KnewsIQ

The problem

I began exploratory work on news-consumption problems in March 2025, out of dissatisfaction with news apps. During the Pahalgam terrorist attack and Operation Sindoor (April to May 2025) I could not picture where events were happening, and that gave the concept its shape: a map-based product showing Indians how power works globally, organised by geography in place of an algorithmic feed.

The design principles followed from that. Closure over engagement, signal over noise, no infinite scroll and no engagement-maximising patterns. A reader chooses by location, reads a summary, the product marks it done, and the reader moves on. The product was mobile-first.

The harder question was who it was for. KnewsIQ had three target users over its life, which means two changes, and each change rebuilt the product:

1. Initially, Indian news consumers interested in geopolitics.
2. From December 2025 to February 2026, UPSC (civil services exam) aspirants as the primary user and busy professionals as the secondary. On-ground research in February 2026 showed that serving aspirants properly would turn KnewsIQ into an ed-tech product, which was outside the vision.
3. From March 2026, professionals whose work depends on geopolitical awareness (executives, investment banking, consulting, finance, risk, international and policy research, think tanks) as the primary user, and enthusiasts as the secondary.

The UPSC phase taught me that interest driven by overlap with study material can look like product-market fit. Remote interviews captured approval without showing whether users would depend on the product daily, so I now treat on-ground validation of persona depth as a condition for structural product changes.

What the product does

KnewsIQ is a daily geopolitical intelligence briefing for India on an interactive world map. Each day the pipeline processes news from the previous 27 hours and does three things to every qualifying story. It structures the story into a 5W1H summary (who, what, when, where, why, how). It classifies the story on two axes: which instrument of power is at play, and which relationship or theatre is involved. It writes “What This Means for India”, an analysis connecting the event to Indian strategic interests.

Each processed story becomes a Story Card pinned to its location. A reader opens the map, clicks a coloured marker and reads the card. Red means unread, blue marks the selected marker, and green means all of a marker’s stories have been read. A progress bar shows how much of the day’s briefing the reader has completed.

Landing page

The product’s promise, value propositions, and signup with email OTP.

Signed-in map

A full-screen map of the last 27 hours, with filters by location, power domain and strategic front.

/today

The full daily briefing without signup, with five free insight reveals before a signup gate. I built it in under two weeks by duplicating the signed-in map, and had the first gating experiment built, deployed, measured and iterated within three weeks.

Story Card pages

One Story Card on its own shareable page, with signup prompts.

What I did, phase by phase

I ran eight testing phases between July 2025 and August 2026 (~14 months) and interviewed 100+ users across the three target users. Until April 2026 I interviewed every user in the closed and controlled betas. From April 2026, as volume passed what one-to-one interviews could cover, feedback became mainly behavioural, with qualitative input continuing through the seeded communities.

The AI pipeline

An hourly automated run fetches new English-language articles through a news API, resuming from where the previous fetch stopped. Code prepares them before any AI is used. It removes articles already fetched, drops sections of the source that never carry relevant news, and skips multi-topic digests such as daily round-ups and bulletins by matching their titles. Removing these before the AI steps cuts cost and keeps each Story Card to a single story.

The remaining articles go to the AI in small batches, and six AI functions turn each relevant one into a Story Card. The first version used a single AI call and seven backend functions; data and user interviews moved it to six AI functions with separate responsibilities. The pipeline ran on OpenAI GPT models in production, up to GPT-5.6.

Relevance

Decides whether the article qualifies, with a three-part test: is a power-wielding actor involved; is power being exercised, contested or disrupted; does the impact matter beyond the local level. It is anchored in six instruments of power and supported by calibration examples and rules for edge cases.

Fact extraction

The 5W1H structure.

Classification

Both taxonomy axes.

Location

Where the event physically happened or was announced, as distinct from the place it concerns (a meeting in Delhi about the Strait of Hormuz is placed in Delhi). Code matches the result against stored locations, asks the AI to complete partial locations, and calls paid geocoding only for places never seen before, which removed repeat geocoding spend.

Headline

A headline for the Story Card.

What This Means for India

A three-part analysis: the underlying power dynamic (The Real Game), a link to Indian interests the reader would not see at once (Non-Obvious Connection), and the next development to monitor (Watch For).

I used context engineering in production to stop drift, hallucination and unintended bias as an article moved from raw text to Story Card.

• Each function’s prompt sets its role, absolute rules and worked examples, and each function sees only what its job needs. Prompts carry calibration examples and are versioned.
• Fact extraction must state facts directly, keep every specific (names, designations, numbers, dates, amounts), and avoid vague or generic phrasing.
• The India analysis takes an analyst’s stance: strategic logic and consequences, no moral judgement, and the same objectivity for every actor. India’s relationships are analysed on India’s own strategic terms, and no single political or Western institutional reading is treated as the neutral default.
• The taxonomy is read from the database into the classification prompt at run time, so taxonomy changes did not require rewriting prompts.
• Randomness is set per task: strict for relevance, classification and location; looser for headlines and analysis.

Each AI function has one job and a fixed output structure, and code between the functions handles everything that does not need a model. Between the steps:

• Every AI step retries once on failure.
• An article leaves the processing queue only when it is fully processed or explicitly judged irrelevant. Any failure keeps it queued for the next run, so no article is silently lost.
• Extracted facts are checked for completeness, and articles that fail are held back.
• Classification may use only the exact category names supplied. A valid sub-category filed under the wrong parent is corrected automatically, and invalid entries are removed.
• Malformed or incomplete AI responses, and responses for articles that were never sent, are rejected.
• Token usage and outcome counts are logged for every step of every batch, which kept cost and failure rates visible.

Evaluation and quality

Every article passes through all six functions, so a change to any one function can alter the final Story Card. I therefore tested any change across the whole pipeline before release. During development, I exported the pipeline’s inputs, accepted outputs and rejected articles and reviewed them against a checklist (accuracy, edge cases, false positives and negatives, output format, domain correctness), revising prompts until further rounds stopped improving results.

In production I ran a weekly human-in-the-loop, semi-automated evaluation from the first backend deployment until work stopped. I tried automated evals and fine-tuning; running them regularly in production cost too much at this stage, so the weekly evaluation took their place.

I wrote a rubric defining each function’s criteria first, then judged each sampled item against it, comparing Story Cards with their source articles. I ran the review with Claude’s assistance. Each week I grouped the failures and added a guardrail per group before the next round; a typical round found 3 to 4 minor issues.

Rejected articles were sampled because an article wrongly rejected at the relevance step never reaches the other five functions. That is by design, to control cost, but it means the error never appears in any output and no user can report it. Sampling rejected articles tests for these false negatives, and I made the relevance step’s false-negative rate the pipeline’s L1 metric.

Filtering at the relevance step also controlled cost. Without it, six AI calls on 2,000+ articles a week would be ~12,000 calls. With it, the pipeline made ~2,000 relevance calls plus five calls on 600+ relevant articles, ~5,000 in total, about 58% fewer AI calls.

The taxonomy

The taxonomy was the largest sustained research investment in the product. I spent weeks of primary research on how practising professionals and academics in India read and interpret geopolitics, including direct conversations with them, and separately on how best to present content to that audience. The result is proprietary to KnewsIQ and was extended as new categories became necessary.

It had three versions, which means two rebuilds: an initial broad version, a single-axis version of 45 categories for UPSC aspirants (January 2026), and the final two-axis version (March 2026). The final version classifies every Story Card on two axes with 84 sub-categories in total, plus a country, state and city location filter.

Power domain, 39 sub-categories

Which instrument of power is at play: Defence and Military Affairs; Diplomacy and Foreign Policy; Intelligence and Security; Economic Statecraft and Geoeconomics; Strategic Technology and Innovation; Governance and Political Power.

Strategic front, 45 sub-categories

Which relationship or theatre is involved: India (internal); India’s Neighbourhood; Key Bilateral Relationships; Global Theatres.

Distribution and experiments

KnewsIQ is a single-page application, so social crawlers and search bots saw generic previews of shared Story Card pages. I built a Cloudflare Worker that fetched each story and injected story-specific preview tags and structured data at the edge, so links shared on WhatsApp, X, LinkedIn and Facebook showed the story’s headline, description and image. I also set up auto-updating sitemaps, structured data across the site and canonical URLs. Search Console indexing worked at first and then stopped; I attempted diagnosis roughly twice a month from December 2025 to August 2026 and never resolved it, so organic search never became a channel. Every visitor arrived through seeded communities or direct sharing.

• Friends shared the link directly with people matching the target user, including in their alumni WhatsApp groups, with no incentives or paid promotion. This grew into the community outreach.
• From April 2026, five tracked community clusters across 40+ WhatsApp groups, all of which I already belonged to. Links went mostly to /today, or to Story Card pages when sharing a specific item. One cluster came through a geopolitics and geoeconomics creator with 151k+ followers, who started as a user before becoming a distribution partner.
• From 24 March 2026, an X account run on written criteria for threads, standalone posts and analytical replies: no fear or panic, no anti-India stance, invite conversation, smart without condescension. Analytics showed that replies to established geopolitics accounts outperformed original threads.
• In an early snapshot of /today, 33 users generated 116 shares, which brought 234 visitors at no cost.

Across all five community clusters, engagement per session was high and unprompted return was low. Visitors returned mostly when prompted by a new share in their community or by an email. Conversion held steady across clusters, so the constraint lay outside audience quality, messaging and channel. Retention was tested from March to August 2026 and confirmed as the binding constraint; the weekday-lunch email was the only mechanism that moved it.

Results

Around April to May 2026 the product recorded 726 content reads and 566 map interactions, with organic reach across 19 countries and 116 cities and no paid acquisition.

The North Star was WAR-3+1: weekly active readers who read 3+ Story Cards and reveal 1+ insight. Counting Story Cards by themselves would miss use of the differentiator, and counting insight reveals by themselves would miss use of the core content. By August 2026 activated visitors in all five clusters met the depth criteria consistently, while unprompted weekly return stayed rare. The metric held, and it identified retention as the binding constraint. L1 metrics covered acquisition, activation, engagement, retention and pipeline quality.

Explored, parked and deferred

Agentic AI

Explored for the pipeline; production cost was too high at this stage.

Event timelines

A backend MVP built with embeddings, a vector database and RAG, which identifies causally connected events and would show users how a story developed. Parked because production cost was not justified at this stage.

Automated evals and fine-tuning

Done, and replaced in production by the weekly human-in-the-loop evaluation because of cost.

OAuth

On the roadmap from the first beta of 10 to 15 users and deferred at every planning cycle, because other items showed higher expected impact. Signup completion held at 80% throughout, so that step was never where users were lost.

B2B intelligence product

Explored and set aside in March 2026, because it needed engineering beyond the no-code stack.

Parked on user request

Show password, remember me, contact us, report a bug, report an error, bookmarks, keeping read Story Cards visible, hyper-personalisation, gamification beyond the progress bar, and a mobile app. The filter system and the Story Card redesign shipped from earlier requests.

My role and the team

Product Manager and Product Builder, working independently from 2025 to 2026. I did the research, problem definition, framework design, pipeline work, frontend integration, analytics, SEO and distribution, and ran every interview.

Stack

WeWeb (no-code frontend, a mobile-first progressive web app), Xano (no-code backend), OpenAI GPT models, Google Maps and Geocoding APIs, PostHog (72+ events), Resend and Brevo for email, NewsAPI AI for news data, a Cloudflare Worker, and an API that fed Story Cards to the X content pipeline.