<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Pranay Suyash on Medium]]></title>
        <description><![CDATA[Stories by Pranay Suyash on Medium]]></description>
        <link>https://medium.com/@pranaysuyash?source=rss-2a54ac04fd72------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZOnyjAbg7ZYvBFBawbBNtQ.png</url>
            <title>Stories by Pranay Suyash on Medium</title>
            <link>https://medium.com/@pranaysuyash?source=rss-2a54ac04fd72------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 18:38:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pranaysuyash/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Five Viewports, Two Missing Controls: A Spatial Editor’s Responsive UX Audit]]></title>
            <link>https://pranaysuyash.medium.com/five-viewports-two-missing-controls-a-spatial-editors-responsive-ux-audit-f5a3967e6144?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/f5a3967e6144</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 16:43:57 GMT</pubDate>
            <atom:updated>2026-10-30T16:43:57.260Z</atom:updated>
            <content:encoded><![CDATA[<p>At 1600px, the spatial editor’s header showed both project and scenario selectors. At 1200px, their labels disappeared, but the selected values remained. At 1024px, the project selector was gone from the header, and by 980px, neither selector was visible there. The canvas still had room, but two controls used to identify and switch the working context had disappeared from the primary navigation.</p><p>The screenshots don’t establish whether those controls became inaccessible. They may have moved into another menu or remained available elsewhere. What they do reveal is exactly where the responsive behaviour needs closer inspection: can someone still tell which project and scenario they’re working in, and switch between them without hunting through the interface?</p><p>This matters more in a spatial editor than it might in a conventional content website. The interface isn’t just presenting information. People are creating layouts, modifying geometry, comparing scenarios, running analyses and inspecting results. The meaning of what appears in the workspace depends on which project, configuration and floor are currently active.</p><p>A responsive header therefore has two responsibilities. It needs to make room for the tools people use, while preserving enough context for them to understand what those tools are operating on.</p><h3>What changed across the five viewport widths</h3><p>I reviewed the same application at 1600, 1200, 1024, 980 and 390 pixels. Rather than evaluating each screenshot independently, I compared which controls remained visible, what information was preserved and how the interface changed as horizontal space became limited.</p><p>At 1600px, the header had enough room for the brand, navigation and context selectors with their labels. The active project and scenario were visible, making it relatively straightforward to understand which workspace was open.</p><p>At 1200px, the application removed the selector labels but retained their values. This is a reasonable compression strategy when the surrounding interface still provides enough context. A recognisable project name may not require the word “Project” beside it, although a scenario called “Default” becomes more ambiguous when the label explaining what it represents disappears.</p><p>The next two transitions were more significant. At 1024px, the project selector was no longer present in the header. At 980px, the scenario selector disappeared as well, and the branding had also collapsed to a compact mark.</p><p>The application wasn’t merely reducing padding or shortening labels anymore. It was changing which pieces of information remained available in the primary navigation.</p><p>At 390px, the application displayed a desktop-experience message instead of attempting to fit the full editing interface onto the phone. That’s a separate product decision and deserves to be evaluated independently from the intermediate desktop widths.</p><p>One detail worth preserving is that the active level was already represented in the bottom status bar and floor map. The application had established another place for users to understand part of their spatial context. Project and scenario selection followed a different pattern, which makes the consistency of the overall navigation worth reviewing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XzNulOHtUp6CvpCM5APGRw.png" /></figure><h3>Workspace context deserves different treatment from ordinary tools</h3><p>A spatial editor might contain dozens of actions: drawing, selecting, moving objects, measuring distances, inspecting properties, changing materials, running simulations and navigating between floors. Not all of these tools need to remain visible simultaneously.</p><p>It makes sense to reduce the prominence of actions that are rarely needed. Some can move into secondary menus, inspector panels or contextual toolbars without significantly affecting how people understand the workspace.</p><p>The active project and scenario serve a different purpose.</p><p>Suppose someone creates two versions of a building layout. One represents the original arrangement, while the other includes changes to internal walls, furniture or ventilation openings. The user runs an analysis against both scenarios to compare the results.</p><p>The geometry and analysis may render correctly in either case. But if the active scenario isn’t clearly identifiable, a user can lose track of which configuration they’re examining. This becomes more concerning when the interface supports several ways of evaluating a scene, since a result may look plausible regardless of whether it belongs to the intended scenario.</p><p>The same issue appears when moving between creation and evaluation. During editing, a user needs to know what they’re modifying. During evaluation, they need to know which configuration produced the result. During verification, they need enough context to determine whether they’re examining the expected state.</p><p>That context shouldn’t depend entirely on memory.</p><p>This doesn’t mean both selectors require full-size dropdowns at every width. Their ability to identify the active workspace is more important than preserving their original appearance.</p><p>A compact header could display the current project and scenario in a single control, while opening that control reveals the individual switching options. The current selection remains visible, and the space required for the full interaction is used only when necessary.</p><p>There is a trade-off. Combining selectors introduces another interaction for someone who changes scenarios frequently. If scenario comparison is a core activity, hiding the switching action behind several steps could be counterproductive.</p><p>The decision should therefore depend on how often users change context, not simply on how many pixels the controls occupy.</p><h3>Three ways to handle the shrinking header</h3><p>The first option is to continue compressing the existing selectors. At wider widths, they retain labels and selected values. As space decreases, labels disappear, padding reduces and long values become truncated.</p><p>This preserves the existing interaction pattern and requires relatively little restructuring. It may be sufficient if the product has a predictable range of project names and the remaining navigation doesn’t compete for too much space.</p><p>The weakness is that truncation eventually makes context difficult to distinguish. Two projects with similar names may appear identical when only their first few characters are visible. A scenario called “Default” may remain visible without telling the user what it is the default of.</p><p>The second option is to introduce a compact workspace-context control. Instead of allocating separate horizontal space to project and scenario selectors, the header displays a concise representation of the current workspace. Opening it reveals the full project and scenario controls, potentially with more room for longer names and descriptions.</p><p>I prefer this direction for the observed 1024px and 980px states because it preserves context without requiring the full selector interfaces to remain permanently open.</p><p>The details would matter. The compact control needs an obvious interactive affordance, a predictable keyboard interaction and clear focus behaviour. The selected context should remain readable when the menu is closed. If someone changes the scenario, the new selection should be visible immediately rather than becoming another hidden piece of application state.</p><p>The third option is to move the selectors into a sidebar or overflow menu. This can free substantial header space, especially for interfaces with many tools, and may work well when project switching is relatively infrequent.</p><p>However, a sidebar is not necessarily a reliable place for persistent context. Spatial editors often allow users to collapse panels to maximise the canvas. If the active scenario disappears with the panel, the interface may become more spacious while also becoming less informative.</p><p>An overflow menu has a similar limitation. It can provide access to controls without occupying much space, but the user may still have to open it simply to discover what is currently selected.</p><p>These options aren’t mutually exclusive. A compact workspace indicator could remain visible while the full controls live in an expanded menu. Lower-frequency utility actions could then move into overflow, preserving more room for the parts of the header that identify the current work.</p><p>The goal isn’t to keep every original control visible. It’s to decide which information needs to survive compression and which interactions can reasonably move elsewhere.</p><h3>The 390px state should be evaluated separately</h3><p>The phone-sized view displayed a message directing users to the desktop experience rather than presenting the complete spatial editor.</p><p>I wouldn’t automatically treat that as a failure.</p><p>A professional spatial application may require precise pointer input, extensive canvas space, multiple inspectors and simultaneous access to information that doesn’t fit comfortably onto a narrow touchscreen. Recreating every desktop capability through smaller controls can produce a technically responsive interface that’s still unpleasant to operate.</p><p>The important question is which mobile tasks the product intends to support.</p><p>Detailed geometry editing may reasonably remain desktop-focused. But someone might still want to open a project on their phone to inspect a scene, review an analysis result, check a status or share a view with another person.</p><p>Those tasks could justify a limited mobile experience without requiring a complete mobile authoring environment.</p><p>For the current product, I would preserve the explicit desktop boundary until those mobile jobs are defined and tested. Simply shrinking the existing interface would introduce a large amount of interaction work without establishing which problems the mobile version is supposed to solve.</p><p>There is also a difference between a phone and a narrow desktop window. A 980px browser on a laptop still has a keyboard, pointer and desktop working environment. It shouldn’t automatically lose important desktop context merely because its viewport approaches a tablet breakpoint.</p><p>That distinction is one reason the intermediate widths deserve closer attention than the 390px screen.</p><h3>How I’d test whether the responsive behaviour actually works</h3><p>The existing screenshots identify where the layout changes. The next useful step is to test the workflow through those transitions.</p><p>I’d use the same project, scenario and level throughout the review. Starting at 1600px, the user would confirm the active context, make a small edit and inspect the result. The viewport would then be reduced through 1200, 1024 and 980 pixels without changing the underlying project state.</p><p>At every width, I’d check whether the current project and scenario remain identifiable, whether switching them is possible through a discoverable route and whether the selected level continues to agree across the canvas, floor map and status bar.</p><p>The most important checks would happen at 1024px and 980px, where the selectors disappear from the header.</p><p>One task could involve switching from the default scenario to a revised layout, viewing an analysis, returning to the original scenario and verifying that the correct configuration is displayed. Another could involve changing floors, modifying geometry and checking that the active level remains consistent wherever it’s represented.</p><p>I’d also repeat the interactions using keyboard navigation. If controls move into a compact menu, they still need a predictable focus order, visible focus indicators and a way to operate them without a mouse.</p><p>Resizing during an active task deserves testing as well. Opening a menu at a wide viewport and then narrowing the browser shouldn’t leave an inaccessible overlay. Changing breakpoints shouldn’t reset the selected scenario or unexpectedly interrupt an editing operation. Browser zoom can also trigger compact layouts, so it belongs in the review rather than being treated as an unrelated accessibility check.</p><p>These tests would establish whether the responsive changes merely rearrange the interface or materially affect someone’s ability to complete the work.</p><p>The current evidence supports a fairly narrow recommendation: preserve an always-visible indication of the active project and scenario at the intermediate desktop widths, potentially through one compact context control. The existing level representation can remain in the bottom bar and floor map rather than being duplicated unnecessarily.</p><p>What remains unverified is whether the project and scenario selectors already have another clearly discoverable route when they leave the header. If that route exists, the main problem may be the loss of visible context. If it doesn’t, the navigation needs to be addressed as well.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f5a3967e6144" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Happens Between a Clinic Booking Request and the Front Desk?]]></title>
            <link>https://pranaysuyash.medium.com/what-happens-between-a-clinic-booking-request-and-the-front-desk-3d4d97a5443b?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d4d97a5443b</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 15:45:26 GMT</pubDate>
            <atom:updated>2026-10-29T16:39:53.584Z</atom:updated>
            <content:encoded><![CDATA[<p>A patient opens a clinic website, chooses an appointment date, selects a time and submits the form. The page displays a success message, and the patient closes the browser assuming the appointment has been arranged.</p><p>Inside the clinic, however, the selected time may still need to be checked. The doctor could already have another appointment, the consultation might require a longer slot, or the clinic may need someone to confirm the request before adding it to the working schedule.</p><p>This creates a problem that is easy to overlook while designing a booking interface. The visitor and the clinic are participating in the same workflow, but they don’t necessarily have the same information or authority.</p><p>A visitor can request a time. The clinic has to decide whether that time can be provided.</p><p>If the application treats those two actions as equivalent, it can create confusion even when every form field and button works correctly.</p><p>I’ve been reviewing this distinction in the context of a visitor-facing website and an operational scheduling system. The interesting problems appear at the point where information moves from one side to the other.</p><h3>A request and a confirmation need different states</h3><p>Appointment interfaces often make date and time selection look like a completed booking. A user sees available-looking slots, chooses one and receives a confirmation-style message.</p><p>That can be appropriate when the software has access to an authoritative schedule and is allowed to reserve the slot immediately.</p><p>But some clinics operate through a request-and-confirmation process. Staff may need to consider the doctor’s schedule, the type of visit, existing commitments and other practical constraints before assigning a time.</p><p>In that situation, the interface should describe what actually happened. “Request received” is accurate when a request has been saved. “Appointment confirmed” should be reserved for the point where the clinic has accepted the arrangement.</p><p>The difference also affects what the user needs next. Someone who has submitted a request may need to know how confirmation will arrive, when they should expect a response and what to do if they need an urgent appointment. Someone with a confirmed visit needs the final time, location and instructions relevant to attending.</p><p>These are different communication requirements.</p><p>A well-designed workflow should preserve the distinction in its underlying data, not merely adjust the wording of the success screen.</p><p>A booking request may move through states such as received, pending review, confirmed, rescheduled, cancelled or declined. A confirmed appointment can then have its own visit states, such as arrived, in consultation, completed or no-show.</p><p>The exact terminology can vary, but the events should correspond to things that have actually happened.</p><p>A request should not become a confirmed appointment merely because its form submission succeeded.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*69NKo_CtW5Tkm9jmtj78ow.png" /></figure><h3>Booking for someone else changes who needs to be contacted</h3><p>A surprisingly important edge case is arranging an appointment for another person.</p><p>This happens frequently in ordinary healthcare settings. Someone may arrange a visit for a parent, child, spouse or another family member. A caregiver may handle the scheduling while the patient is the person who will actually attend.</p><p>If the form contains only one name and one phone number, the system may not know whether those details belong to the patient or the person arranging the appointment.</p><p>That becomes a problem when the clinic needs to call back, confirm a changed time or clarify the booking.</p><p>The solution isn’t necessarily to collect more information indiscriminately. It is to understand which information is required and whose contact details should be used for which purpose.</p><p>A booking workflow can offer a clear option for arranging a visit for somebody else. The patient and arranging person’s identities can remain distinct, with an explicit contact preference for scheduling updates.</p><p>That also helps avoid sending unnecessary personal information to someone who shouldn’t receive it.</p><p>For clinics that don’t need medical intake before a consultation, the public booking form shouldn’t turn into a long questionnaire about symptoms or personal medical history. The initial job is arranging a visit, and the data collected should reflect that limited purpose.</p><p>Keeping the request small can also reduce friction on mobile, where complicated forms are harder to complete.</p><h3>The schedule has to be checked against reality</h3><p>Date selection is another area where a visually convincing interface can hide operational problems.</p><p>Suppose two people request the same time within a few seconds of each other. Both browsers may have loaded the list of available options before either request was submitted.</p><p>If the application simply trusts the displayed availability, it may allow both requests to appear confirmed.</p><p>A system that reserves slots automatically needs an authoritative capacity check at the point of reservation. A request-based system can accept both preferences but must still distinguish them from confirmed capacity.</p><p>The calculation may also depend on more than a fixed number of appointments per hour.</p><p>Different visit types can require different durations. The doctor may have a break, an emergency, an unavailable period or an appointment that runs longer than expected. The clinic may also operate with buffers or limits on how many requests staff can manage during a particular period.</p><p>Those rules should be defined in the scheduling system rather than inferred by a decorative calendar in the visitor interface.</p><p>The user also needs a sensible recovery path. If the preferred slot is no longer available, the clinic should be able to propose another time without losing the original request or asking the patient to fill in everything again.</p><p>A change in the appointment time should remain visible as a change, particularly when it affects what the patient has already been told.</p><h3>The staff queue is part of the product</h3><p>A visitor-facing booking flow can be beautifully designed while the front desk struggles to process the requests it produces.</p><p>The administrative side needs to show which requests are waiting, which ones have been assigned, which ones require a callback and which ones have already been confirmed.</p><p>Ordering matters. If a new request arrives while staff are working through the queue, the application needs a clear rule for where that request appears. Filters, pagination and result limits must not silently hide older requests.</p><p>A technical problem can arise when software retrieves only a limited subset of requests and then sorts that subset in the browser. The list can appear correctly ordered while excluding older records that were never retrieved.</p><p>For a waiting queue, sorting and filtering should happen before the result limit, using a stable ordering rule so records don’t unexpectedly move between pages.</p><p>There are also practical interface questions. Can staff find a booking by patient name or phone number? Can they distinguish a pending request from a confirmed appointment without opening every record? Can they see the preferred time and the actual assigned time separately? Can they recover from an accidental status change?</p><p>These details may be less visually exciting than a visitor-facing calendar, but they affect whether the booking system can be used throughout a working day.</p><p>The administrative interface should be evaluated using the tasks staff actually perform rather than by counting how many fields or dashboard features it contains.</p><h3>Notifications are another place where requests can get lost</h3><p>A saved appointment request and a successfully delivered notification are separate events.</p><p>The application may save the request correctly but fail to send an email or message to the staff. A notification provider can become unavailable, reject a message or experience a temporary network problem.</p><p>If the workflow depends entirely on an email reaching someone’s inbox, the request may remain unhandled even though the website reported success.</p><p>A more reliable design preserves the request in the clinic’s own system and treats notifications as a separate delivery process.</p><p>Staff should be able to find pending requests without needing the corresponding notification. A failed notification can be retried, and the request remains available until somebody deals with it.</p><p>This is an established distributed-systems problem. The <a href="https://docs.aws.amazon.com/prescriptive-guidance/latest/cloud-design-patterns/transactional-outbox.html">transactional outbox pattern</a> addresses the difficulty of keeping a database change and a downstream notification consistent.</p><p>Not every small clinic needs a complicated messaging architecture, but the underlying principle is useful: the request should survive even when the mechanism used to alert somebody about it fails.</p><p>The same distinction should exist for patient-facing communication. Sending a confirmation message does not necessarily establish that it was delivered, read or understood. The system should avoid silently treating all those states as equivalent.</p><h3>Mobile and language support need to be tested as workflows</h3><p>In India, a clinic website may need to work for visitors who are more comfortable reading Hindi or another regional language, as well as people using English.</p><p>That involves more than translating headings.</p><p>Field labels, form instructions, date selection, error messages and confirmation states all need to remain understandable. A layout that works with short English labels may become cramped once the text expands. A message that is technically correct but uses unfamiliar terminology can still leave the user unsure whether the appointment is confirmed.</p><p>Phone-based booking also introduces practical constraints. The keyboard can cover fields and buttons, date pickers may behave differently from desktop controls, and validation errors can be difficult to locate when the form is long.</p><p>A useful test involves completing the full request using a small phone, deliberately making a mistake, correcting it and checking that the final request contains the intended information.</p><p>I’d also test the booking-for-someone-else flow in the same conditions, because distinguishing the patient’s details from the arranging person’s details can become confusing when the screen is crowded.</p><p>Time handling deserves attention as well. A booking interface should use the clinic’s local calendar consistently, refresh the applicable date when necessary and avoid presenting yesterday as the default date merely because a page remained open across midnight.</p><p>These are ordinary cases, but they reveal problems that static screenshots rarely show.</p><h3>Measuring the complete booking process</h3><p>The obvious metric for a booking page is how many visitors submit a request. That is useful, but it only measures the first part of the workflow.</p><p>For a request-based clinic system, I’d also want to know how many submitted requests are received by the operational system, how long they remain pending, how many are confirmed, how often the preferred time changes and whether staff can identify requests requiring attention.</p><p>The goal isn’t to turn the clinic into a dashboard full of performance indicators. The measurements should help identify practical problems.</p><p>A large number of submissions with a long confirmation delay may indicate a staffing or scheduling bottleneck rather than an issue with the public website. Requests that remain unhandled because notifications failed point to a different problem. Frequent rescheduling may suggest that the available-looking slots don’t reflect actual capacity.</p><p>These issues require different fixes, so grouping them into one booking conversion figure would make diagnosis harder.</p><p>The same applies to usability testing. I’d evaluate the visitor and staff journeys together, using the same booking case. The visitor submits a request, staff locate it, check the schedule, propose or confirm a time, and the appropriate person receives the update.</p><p>That test should also include failures: a preferred slot becoming unavailable, someone booking for a family member, an unreachable contact, a duplicate request and a notification that doesn’t arrive.</p><p>The immediate behaviour to check is straightforward. A submitted request remains recoverable, the front desk can see and act on it, and the visitor is never shown a confirmed appointment until the clinic has actually assigned one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d4d97a5443b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When a Corrected Transcript Leaves an AI Review Out of Date]]></title>
            <link>https://pranaysuyash.medium.com/when-a-corrected-transcript-leaves-an-ai-review-out-of-date-1d1b1957acf2?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d1b1957acf2</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 15:43:59 GMT</pubDate>
            <atom:updated>2026-10-29T15:43:59.961Z</atom:updated>
            <content:encoded><![CDATA[<p>A single word can completely change the meaning of a recorded sales conversation. Compare “I can waive that fee” with “I can’t waive that fee.” The sentences are nearly identical, but the first could be an unauthorised promise while the second might be a salesperson correctly explaining a policy restriction.</p><p>Suppose an automated call-review system transcribes the first version and flags it as a policy violation. A manager listens to the audio, notices that the transcription missed the negation and corrects the sentence. The words on the screen now match what was spoken.</p><p>The problem appears when the rest of the application still contains findings, quotations and timestamps produced from the earlier transcript.</p><p>While reviewing a post-call assessment workflow, I found that handling transcript corrections involves more than updating the displayed text. The system needs to understand which downstream decisions depend on the original wording, which evidence remains valid and what should happen to conclusions that somebody has already reviewed.</p><p>Some of that work is straightforward when the entire assessment can simply be run again. It becomes harder when the application also needs to preserve the original machine output, retain human review decisions and explain why a previously accepted finding has changed.</p><h3>The transcript and its timestamps are separate pieces of information</h3><p>Speech-recognition systems don’t necessarily produce the transcript and its word-level timing information through the same process.</p><p>For example, <a href="https://github.com/m-bain/whisperX">WhisperX</a> combines speech recognition with an alignment stage that estimates where recognised words occur in the audio. The resulting timestamps make it possible to click a word or sentence and jump to the relevant part of the recording.</p><p>That is particularly useful for evidence review. A manager examining a flagged statement can listen to the original words rather than relying entirely on the model’s interpretation.</p><p>But those timestamps correspond to the text that was aligned at the time.</p><p>If someone later corrects a sentence, the text editor can replace the words immediately. That does not automatically mean the alignment has been recalculated against the audio. A quotation might display the corrected sentence while its individual word timing still comes from the earlier transcription.</p><p>The distinction matters most when the correction changes the number or sequence of words. Adding a missing negation, deleting a repeated word or splitting a phrase can make the original token-to-time mapping unreliable.</p><p>A sensible approach is to preserve the original alignment, associate it with the transcript revision that produced it and mark affected timing references for review when the text changes.</p><p>The system may be able to realign only the relevant audio segment. In other cases, it may be safer to fall back to a broader sentence-level timestamp until the corrected words have been verified.</p><p>Simply copying the old timestamps onto new words is convenient, but it creates an impression of precision the system hasn’t actually established.</p><h3>Existing findings need to know which transcript produced them</h3><p>Now consider the assessment itself.</p><p>A call-review application might evaluate a conversation against several approved company policies. One requirement could concern discount approvals, another product claims, another objection handling, and another the next steps agreed with the customer.</p><p>For every finding, the system may store a verdict, explanatory text, supporting transcript passages and a reference to the applicable policy.</p><p>Those records are only meaningful if their inputs remain identifiable.</p><p>If the transcript changes, the application needs to know which findings were produced from the previous version. A missing negation might invalidate the discount-policy finding while leaving unrelated portions of the review unaffected.</p><p>One implementation option is to reassess the entire call after every meaningful transcript correction. That is relatively simple to reason about because all findings are produced from a common source version. It may also be expensive or unnecessary for minor edits.</p><p>Another approach is to track dependencies more precisely. Each finding records the transcript revision, relevant evidence spans and policy revision used to produce it. When a correction touches those spans, the affected finding is marked stale and sent for reassessment.</p><p>The second approach can reduce unnecessary processing, but it introduces a harder engineering problem: determining which dependencies a particular correction affects. If the dependency mapping is incomplete, the system may leave an invalid finding looking current.</p><p>For an early product, I would favour a conservative reassessment policy over a highly optimised one that is difficult to verify.</p><p>The scope can become more selective once the system has sufficient tests showing that unaffected findings really remain unaffected.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oBW9_aLlb3y9oHkVRQhsFA.png" /></figure><h3>Human review adds another layer of responsibility</h3><p>The problem becomes more complicated after somebody has accepted or rejected a finding.</p><p>Suppose a manager accepts the original assessment that the salesperson made an unauthorised discount promise. That decision is saved as part of the review record.</p><p>Later, the manager or another authorised reviewer corrects the transcript. The relevant sentence now contains the missing negation, and reassessment produces a different conclusion.</p><p>What should happen to the original accepted decision?</p><p>Overwriting it would make the current screen look cleaner, but it would lose important history. The manager really did accept the earlier finding, based on the evidence available at that time.</p><p>Keeping the accepted decision permanently active would create a different problem. The current review would continue reporting a conclusion that no longer matches the corrected transcript.</p><p>The application needs to preserve both facts: what was decided earlier and which decision applies to the current source revision.</p><p>That suggests separating immutable reviewed records from the current working assessment. A completed review can retain the transcript and policy revisions it was based on. A later correction creates a new revision and marks any dependent findings as requiring reassessment.</p><p>Once the new findings have been examined, a new reviewed revision can be saved without altering the earlier one.</p><p>This is useful for quality and compliance workflows because it makes disagreements reconstructable. Somebody investigating the review later can see the original machine transcript, the correction, the previous finding, the revised finding and the human decisions involved.</p><p>The history should also distinguish a model changing its recommendation from a human overturning that recommendation. Those are different events, even when they produce the same final outcome.</p><h3>Policy changes can invalidate findings too</h3><p>Transcript corrections are only one source of change.</p><p>Imagine that a company revises its approved sales policy. A discount that previously required manager approval may now be allowed within a defined range, or a formerly acceptable claim may become prohibited.</p><p>Even if the transcript hasn’t changed, assessments produced using the previous policy may no longer describe the result under the new rules.</p><p>That doesn’t necessarily mean every historical review should be rewritten. Historical assessments may need to remain tied to the policy that was in force when they were completed.</p><p>But the application should be able to distinguish historical compliance from a new assessment under the current policy.</p><p>This is another reason to preserve policy revisions explicitly instead of retrieving whatever the latest policy text happens to be whenever someone opens an old finding.</p><p>A useful review record needs more than an assessment date. It should identify the source recording, transcript revision, policy revision, assessment version and the human review outcome.</p><p>That combination makes it possible to determine exactly which inputs produced a particular decision.</p><h3>Testing what becomes stale</h3><p>A meaningful test suite for this workflow should include corrections that actually change assessment results.</p><p>Start with a transcript where the machine recognises “I can waive the fee,” and the policy requires prior approval. The system generates a finding and attaches the supporting quote. A reviewer then changes the sentence to “I can’t waive the fee.”</p><p>The expected behaviour should be explicit. The corrected transcript becomes a new revision, the previous assessment remains available as history, the affected current finding becomes stale, and the review cannot be treated as final until the necessary reassessment is completed.</p><p>The timing reference also needs attention. If the application cannot establish word-level alignment for the edited phrase, it should not continue representing those timings as verified.</p><p>Other tests should examine different kinds of edits.</p><p>Changing punctuation without altering the meaning may not require every finding to be regenerated. Moving a statement from the customer to the salesperson can change who is responsible for a commitment. Correcting a product name may affect an approved-claims assessment. Removing a repeated sentence can shift the positions of later quotations.</p><p>Concurrent edits also deserve testing. If one reviewer changes the transcript while another is accepting an older finding, the application needs to detect that their actions were based on different source revisions.</p><p>Without a version check, it is possible to save a human approval against evidence that has already changed underneath the reviewer.</p><p>These cases are easier to reason about when the system treats corrections, reassessments and final reviews as explicit state transitions rather than ordinary updates to a mutable document.</p><h3>The reviewer interface has to expose the consequences</h3><p>None of this is useful if the interface hides the relationship between corrections and assessments.</p><p>A manager correcting a transcript needs to know whether the change affects existing findings. If a previous verdict is now stale, that state should be visible beside the relevant evidence rather than buried in a technical log.</p><p>The interface should also make the original audio easy to inspect. The corrected transcript is a human interpretation of what was heard, and the recording remains the underlying source.</p><p>For an important correction, I would want the reviewer to see the changed text, the affected findings and the status of any required reassessment. Once the new result is available, the previous and revised decisions should be easy to compare.</p><p>A useful design also allows the reviewer to disagree with the reassessment. Correcting the source does not mean a model automatically gets the final decision right.</p><p>The model can propose findings and identify evidence. The human reviewer still needs an intelligible way to accept, reject or mark the result as not assessable.</p><p>This is especially important when the evidence itself is ambiguous. A poorly recorded conversation may not contain enough information to determine whether a commitment was made. The application should preserve that uncertainty rather than force every policy question into a positive or negative verdict.</p><p>The implementation doesn’t need to start with complicated automation. A reliable versioned transcript, explicit stale states, conservative reassessment and preserved review history would already address much of the problem.</p><p>The remaining work is largely about testing these transitions. A corrected word should change the appropriate evidence and findings, while unrelated findings should remain intact only when the system can establish that they are unaffected.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d1b1957acf2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Banking Calls, Hinglish Tutorials, and the Gap in Sales-Call Evaluation]]></title>
            <link>https://pranaysuyash.medium.com/banking-calls-hinglish-tutorials-and-the-gap-in-sales-call-evaluation-0f286d0da53f?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f286d0da53f</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 19:03:19 GMT</pubDate>
            <atom:updated>2026-10-27T19:03:19.351Z</atom:updated>
            <content:encoded><![CDATA[<p>Finding speech data for an AI application seems straightforward until you specify what the application is actually supposed to do.</p><p>Suppose the goal is to evaluate recorded sales conversations in English, Hindi and Hinglish. The system needs to understand what the salesperson said, which commitments were made, whether particular policies were followed and what evidence supports each assessment. It also needs to distinguish the salesperson’s actions from the customer’s statements.</p><p>A collection of transcribed speech should provide a starting point. There are publicly available speech datasets with audio, transcripts and established licences, including resources containing real human voices rather than entirely generated conversations.</p><p>But having speech data and having the right evaluation data are very different things.</p><p>Two datasets illustrate the difficulty particularly well. HarperValleyBank provides human-spoken banking conversations with transcripts and separate audio tracks for the participants. OpenSLR 104 contains Hindi-English code-switched speech with transcripts. Both are useful resources, and both can contribute to an evaluation workflow.</p><p>Neither, by itself, establishes whether a system can accurately assess complete sales conversations across English, Hindi and Hinglish.</p><p>The difference becomes important once we stop measuring transcription alone and start evaluating the decisions made from the transcript.</p><h3>Two useful datasets with very different limitations</h3><p><a href="https://github.com/cricketclub/gridspace-stanford-harper-valley">HarperValleyBank</a> is an interesting starting point because its conversations involve human speakers playing agent and customer roles. Each conversation has separate caller and agent audio files, alongside transcripts and metadata.</p><p>The repository is distributed under CC BY 4.0, making it more accessible for reuse than many private call-centre recordings, subject to the applicable licence terms. It also distinguishes human-corrected transcripts from machine-generated information such as emotion and dialogue-act predictions.</p><p>That distinction matters. A human-corrected transcript is useful as a speech-recognition reference. A model-generated emotion label should not automatically become ground truth for evaluating another emotion classifier.</p><p>The larger limitation is the domain. HarperValleyBank contains simulated banking calls, not genuine sales negotiations. The speakers are human, but the conversations follow role-play scenarios designed for a particular research setting.</p><p>That makes the dataset suitable for examining aspects of conversational speech recognition, turn-taking and speaker-related processing. It does not establish how an assessment system performs when salespeople negotiate prices, discuss discounts, handle objections or make commercial commitments.</p><p><a href="https://openslr.org/104/">OpenSLR 104</a> provides another useful but incomplete piece. Its Hindi-English material is distributed under CC BY-SA 4.0 and comes from spoken technical tutorials. The code-switching is related largely to the subject matter of those lectures.</p><p>That can be valuable when testing speech recognition across Hindi and English vocabulary. But a person explaining a technical topic behaves differently from two people negotiating a sale.</p><p>The vocabulary differs, but so do interruption patterns, conversational goals, speaking styles and the consequences of misunderstanding particular words.</p><p>The licences also need to be considered individually. CC BY and CC BY-SA both permit commercial reuse under their conditions, but attribution is required, and ShareAlike imposes additional requirements when adapted material is shared. Those obligations should be checked against the intended distribution and use rather than reduced to a simple commercial-use checkbox.</p><p>More fundamentally, a dataset’s licence tells us something about permitted reuse. It does not tell us whether the data represents the conversations we need to evaluate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E2ENDSuuQfrdPS2xwmiEuw.png" /></figure><h3>A transcription can be mostly correct and still change the outcome</h3><p>Imagine a salesperson telling a customer, “I can’t waive that fee without approval.”</p><p>The speech-recognition system transcribes it as “I can waive that fee without approval.”</p><p>Most of the sentence is correct. The product, fee and approval context have survived transcription. But one missing negation changes what the salesperson appears to have promised.</p><p>For a generic transcription benchmark, the difference contributes to a word-level error count. For a policy-assessment system, it may change the final verdict.</p><p>Speaker attribution creates another category of failure.</p><p>Suppose the customer says, “I’ll send the documents tomorrow.” If that statement is incorrectly assigned to the salesperson, the assessment system may record an agent commitment that was never made.</p><p>These errors are particularly troublesome because the resulting transcript can remain perfectly plausible. A language model reading it has no obvious reason to question the sentence unless it has access to conflicting evidence.</p><p>This is one limitation of relying entirely on word error rate, or WER, to evaluate speech recognition.</p><p>WER measures substitutions, deletions and insertions relative to a reference transcript. It is useful for understanding transcription performance, but different errors can have very different consequences for downstream tasks.</p><p>Research has examined this problem directly. The paper <a href="https://aclanthology.org/2023.bionlp-1.6/">Evaluating and Improving Automatic Speech Recognition using Severity</a> discusses how ordinary word-level error measures can miss differences that matter to applications consuming the transcript.</p><p>For a sales-call evaluator, the important errors include changed quantities, lost negations, confused speakers, misheard dates and altered commitments. A single mistake in one of those categories may matter more than several harmless transcription differences elsewhere.</p><p>That changes what should be tested.</p><h3>The assessment needs two evaluation paths</h3><p>A useful starting point is to separate the assessment system from the speech-recognition system.</p><p>First, take a conversation with a trusted reference transcript. Preserve the speaker identities and annotate the statements relevant to the assessment. Run the policy-evaluation system against that reference.</p><p>This establishes how well the assessment logic performs when it receives the intended words.</p><p>For example, can it identify a salesperson promising a discount that requires approval? Can it distinguish a customer request from an agent commitment? Does it quote the correct portion of the conversation when explaining a decision?</p><p>Next, run the same audio through the intended speech-recognition pipeline. Feed the resulting transcript into the same assessment system, using the same policies and scoring criteria.</p><p>Now the two results can be compared.</p><p>If the assessment is correct on the reference transcript but wrong on the ASR transcript, transcription or speaker attribution may have introduced a consequential difference.</p><p>If the assessment is wrong on both, the problem may sit in the evaluation logic, the policy interpretation or the quality of the reference labels.</p><p>If the final verdict remains correct despite transcription errors, those errors may be less important for that particular assessment, although they may still matter for other uses of the transcript.</p><p>This separation makes diagnosis considerably easier. Without it, a single final accuracy score hides the reason an assessment succeeded or failed.</p><p>I’d also require every assessment to identify the evidence supporting it. A verdict without the relevant statement is difficult to review, especially when several participants discuss similar subjects across a long call.</p><p>That evidence should preserve the speaker, relevant words and position in the conversation. It should also distinguish an explicit statement from an inference made by the assessment model.</p><h3>Complete conversations matter</h3><p>Speech datasets are often divided into short segments because those are convenient for transcription training and evaluation. A business assessment may need much more context.</p><p>Consider a salesperson who initially says a discount isn’t possible, later offers to request approval, and eventually confirms that approval has been granted.</p><p>Evaluating only the first statement would produce an incomplete view of the interaction.</p><p>Another conversation might include a tentative promise that is corrected several minutes later. An evaluator needs to recognise the correction rather than count the original statement and the final statement as independent commitments.</p><p>This is why the intended evaluation unit should be specified carefully.</p><p>If the product assesses complete calls, the dataset should contain complete or sufficiently coherent conversations. Isolated utterances can be useful for testing particular recognition errors, but they cannot establish whether the system follows commitments, corrections and changes of position across the conversation.</p><p>The reference labels need to reflect that context as well.</p><p>For each relevant policy question, a reviewer should be able to record the expected decision, supporting evidence, speaker responsible and any contextual information necessary to interpret the statement. Disagreements between reviewers should be preserved and resolved deliberately instead of being hidden inside a single label.</p><p>Evaluation examples should also be withheld from the implementation and prompt-tuning process. Otherwise, repeated adjustments can gradually optimise the system for familiar scenarios without improving its ability to handle unfamiliar calls.</p><h3>Where original role-plays fit</h3><p>There is a practical problem with relying only on existing human sales-call recordings: rights-compatible audio with complete transcripts, suitable language coverage and relevant commercial scenarios is not always available for the specific evaluation being designed.</p><p>A promising-looking dataset may contain only transcripts. Another may have audio but no clear licence covering the intended commercial use. A third may contain useful conversations while lacking Hindi-English code-switching or the policy situations being assessed.</p><p>One possible fallback is to create original sales role-plays.</p><p>The scenarios can be designed around known failure conditions: discount approvals, payment commitments, customer objections, follow-up promises, changed dates, Hindi negation and mixed-language conversations.</p><p>Speakers can be recorded separately, or suitable synthetic voices can be used where their usage rights permit it. Separate audio channels make it easier to preserve speaker identity before mixing the conversation into a realistic call format.</p><p>The reference transcript can be frozen before the speech-recognition system processes the audio.</p><p>This creates a useful controlled test environment. The intended words, speakers, actions and policy outcomes are known. If the ASR system changes a negation or the assessor assigns a commitment to the wrong speaker, the failure can be reproduced and examined.</p><p>But synthetic conversations have limitations that should influence how their results are interpreted.</p><p>Generated voices may be cleaner than ordinary telephone speech. They may contain fewer interruptions, overlapping turns, hesitations and unexpected phrases. The scenarios are created by people who already know what they intend to test, which can leave out behaviours that occur naturally in real sales conversations.</p><p>A system could perform very well on this suite and still struggle with genuine calls.</p><p>Original role-plays are therefore useful for controlled stress tests and regression testing. They should not be used as evidence of actual customer-call accuracy without further evaluation on relevant human conversations.</p><h3>What I would report</h3><p>I would keep several results separate rather than compress them into one overall accuracy figure.</p><p>The first is assessment accuracy against reference transcripts. This measures the system’s ability to interpret known statements under the chosen policy framework.</p><p>The second concerns transcription performance, including the frequency and severity of errors that affect the intended assessment. Ordinary WER remains useful, but it should be supplemented with targeted checks for negation, numbers, speaker attribution and other consequential information.</p><p>The third is the final assessment performance when the system operates on ASR transcripts. Comparing that result with the reference-transcript assessment shows which failures are associated with the speech-processing stage.</p><p>Those results should also be broken down by language and scenario type. A system may handle English conversations comfortably while struggling with natural Hindi-English switching. It may recognise pricing statements correctly but confuse follow-up commitments, or perform well when speakers take turns cleanly while failing during interruptions.</p><p>For every incorrect assessment, the report should retain the relevant audio segment, reference transcript, ASR output, expected policy decision and generated supporting evidence. That record makes it possible to inspect failures and compare later system changes without reconstructing the entire experiment.</p><p>Until a representative set of rights-cleared human sales conversations is available, the controlled evaluation can establish whether specific failure cases are handled. Real-call performance remains a separate result that still needs measurement.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0f286d0da53f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why a Container Tracker Doesn’t Automatically Become a Warehouse Platform]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/why-a-container-tracker-doesnt-automatically-become-a-warehouse-platform-3846a59d1e2d?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1600/1*-ikfSfAYdaYZs15xCvtjpA.png" width="1600"></a></p><p class="medium-feed-snippet">An outdoor container yard and a warehouse have something obvious in common: both need to know where things are. There are physical&#x2026;</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/why-a-container-tracker-doesnt-automatically-become-a-warehouse-platform-3846a59d1e2d?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/why-a-container-tracker-doesnt-automatically-become-a-warehouse-platform-3846a59d1e2d?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/3846a59d1e2d</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 16:33:20 GMT</pubDate>
            <atom:updated>2026-10-26T16:33:20.764Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Animating a Flat AI Character in EffectCraft: What Worked and What Broke]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/animating-a-flat-ai-character-in-effectcraft-what-worked-and-what-broke-d85a09ed47b1?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1850/1*Zltz5rQeahof3TR2icQwyw.png" width="1850"></a></p><p class="medium-feed-snippet">An AI-generated character can look ready for animation long before its underlying artwork is actually usable for it. The image might have&#x2026;</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/animating-a-flat-ai-character-in-effectcraft-what-worked-and-what-broke-d85a09ed47b1?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/animating-a-flat-ai-character-in-effectcraft-what-worked-and-what-broke-d85a09ed47b1?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/d85a09ed47b1</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 12:35:32 GMT</pubDate>
            <atom:updated>2026-10-26T12:35:32.525Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[An Editable Blender File Isn’t Necessarily a Customizable Product]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/an-editable-blender-file-isnt-necessarily-a-customizable-product-023bc79eca6a?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1600/1*C9cMx0KndPORGrh6VWBRmQ.png" width="1600"></a></p><p class="medium-feed-snippet">Buying a 3D model sounds straightforward. You look at the renders, check the file formats, purchase the asset and download it. If the&#x2026;</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/an-editable-blender-file-isnt-necessarily-a-customizable-product-023bc79eca6a?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/an-editable-blender-file-isnt-necessarily-a-customizable-product-023bc79eca6a?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/023bc79eca6a</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 08:40:51 GMT</pubDate>
            <atom:updated>2026-10-24T08:40:51.494Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why Background Removal Can Leave a White Halo: An OpenCV and Python Walkthrough]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/why-background-removal-can-leave-a-white-halo-an-opencv-and-python-walkthrough-f098ce518c66?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1680/1*wX6hm2hOoFryzepjE_ZXRQ.png" width="1680"></a></p><p class="medium-feed-snippet">Removing the background from an image seems straightforward until you place the result against a different colour. A cutout that looks&#x2026;</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/why-background-removal-can-leave-a-white-halo-an-opencv-and-python-walkthrough-f098ce518c66?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/why-background-removal-can-leave-a-white-halo-an-opencv-and-python-walkthrough-f098ce518c66?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/f098ce518c66</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 08:34:05 GMT</pubDate>
            <atom:updated>2026-10-24T08:34:05.344Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Boxes Fit. That Doesn’t Mean the Load Is Ready.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/the-boxes-fit-that-doesnt-mean-the-load-is-ready-0373106b4f3a?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1600/1*NfMJMc6GkmiVqZ5peXpqEg.png" width="1600"></a></p><p class="medium-feed-snippet">A truck appears on the screen as a large rectangle. Boxes of different sizes are arranged inside it, and a planning algorithm finds a&#x2026;</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/the-boxes-fit-that-doesnt-mean-the-load-is-ready-0373106b4f3a?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/the-boxes-fit-that-doesnt-mean-the-load-is-ready-0373106b4f3a?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/0373106b4f3a</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 11:28:36 GMT</pubDate>
            <atom:updated>2026-10-22T11:28:36.215Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Someone Already Built It. Should You Still?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pranaysuyash.medium.com/someone-already-built-it-should-you-still-d6489884b55a?source=rss-2a54ac04fd72------2"><img src="https://cdn-images-1.medium.com/max/1600/1*2DxbjwoXshu8s9cVk6ulVw.png" width="1600"></a></p><p class="medium-feed-snippet">Finding an existing product or research paper that does something similar to your idea can be discouraging.</p><p class="medium-feed-link"><a href="https://pranaysuyash.medium.com/someone-already-built-it-should-you-still-d6489884b55a?source=rss-2a54ac04fd72------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://pranaysuyash.medium.com/someone-already-built-it-should-you-still-d6489884b55a?source=rss-2a54ac04fd72------2</link>
            <guid isPermaLink="false">https://medium.com/p/d6489884b55a</guid>
            <dc:creator><![CDATA[Pranay Suyash]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 17:52:47 GMT</pubDate>
            <atom:updated>2026-10-21T17:52:47.153Z</atom:updated>
        </item>
    </channel>
</rss>