Himanshi Agrawal — AI Product Manager Portfolio

Product Case Studies

Circuitry.ai · Service AIdvisor · Jul 2024

How do you get a field technician the right answer in under 30 seconds?

From broken knowledge retrieval to a 45% faster resolution time across Fortune-500 deployments

🔴 The Problem

Field technicians servicing heavy equipment were calling support lines or flipping through 500-page PDFs to diagnose faults. Resolution times were high, first-time fix rates were low, and manufacturers were losing client trust. Existing keyword search returned irrelevant pages — it couldn't understand intent.

🔵 My Product Decision

I defined the core user need: "the right answer to a specific fault, in plain language, in under 30 seconds." I chose RAG over fine-tuning because manufacturer documents change frequently — we needed retrieval, not memorisation. I scoped the MVP to Service + Parts before expanding, so we could validate quality with one client before scaling.

🟢 The Outcome

Deployed across Takeuchi (heavy equipment). Resolution time dropped 45%. Technician productivity lifted 35% vs. keyword-search baseline. LLM error fallbacks reduced 23% after iterating on chunking strategy and prompt configuration. Product became the anchor for 3 Fortune-500 client expansions.

45%

Faster resolution time

35%

Productivity boost

23%

LLM fallbacks reduced

Fortune-500 clients

Key Tradeoffs I Made

✅ Chose

RAG over fine-tuning — manufacturer docs update constantly; retrieval is more maintainable and accurate for dynamic knowledge bases.

❌ Didn't do

Fine-tuning the model on client data — would have required retraining on every document update, too slow and expensive for enterprise clients.

✅ Chose

Custom chunking per document type — parts manuals, service guides, and warranty docs have different structures; one chunking strategy broke quality.

❌ Didn't do

Fixed chunking across all docs — initial approach caused 23% fallback rate; we iterated to semantic + hierarchical chunking to fix it.

PRD Snapshot — Service AIdvisor

Problem

Technicians spend 20+ minutes per fault diagnosis using static PDFs and support lines, resulting in high resolution times and low first-time fix rates.

Goal

Enable a technician to get the correct fault diagnosis answer in under 30 seconds, in plain language, without calling support.

Success Metric

Primary: Resolution time per ticket. Secondary: First-time fix rate, technician satisfaction score, DAU/MAU ratio.

Out of Scope

Predictive fault detection, IoT sensor integration, automated parts ordering — deferred to Parts AIdvisor v2.

Key Risk

Wrong answer in a field situation = safety risk. Guardrail: fallback to source document + confidence score threshold before response is surfaced.

RAGMilvusCustom ChunkingPRDEnterprise B2BHeavy Equipment

Circuitry.ai · Product AIdvisor · Apr 2024

How do you turn a product catalogue into a sales conversion engine?

From static product pages to a 30% sales conversion uplift — published Business Wire, Apr 2024

🔴 The Problem

Manufacturer sales reps and end customers were unable to navigate complex product catalogues to find the right product for their use case. Sales cycles were long, reps were overwhelmed with repetitive product queries, and customers were dropping off at the selection stage. The knowledge existed — it just wasn't accessible in the moment of decision.

🔵 My Product Decision

I framed this as a decision support problem, not a search problem. The user doesn't want results — they want a recommendation with a reason. I defined the MVP around conversational product guidance with cited sources so sales reps could trust and verify answers. I prioritised time-to-first-value: a rep needed to get value in their first session, not after a 2-week onboarding.

🟢 The Outcome

30% sales conversion uplift published on Business Wire (Apr 2024). Reduced repetitive product queries to sales reps, freeing them for higher-value conversations. Became the flagship product used in all subsequent enterprise POCs and the anchor for Circuitry.ai's Series A positioning.

30%

Sales conversion uplift

Apr 2024

Business Wire published

15+

Enterprise POCs run

Key Tradeoffs I Made

✅ Chose

Conversational guidance with cited sources — reps needed to trust and verify answers before sharing with customers. Trust was the product.

❌ Didn't do

Pure search results UI — returning a list of products without a recommendation doesn't solve the decision problem. It just moves the work to the user.

✅ Chose

Narrow MVP scope — product catalogue only, not full sales workflow. Validated quality and trust with one client before expanding to warranty and parts.

❌ Didn't do

Full CRM integration in v1 — would have tripled scope and delayed launch by 3+ months. Deferred to v2 after proving core value.

PRD Snapshot — Product AIdvisor

Problem

Sales reps spend disproportionate time on repetitive product selection queries; customers drop off at the product selection stage due to catalogue complexity.

Goal

Enable a sales rep or end customer to get a confident, cited product recommendation in under 60 seconds.

Success Metric

Primary: Sales conversion rate. Secondary: Time-to-recommendation, rep satisfaction score, repeat session rate.

Out of Scope

CRM integration, automated quote generation, pricing engine — all deferred post-MVP validation.

Key Risk

Wrong product recommendation = lost sale + damaged trust. Guardrail: source citation mandatory on every response; confidence threshold before surfacing answer.

Sales ConversionDecision IntelligencePRDB2B SaaS0→1 Launch

Circuitry.ai · Platform Innovation · 2025

Thousands of hours of video knowledge — completely unsearchable. What do you do?

Ideating a new knowledge modality that no competitor offered

🔴 The Problem

Manufacturing clients had thousands of hours of video content — assembly walkthroughs, repair guides, training demos. None of it was searchable. The standard assumption was: no transcript = no retrieval. Generating transcripts for all videos was expensive, slow, and required client cooperation on audio quality. The knowledge was locked.

🔵 My Product Decision

I questioned the core assumption. A video doesn't need a transcript to have retrievable structure — it has a title, description, chapter markers, duration, upload context, and playlist position. I defined a metadata-inference strategy to create knowledge representations from this signal alone, unlocking retrieval without touching the audio stream. I scoped it to YouTube-hosted content first for fastest validation.

🟢 The Outcome

Shipped Transcriptless Video Retrieval — an entirely new knowledge modality. Extended with YouTube playlist ingestion and timestamp-level citations. Field technicians could now query video content and be pointed to the exact timestamp relevant to their fault. No competitor offered this. Became a key differentiator in enterprise sales conversations.

Key Tradeoffs I Made

✅ Chose

Metadata-inference over transcript generation — faster to ship, no dependency on audio quality, works immediately on existing video libraries.

❌ Didn't do

Auto-transcription pipeline — expensive at scale, slow to process large libraries, and broke on low-quality audio common in factory environments.

✅ Chose

YouTube-first scope — clients already hosted training content there; zero migration cost and instant library access via YouTube API.

❌ Didn't do

Build a video hosting platform — massive scope creep, not our core problem. Use what exists.

Product InnovationMetadata InferenceMulti-modal RetrievalFirst PrinciplesYouTube API

Circuitry.ai · Noahmed Medical Devices POC · 2025

When accuracy isn't a nice-to-have — it's a compliance requirement

Designing a zero-hallucination AI pipeline for medical device field service

🔴 The Problem

Noahmed's field service engineers needed AI-assisted guidance for medical device maintenance procedures — but the compliance bar was extreme. A hallucinated answer in a medical context isn't a bad user experience. It's a regulatory and safety risk. Standard RAG with LLM summarisation was not safe enough. The product needed to be redesigned from the ground up for zero-hallucination tolerance.

🔵 My Product Decision

I made a deliberate call to prioritise accuracy over convenience. Instead of AI-generated summaries, I designed the pipeline to surface the exact source document section as a downloadable PDF slice — the AI identifies the right page, the engineer reads the original. Gemini with custom tool-calling was used to navigate to the right section. No paraphrasing. No summarisation. Source only.

🟢 The Outcome

Delivered a zero-hallucination pipeline that passed compliance review for the medical device POC. Engineers got pinpoint navigation to the exact page of 23-page compliance procedures — faster than manual search, with full source traceability. Opened a new vertical for Circuitry.ai in regulated industries.

Key Tradeoffs I Made

✅ Chose

Sliced PDF delivery over AI summarisation — in regulated contexts, the source document IS the answer. Any paraphrasing introduces compliance risk.

❌ Didn't do

LLM-generated summaries — convenient, but legally and regulatorily unacceptable for medical device maintenance. Accuracy over experience.

Medical DevicesZero-HallucinationComplianceGeminiProduct Judgment

Circuitry.ai · Exedy Automotive · 2025

A client's data was too complex for any standard pipeline. I had to think differently.

Using an LLM as a data normalisation layer — unblocking a deployment in hours, not weeks

🔴 The Problem

Exedy Automotive supplied torque specifications for F1-grade vehicles across dozens of complex, inconsistently structured multi-sheet Excel files. No two files had the same column headers, units, or structure. Standard data ingestion pipelines failed completely — the variance was too high for rules-based normalisation. The deployment was blocked. Manual data engineering would have taken weeks.

🔵 My Product Decision

I reframed the problem. This wasn't a data engineering problem — it was a comprehension problem. I proposed using Claude as an LLM-based normalisation layer: pass each sheet to the model with a structured output schema, let it interpret the intent of each column regardless of how it was labelled, and extract clean structured data. Nobody on the team had framed it this way. I defined the prompt schema, tested it across 5 sheet variants, and validated output accuracy before recommending it as the production approach.

🟢 The Outcome

Client data was normalised and ingested in hours — not weeks. Deployment unblocked. The approach became a reusable pattern for other clients with messy structured data. This is the story I use when asked about creative problem-solving under constraints: the best solution wasn't a new tool — it was applying an existing one in a way nobody had thought of yet.

Key Tradeoffs I Made

✅ Chose

LLM as normalisation layer — handles structural variance that rules-based systems cannot. Interpreting intent is exactly what LLMs are built for.

❌ Didn't do

Rules-based ETL pipeline — would require a custom rule set per file variant. With dozens of inconsistent files, this was weeks of work with no guarantee of coverage.

✅ Chose

Validate across 5 sheet variants before recommending for production — accuracy in an automotive context is non-negotiable. Test before you commit.

❌ Didn't do

Ship immediately without validation — speed mattered, but a wrong torque spec in an F1 context has real consequences. One pass of validation was worth the delay.

Creative Problem SolvingAutomotiveLLM Data NormalisationAnthropic ClaudeProduct Instinct

Intellectual Property · 2024

Industrial Design Patent — Transformable Multi-Door Refrigerator

Invented and patented a modular multi-door refrigerator that allows users to reconfigure compartment layouts based on storage needs — solving a flexibility problem that standard fridge form factors hadn't addressed. Demonstrates the same first-principles product thinking I apply to software: start with the real user constraint, not the existing form factor.

Industrial DesignFirst PrinciplesModular Architecture

Patent No. 04346-001 · Granted 2024

Building AI products
that work in production

3 years on a Fortune-500 account before becoming a PM

I define the problem before I define the solution.

Let's build something together.

Building AI productsthat work in production

3 years on a Fortune-500 account before becoming a PM

I define the problem before I define the solution.

Let's build something together.

Building AI products
that work in production