PPawanchander.
--:--:-- ISTPROJECTS: 5TEARDOWNS: 2DOMAINS: 4
Selected work/05 cases

Products led end to end

Each case moves from the real problem to the insight that changed the plan, the decisions and their tradeoffs, the outcomes, and an honest note on what I would revisit.

01Consumer EdTech · AI guidance · 2025

Steerly

Steerly logo

An AI career-discovery platform that helps students in India choose a path with clarity, before they commit years and money to it.

steerly.me
Role
AI PM (consulting)
Timeframe
Current
Surface
Web · India
Stage
Live, 5,000+ students
Context

In India, a teenager picks Science, Commerce or Engineering under enormous pressure and almost no real information. That single choice sets up years of study and a large financial bet for the family. Steerly turns that leap of faith into a structured, personal, data-backed exploration.

The real problem

Career choices were being made on hearsay, coaching-center defaults and parental anxiety. Generic interest quizzes told students a label but never let them feel the actual work, and none of the advice was grounded in India-specific salary, growth or AI-impact data.

Insight that changed the plan

Watching families use early versions, two things were obvious. The real decision-maker is often the parent worried about wasted investment, and interest surveys do not build conviction. People commit only after they experience the work, so the product had to let them try a career, not just read about one.

Decisions & tradeoffs
  1. 01

    Simulations over static career lists

    Built scenario-based "week in the life" missions where a student makes real decisions and sees whether they enjoy the work style, instead of scrolling salary tables. Experiencing the job is what moves someone from curious to convinced.

    Tradeoff

    Each career simulation is far heavier to design and build than another quiz question, so breadth of coverage grows slowly in exchange for genuine conviction.

  2. 02

    The AI is a guide, not an oracle

    AI future-impact and trend indicators are framed as forecasts with ranges and reasoning, tied to India-focused career and university data, never a single magic answer a family would over-trust.

    Tradeoff

    Honest ranges feel less impressive than one confident number, but they protect a decision that a family cannot easily undo.

  3. 03

    Design for two people, not one

    The student explores, assesses and simulates; the parent gets an investment-path view with salary and job-outlook data, bookings and secure payments. Two surfaces, one decision they make together.

    Tradeoff

    Serving the parent doubled the surface to design, but it matched who actually pays and who actually worries.

Outcomes
5,000+
Students
onboarded to date
10 min
To a career profile
median first session
4
Steps to clarity
discover to expert call
3
Profiles
student, parent, expert
What I'd underline

Refusing the one-shot answer is the call I defend most. In a category that sells certainty, honest exploration was slower to build and is the only version a family should rely on.

What I'd revisit

Hand-designed simulations are the moat and the bottleneck. Coverage grows slowly, and I am still unsure whether a semi-generated simulation could keep the conviction without the authoring cost. The honest open question is how far quality survives automation here.

02Marketplace · Applied AI · 2025

Ollie

Ollie logo

A two-sided marketplace connecting Madrid dog owners with vetted local pet-care pros, with a Groq-powered assistant handling the questions in between.

Role
AI PM (consulting)
Timeframe
0 to 1
Surface
Web · Madrid
Stage
Live across 21 districts
Context

Finding someone you trust to walk, board, groom or train your dog in Madrid meant piecing together word of mouth, classifieds and neighborhood chats. Ollie set out to be the single trustworthy place for that, with real people and real vetting behind every match.

The real problem

Owners had no reliable way to judge whether a walker or groomer was safe and right for their dog, and small local businesses had no structured channel for qualified leads beyond their own network. Both sides were guessing.

Insight that changed the plan

In pet care, trust is not an average of star ratings. It is a person vouching for another person. That single belief shaped the whole product: match by hand for quality, and use AI to absorb the heavy load of questions and local nuance so human matching could stay the special part.

Decisions & tradeoffs
  1. 01

    Match by hand, not by algorithm

    At zero to one a marketplace has too little data for a good algorithm, so matches are made by a person based on the dog and the neighborhood. It set the quality bar and generated the data a future algorithm would need.

    Tradeoff

    Hand-matching does not scale yet, which was a deliberate early choice in favor of trust and learning over raw throughput.

  2. 02

    Scope the AI assistant to guidance, not booking

    The assistant, built on the Vercel AI SDK with Groq running LLaMA 3.3 70B, streams answers about pet care and Madrid-specific rules and dog-friendly areas. It informs owners but never closes the match, so humans stay accountable for trust.

    Tradeoff

    The AI does not complete the transaction, so it looks less autonomous, but correctness and trust stay high where it matters.

  3. 03

    Meet a local audience where they already are

    Fully bilingual English and Spanish, with WhatsApp as a first-class contact path next to the structured request form and a reference-number status tracker.

    Tradeoff

    Supporting two languages and an informal channel added surface, but it removed the friction that would have lost local owners.

  4. 04

    Petfolio as a retention wedge, validated first

    A digital passport for a dog, its health records, preferences and history, shareable with any carer in one tap, launched as a waitlist to test pull before building the full product.

    Tradeoff

    Deferring the build to a waitlist slowed the roadmap but avoided investing in a feature before demand was proven.

Outcomes
21
Madrid districts
served at launch
4
Care categories
walks, daycare, grooming, training
100%
Hand-matched
personal match, not an algorithm
EN / ES
Fully bilingual
local by default
What I'd underline

The instinct in a marketplace is to automate matching on day one. Doing it by hand, and letting AI carry only the guidance, is what kept quality high while the data caught up.

What I'd revisit

Hand-matching gave us quality but capped supply, and I underestimated how quickly that becomes the growth ceiling. If I ran it again I would instrument the manual matches from week one as labeled training data, so the handoff to an assisted match had a real dataset to stand on instead of intuition.

03Consumer · Edge AI · 2024-25

Sargam AI

An on-device AI tutor that turns any skill into a personal, adaptive practice loop, with feedback fast enough to keep people coming back.

Role
AI PM (consulting), 0 to 1
Timeframe
14 months
Surface
iOS / Android
Stage
Public beta
Context

Learners abandon skill apps not because content is missing, but because feedback is generic and arrives too late. Sargam AI was built on a bet: if the feedback loop lives on the device, it can be instant, private and personal enough to keep someone coming back.

The real problem

Cloud-based tutoring felt slow and impersonal, and every interaction shipped a learner voice and mistakes to a server. Retention past week two collapsed. The real problem was trust and latency, not model quality.

Insight that changed the plan

When I watched 30 first sessions, the drop-off was not at the hard content. It was at the half-second pause before feedback. Perceived responsiveness, not accuracy, was the retention lever.

Decisions & tradeoffs
  1. 01

    On-device inference as the default

    Shipped a quantized model that runs locally so feedback returns in under 300ms and a learner data never leaves the phone. The cloud became the exception, used only for heavy generation with explicit consent.

    Tradeoff

    Gave up model size and some answer quality for latency and privacy, and accepted a harder engineering path to protect the core loop.

  2. 02

    One adaptive loop, not a feature menu

    Cut a roadmap of 11 features down to a single practice, feedback and adjust loop. Everything that did not tighten that loop was deferred.

    Tradeoff

    Launched looking too simple to some stakeholders, betting depth in one loop beats breadth across many.

  3. 03

    Confidence-gated feedback

    The model only corrects when its confidence clears a threshold. Below it, it asks a question instead of asserting. This kept trust intact when the small model was unsure.

    Tradeoff

    Fewer corrections surfaced, but the ones shown were trusted, measured by a 3x drop in "that is wrong" reports.

Outcomes
<300ms
Feedback latency
p50, on-device
+41%
D14 retention
beta cohort vs cloud baseline
0
Voice data to server
default on-device path
4.7
Store rating
public beta, ~2.1k reviews
What I'd underline

The winning move was subtractive. Cutting ten features to perfect one loop was the hardest call to defend, and the one that moved every metric that mattered.

What I'd revisit

The retention lift came from a beta cohort of motivated early users, so I hold that number loosely until it survives a colder audience. I also over-indexed on latency and under-invested in content depth. Fast feedback kept people in week two, but by week six the thin skill library, not speed, was the real churn driver.

04Consumer · Mobile safety · 2024

HoldOn

HoldOn logo

A proactive, multi-layer phone-theft prevention app for Europe, distributed through the insurers who benefit most when theft goes down.

Role
AI PM (consulting)
Timeframe
11 months
Surface
Android / iOS · Flutter
Stage
Piloted in Spain
Context

European phone theft is relentless. Barcelona sees hundreds of thefts a day and London tens of thousands a year, against a phone-insurance market worth over seven billion euros with rising claims. Every existing tool was reactive: it helped you find a phone after it was already gone.

The real problem

Recovery is too late for the victim and too expensive for the insurer. There was no comprehensive, European-focused product that tried to stop the theft as it happened, and a pure consumer app faced brutal acquisition costs against a worried but price-sensitive audience.

Insight that changed the plan

The person most motivated to pay was not the anxious owner, it was the insurer paying out claims. Reframing HoldOn from an app you sell to nervous users into claims-reduction you sell to insurers solved distribution and cost at once, because prevention is something an insurer will happily subsidize.

Decisions & tradeoffs
  1. 01

    Proactive multi-layer detection over recovery

    Combined motion sensing at a tuned 3.5g threshold, location hotspot awareness and unlock-timeout monitoring to catch the moment of theft, covering far more scenarios than any single-signal approach.

    Tradeoff

    Multiple signals mean more false positives to tune and more engineering than a simple find-my-phone feature, spent to actually prevent rather than console.

  2. 02

    Insurance partnerships as the go-to-market

    Rather than buy users one by one, HoldOn goes to market through insurers with revenue shared on the premium discounts prevention unlocks, turning the people who lose money on theft into the distribution channel.

    Tradeoff

    Partnership sales are slower and relationship-heavy, chosen over fast but expensive direct-to-consumer growth for a far healthier lifetime-value-to-cost ratio.

  3. 03

    European-first and GDPR-native

    Localized for Spanish, Italian and French tourists and built for European privacy law from the start, deliberately targeting a region global players would not bother to localize deeply for.

    Tradeoff

    Narrowing to Europe first gave up the appearance of a global launch in return for a defensible, winnable beachhead.

  4. 04

    Prove it with a 48-hour Flutter MVP

    Shipped a cross-platform MVP in two days to put real theft-prevention flows in front of pilot users and insurers before committing to a heavier build.

    Tradeoff

    The MVP had rough edges, accepted in exchange for fast, real signal on whether the core idea held.

Outcomes
90%+
Theft scenarios
internal coverage estimate
10-20x
LTV / CAC
modeled target via partners
3.5g
Motion threshold
tuned in pilot
€7.74B
Insurance market
European TAM
What I'd underline

The hardest call was refusing the consumer-first instinct everyone expected. The durable moat was the insurance relationship, not app-store installs, and choosing the slower channel is what made the model real.

What I'd revisit

The insurer channel fixes acquisition cost but hands your roadmap to a partner with a long sales cycle, and I under-planned for that dependency. The unresolved tension is detection sensitivity: the 3.5g threshold that catches a grab also fires on a dropped phone, and I never got the false-positive rate low enough to feel truly shippable at scale.

05Enterprise · Cloud NLP · 2022-23

FARO

FARO logo

A trust layer for enterprise NLP that turns black-box model output into decisions a team can defend to a regulator.

Role
AI PM (consulting)
Timeframe
16 months
Surface
Cloud platform
Stage
GA
Context

Enterprises do not buy accuracy, they buy defensibility. FARO, named for a lighthouse, wrapped NLP pipelines in the evidence, controls and audit trail that let a risk-averse organization actually ship a model into a regulated workflow.

The real problem

Models tested well and stalled in procurement. Legal and risk teams could not sign off on output they could not explain or trace. The bottleneck was governance, not machine learning.

Insight that changed the plan

The buyer was not the data scientist. It was the person who would be blamed if the model was wrong. FARO was built for the signer, not the builder.

Decisions & tradeoffs
  1. 01

    Evidence attached to every output

    Each prediction ships with its sources, confidence and the rule path that produced it, so a reviewer can audit any single decision.

    Tradeoff

    Heavier payloads and more surface to design, exchanged for the explainability that unblocked legal sign-off.

  2. 02

    Human-in-the-loop as a first-class state

    Low-confidence output routes to a review queue by design, with SLAs and ownership. Escalation is a feature, not a failure.

    Tradeoff

    Slower full automation, in return for a path to production that risk teams would actually approve.

  3. 03

    Eval harness before capability

    Shipped the offline evaluation and drift-monitoring harness before adding new model capabilities, so every claim was measurable.

    Tradeoff

    Delayed flashy features by a quarter to build the boring infrastructure that made the platform trustworthy.

Outcomes
-73%
Time to sign-off
first pilot accounts
100%
Decisions traceable
by design
5
Regulated deploys
first year
2.4x
Net expansion
early cohort
What I'd underline

In enterprise AI, the roadmap is a trust-building sequence. We won by shipping governance before capability, the opposite of what the demo instinct wanted.

What I'd revisit

Leading with governance won procurement but slowed our early feature velocity, and we lost two fast-moving prospects to flashier demos while we built the audit trail. I still debate whether a lighter, faster proof-of-value first, with the compliance layer close behind, would have widened the top of funnel without costing us the deals that needed defensibility.