Keenan PereiraInsights
64 problems from real projects: what broke, what fixed it, and the number that shows it worked.
| Insight | Problem | Fix | Key number |
|---|---|---|---|
| Non-brand clicks are the SEO number that matters | ProblemMost Google clicks came from people already searching “AHA”, which hid real SEO progress. | FixA Search Console scorecard that splits brand from non-brand and tracks rank and blog share. | +183% non-brand clicks |
| A blog hub that builds its own topic pages | ProblemA hand-written list of 11 posts, with no topics and almost no links to sales pages. | FixOne post list and one taxonomy generate the hub, topic pages and schema. | 11 → 48 posts in the Journal |
| A thank-you page that lied | ProblemFailed leads still showed “thank you” and counted as ad conversions. | FixEvery form now confirms the lead saved, with tests that force each failure. | 7 / 7 forms verify the save |
| The first-week SEO audit that read the site like a crawler | ProblemCrawlers saw duplicate titles, no FAQs, no form and facts that disagreed. | FixFix the HTML that ships, then have every build check it before release. | 21 → 0 pages sharing a duplicate title |
| Turning showroom pages into local landing pages | ProblemThin centre pages, one unknown to Google, and three pages chasing one Mumbai query. | FixA real local page per centre, and one page that owns the Mumbai query. | 362 → 999 words, Lower Parel page |
| The slow part was the server, not the page | ProblemPages failed Core Web Vitals for real users while lab tests passed. | FixDo the arithmetic first, fix what the repo controls, and test every page on every device. | 96% of the LCP wait came before first paint |
| Measuring whether AI assistants recommend you | ProblemNobody could say whether AI assistants named AHA when buyers asked. | FixA fixed 12-prompt panel on two engines, then crawler access, llms.txt, one Organization and Bing. | 5 → 7 of 24 panel answers name AHA |
| Where did this lead come from? | ProblemNo website lead in Zoho could be traced to the page it came from. | FixEvery form now sends its page, the visit’s first page and a callback time; 90-day ad tags are going live. | 7 / 7 forms record the lead’s page |
| The sitemap that quietly hid the whole blog | ProblemA hand-kept sitemap drifted, then silently dropped the entire blog. | FixGenerate it from the built pages and blog data, and check every URL on deploy. | 38 → 108 URLs in the sitemap |
| What a fortnight of Google Business Profile posts taught us | ProblemBusiness Profile posting was ad hoc, and one early reading looked like a long tail. | FixA tagged post bank on a fixed cadence, read three times before the plan changed. | +93% → 0% first four posts: ten days, then five |
| The hero image that waited in line | ProblemHero images took 5.4 s on phones, queued behind 23 other images. | FixPreload the hero, lazy-load the rest, and content-hash every asset. | −80% LCP on /solutions |
| Product pages led by specifications | ProblemProduct pages repeated their facts as bullets and had no specifications. | FixOne data file renders each page’s spec table, related products and PDF catalogue. | 2,037 px of bullets, and no spec table |
| Writing about Apple without breaking Apple’s rules | ProblemManagement wanted the site to feel Apple-ish, with none of the trademark risk. | FixEncode Apple’s trademark rules in data, templates and a wording check before writing copy. | 6 of 6 Apple pages indexed within days |
| A path-traced showroom in a TV browser | ProblemThe website concept’s 3D tour of the experience centre was hard to drive in a live demo. | FixPath-trace every room offline, and let the TV only turn the camera and crossfade the light. | 22 renders, path-traced offline in about 2 hours |
| A sales pitch that plays itself | ProblemEvery scene of the showroom pitch waited for someone to press on. | FixSlides that play themselves and pause only on a To remember slide. | 15 → 5½ min walkthrough, with a presenter clicking |
| Product films recorded from the app, with no ffmpeg | ProblemDemo videos were wanted to send, in lengths that kept changing from 20 seconds to 90. | FixRecord the app’s own films in headless Chrome on a virtual clock, and encode them in Chrome. | 87 s lock film, no feature left out |
| The BACK button that kept walking out of the app | ProblemA few BACK presses in a row could leave the showroom app for an old cached page. | FixBank one history step per real button press, and serve every page fresh. | 1 per press history steps banked |
| From sleep loop to portable screensaver | ProblemThe TV’s idle loop was wanted inside another app, without the login, the map or the CRM. | FixBuild a standalone page from the TV’s own code, and stop the build if that code changes shape. | 22 slides a pass on the TV, about 4½ minutes |
| Building for a 2 GB TV and a remote with no pointer | ProblemA map of 500 buildings had to run on a 2 GB TV, driven by a D-pad. | FixA TV mode: zone jumps, overscan calibration, a lighter renderer and a 6-minute story. | −35% browser memory |
| Five bugs a data audit caught before anyone made a decision on them | ProblemDashboards looked plausible, but leads, spend and channels were each off. | FixRecompute every number by a second route before trusting it. | 2.00× → 1.00× Meta lead count |
| The shared token that switched off the website’s leads | ProblemAbout nine systems shared one Zoho refresh token, and any could cancel another’s access. | FixThe BI got its own refresh token and shares one access token an hour. | ~30 min website leads stopped, once |
| A 15-minute warehouse that runs on one PC | ProblemThe BI read a hand-synced copy that reported failed syncs as successes. | FixRead-only pullers feed a private warehouse every 15 minutes, from one PC. | 15 min refresh, while the PC is on |
| One meaning per number, so the screens stop arguing | ProblemTabs and the Ask chat used different rules for the same number. | FixWrite each rule once with its date, and check every copy against it on every change. | 54 metric formulas, one definition each |
| A phone page that tells consultants who to call next | ProblemConsultants needed to know who to call next, not another screen of totals. | FixA read-only phone page per consultant, copied from a whiteboard sketch, with a ranked call list. | 7 days with nothing logged puts a deal on the list |
| A wrong pin is worse than no pin | ProblemCRM building names were messy, and some deals were pinned in the middle of India. | FixMerge spellings by hand, take the pin the data agrees on, and accept new pins only after four checks. | 507 buildings on the map |
| Planning billboards around where the leads are | ProblemA billboard inventory listed sites and rates, but said nothing about where AHA’s leads are. | FixGeocode every site, then rank it by the leads within 2 km. | 93 sites geocoded from a spreadsheet |
| Electrical plans that open fast and stay sharp on a TV | ProblemA plan’s PDF took seconds to open on the TV, and zooming in blurred the drawing. | FixPre-render every page as an image, and redraw from the PDF only when someone zooms. | 1,800 px pre-rendered page, about 200 to 350 KB |
| Giving a TV remote a mouse pointer | ProblemA map wants a mouse, and a TV remote has four arrows and OK. | FixA pointer that eases, curves and settles, with OK and BACK moving one level at a time. | 3× drag threshold for the air mouse |
| A phone as the TV remote, with voice search | ProblemA full-screen map on a TV needed a way to find one building by name from across the room. | FixA phone page sends voice searches over Supabase Realtime, and a fuzzy matcher finds the building. | 4 speech guesses scored per search |
| What buyers really ask on sales calls | ProblemCall transcripts held what buyers ask, in noisy Hinglish nobody could read at scale. | FixHaiku pulls question-and-answer pairs from each call; Sonnet clusters them across calls. | 5 to 6 buyer questions per call |
| Asking AI assistants about AHA without saying AHA | ProblemNobody knew whether AI assistants named AHA when buyers asked them for a company. | FixAsk 31 buyer questions without AHA’s name, score every answer, and brief the gaps. | 31 buyer questions, AHA’s name removed |
| Recovering building names from addresses and plan PDFs | ProblemHundreds of deals had no usable building name, so the map couldn’t group homes by building. | FixMine the address text with a geocoder, then have Claude read the plan PDFs for the rest. | 633 / 711 building names recovered from address text |
| An ads agent that grades its own advice | ProblemAds advice is only useful if someone checks, a week later, whether it worked. | FixA Claude agent suggests at most 8 changes, a person applies them, and the next run grades each one. | ≤ 8 recommendations per weekly run |
| A nightly sync that survives its own failures | ProblemOne flaky API or expired token could hold back the whole day’s data. | FixCore steps stop the run, optional steps only warn, and 28 checks grade every run. | 16 of 25 steps that can fail without stopping the run |
| Mapping a whole market when the API stops at 60 results | ProblemGoogle’s Places search stops at 60 results per query. | FixA quadtree sweep with ID-only requests, then a greedy set cover for details. | 1,176 firms mapped in Mumbai |
| Telling an automation firm from a copier dealer | Problem“Automation” also matches copier dealers, software houses and institutes, and many real firms have vague names. | FixCheap name and category rules first, then score each firm’s own home page. | 27% of the Mumbai list got in on website text |
| Buying place details with a greedy set cover | ProblemPaying for details on every search would have bought most places several times over. | FixSweep for free IDs, then buy the result pages that cover the most places still needed. | 49,695 sweep searches, all the free ID-only kind |
| Thirteen keywords in Mumbai, six for the rest of India | ProblemThirteen keywords at Mumbai’s depth across India meant far more squares and places to pay for. | FixMeasure each keyword’s unique finds in Mumbai, then test cheaper split rules on the saved results. | 6 of 13 keywords kept for India |
| A review sheet instead of a silent filter | ProblemA filter that drops every doubtful result hides its own mistakes. | FixGive doubtful firms their own sheet, and record on every row why it’s there. | 787 doubtful Mumbai firms kept for review |
| Tracking every enquiry without slowing the page down | ProblemTag Manager slows the first screen, and a tap on Submit isn’t a lead. | FixLoad tags on the first tap, queue events in the data layer, and signal success separately. | First tap when tags load, 7 s at most |
| Phone and pincode checks that make every lead CRM-ready | ProblemA lead with a wrong phone number or no last name is useless to sales, or refused by the CRM. | FixThe browser’s own checks, then a small script that maps each answer to the field sales uses. | 7 → 12 answers to CRM fields |
| Designing a landing page for people who just clicked an ad | ProblemAd visitors arrive cold, on phones, for a purchase nobody makes on impulse. | FixProof before the form, a contact route always in reach, and copy in the words people search with. | 12 sections, one form |
| Lead capture without handing the browser your CRM keys | ProblemMy first lead form would have put the CRM keys in the page source. | FixA tiny server endpoint holds the keys; the browser only posts the form. | 0 secrets in the page |
| From 3.3 MB to 180 KB, the image diet behind a fast landing page | ProblemThe ads landing page carried 3.3 MB of photos. | FixA sharp pipeline: WebP, exact display sizes and lazy loading. | −94% image weight |
| Turning “it’s too hot” into an AC command without an LLM | ProblemPeople say “it’s too hot”, not “turn on the bedroom AC”. | FixPattern tables map complaints, comparatives and aliases to device actions. | 10 comfort-complaint patterns |
| Keeping locks and alarms out of “everything” | ProblemTaken literally, “turn off everything” includes the alarm and the locks. | FixThree layers drop security devices, and a guardrail decides what needs a person. | 6 of 8 device types “everything” can touch |
| Console, offline voice and WhatsApp on one command engine | ProblemThree ways to talk to the house could mean three different answers. | FixOne parser, guardrail and device state behind every channel. | 3 channels on one parser and guardrail |
| 48 test commands, each with a reason to exist | ProblemA parser that works on easy sentences says nothing about hard ones. | Fix48 labelled commands, reported by intent, layer and confidence. | 48 test commands, each with a reason |
| Rules first, LLM second | ProblemVague voice commands are slow, and unsafe, to send straight to an LLM. | FixRules answer first, a local LLM handles the rest, and guardrails decide. | 0.9 confidence for the rules to answer |
| The model that writes the CRM briefing never sees a customer | ProblemPosting CRM records to a hosted AI service was the easy route, and the wrong one. | FixGive a local model totals only, and list every place the data does go. | 0 customer records sent to an outside AI service |
| Moving the CRM maths out of Python and into Postgres | ProblemEvery dashboard load pulled rows into Python and counted them in loops. | FixTwelve Postgres functions do the counting and return one JSON object each. | 33 → 13 database calls per dashboard load |
| Fitting a daily CRM briefing into one WhatsApp message | ProblemThe full report was too long and too heavily formatted for a phone. | FixA fixed template in the prompt, and a 1,500-character cap enforced in code. | ≤ 1,500 characters, cut at a line break |
| Don’t let the LLM do the maths | ProblemAn LLM summarising CRM data dropped digits and mislabelled fields. | FixSQL computes every figure; the model only narrates them. | ~255 lines of Python maths moved to SQL |
| A Zoho CRM copy that only fetches what changed | ProblemRe-downloading the whole CRM every day gets slower as it grows, and a fixed field list drops data. | FixFetch only changed records, upsert them by Zoho ID, and keep each full record as JSON. | 4 Zoho modules copied on every run |
| Why FER-2013 is hard, and what 62% really means | ProblemThe faces are tiny, grey, noisily labelled and badly unbalanced. | FixCount every class before training, and read accuracy against simple baselines. | 7,215 vs 436 happy vs disgust training faces |
| A CNN that shrinks a face to 512 numbers | ProblemExpression cues live in a few pixels of a 48 × 48 grey face. | FixFour widening convolution blocks pool the face down to 512 features, with dropout throughout. | 4 convolution blocks |
| When more epochs stopped helping | ProblemTraining accuracy kept rising while validation stalled and validation loss turned upward. | FixWatch validation loss and stop where it bottoms out, around epoch 85. | ~82% training accuracy at epoch 200 |
| From webcam to Flask, one frame a second | ProblemA notebook model can’t see a webcam, and live frames must look like its training faces. | FixThe page posts one JPEG a second to Flask, which repeats the training preprocessing exactly. | 1 s between frames |
| Turning an emotion label into a colour and a tip | ProblemA bare label like “sad” tells a person nothing they can use. | FixMap each emotion to a background colour and a short, kind tip. | 7 emotions, each with a response |
| Filling a customer’s gaps from their own other months | ProblemThousands of blanks and placeholders, in data where every customer appears eight times. | FixFill each blank from the same customer’s other months, never from a column average. | 12,500 customers behind 100,000 rows |
| Cleaning ages of -500 with rules, not formulas | ProblemImpossible values everywhere, from ages of -500 to 1,798 bank accounts. | FixA plain limit per column, set by what each value can be, applied after placeholders become blanks. | -500 to 8,698 ages before the rules |
| Four feature selectors, three signals they all kept | Problem44 candidate columns, most of them month and occupation flags. | FixLet four models each shortlist features, then trust what they agree on. | 44 → 5 features in the final forest |
| Four credit models compared on more than accuracy | ProblemAccuracy alone can’t rank models when half the customers share one band. | FixCompare on one split against a baseline, recall per band and two-band errors. | 76.6% random forest accuracy |
| Split credit data by customer, not by row | ProblemA random row split put the same customers in both the training and test data. | FixSplit by customer, so every test row belongs to someone the model has never seen. | 8 monthly rows per customer |
Nothing matches that.