instagram personas built on custom-trained DALL-E, Flux, and ElevenLabs models to study how AI disclosure changes audience behavior, gender dynamics in AI engagement, and creator-safety gaps platforms don't address.
Built an early-stage food-tech startup from zero: brand, product, hiring, ops, investor materials, governance, and the hard leadership calls through dissolution.
Designed reviewer accountability and decision flows for A2i, used by teachers across twelve countries with a 5× reading comprehension lift versus controls. Continued the work inside Scholastic after the acquisition.
Designed the Favorite Friends feature end-to-end — research, personas, prototypes, and UI specs — for a product at global consumer scale.
Led product design and QA for property management software. Shipped a forms feature that grew daily new-user signups from 25 to 400.
M.Phil. in linguistics and literary criticism; taught English literature and founded education and public-health initiatives in rural Punjab.
JTBD, ethnography, usability research, PRDs, roadmap planning, eval design, HITL workflows, RAG knowledge systems, prompt behavior, trust & safety, and executive communication.
All engagements run through NS Studio LLC. Most teams start with a Strategy Sprint.
A read on your AI roadmap: what to keep, what to kill, what's missing. Written diagnostic plus a 90-minute working session with the team.
Pick a single AI product or feature. We pressure-test the bet, design the eval, write the kill criteria, and ship a 90-day plan you can hand to engineering.
Ongoing advisory on AI product decisions — weekly working session, async review, and direct input on roadmap, evals, and ship/kill calls.
A short field guide to the seven signals that an AI pilot is quietly dying — and the decision tree I use with F500 teams to call it before the budget runs out.
The framework I use with enterprise teams to decide whether a pilot graduates, gets re-scoped, or gets killed. ~12 pages.
Read the guide →A 2025 field report on AI persona deployment, disclosure dynamics, gender risk, and oversight gaps. Built pre-FLUX.2 with DALL-E, Gemini, Flux, and ElevenLabs. ~6 pages.
Read the case study →More writing at nsstudiollc.com/blog.