[{"data":1,"prerenderedAt":953},["ShallowReactive",2],{"cv-profile":3,"cv-experience":49,"cv-projects":141,"cv-skills":785},{"id":4,"title":5,"availability":6,"body":7,"description":13,"email":24,"extension":25,"links":26,"location":39,"meta":40,"name":41,"navigation":42,"path":43,"role":44,"seo":45,"stem":46,"tagline":47,"__hash__":48},"profile\u002Fprofile.md","Profile","Open to AI engineering roles, remote and full time",{"type":8,"value":9,"toc":20},"minimark",[10,14,17],[11,12,13],"p",{},"I am an AI engineer who likes the unglamorous half of machine learning: getting a model out of the\nnotebook, putting it behind a real endpoint, watching what it does to real users, and then going\nback to improve it.",[11,15,16],{},"Four years in, most of that work sits where ML meets web: image generation for batik patterns,\ncomputer vision on satellite imagery, face embeddings for authentication, an autonomous agent that\nsits on top of clinical paperwork, and the Nuxt\u002FGo\u002FPostgres plumbing that makes all of it usable.\nRight now that means leading development at RAINS Corp and building the e-commerce and AI stack at\nMayfeyr, a multi-brand luxury fashion retailer whose catalogue pipeline I designed and still own.",[11,18,19],{},"The research half is not a separate track. My GenBatik work is a first-author paper in IEEE Access\n(Q1), the framework holds registered IP, and I review for the journal now. I also write about the\nengineering side at invasikode.com, because explaining a system is a good test of whether I actually\nunderstand it.",{"title":21,"searchDepth":22,"depth":22,"links":23},"",2,[],"daffaizzuddin.w@gmail.com","md",[27,30,33,36],{"label":28,"href":29},"GitHub","https:\u002F\u002Fgithub.com\u002Fizzudd",{"label":31,"href":32},"LinkedIn","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fdaffa-izzuddin\u002F",{"label":34,"href":35},"Writing","https:\u002F\u002Fwww.invasikode.com",{"label":37,"href":38},"CV","\u002Fcv","Malang, Indonesia (UTC+7)",{},"Daffa Izzuddin",true,"\u002Fprofile","AI Engineer",{"description":13},"profile","I take models from a training run to something people can actually use: train, serve, monitor, iterate.","mW8qmETNc8WFdEWUOAaEjN5Mb-oO_O0sUksxpr6vuss",[50,86,116],{"id":51,"title":52,"body":53,"company":60,"companyUrl":61,"current":42,"description":57,"end":62,"extension":25,"highlights":63,"location":69,"meta":70,"navigation":42,"path":71,"period":72,"role":73,"seo":74,"stack":75,"start":81,"stem":82,"summary":83,"type":84,"__hash__":85},"experience\u002Fexperience\u002Frains.md","Rains",{"type":8,"value":54,"toc":58},[55],[11,56,57],{},"The agent runs alongside a person's day instead of asking them to change it: it logs the workflow,\ndrafts the report, and the person corrects what it got wrong. The corrections matter, because the\nagent only earns its place once someone trusts the output enough to stop writing it twice.",{"title":21,"searchDepth":22,"depth":22,"links":59},[],"RAINS Corp","https:\u002F\u002Frainscorp.id",null,[64,65,66,67,68],"Architected and deployed an autonomous shadow agent that now runs alongside clinical and administrative staff, across physician, healthcare worker and front-office roles.","Built the inter-agent query routing that holds a 92% query success rate, and I am benchmarking latency against a sub-5s target.","Lead a squad of 2 engineers, ship to staging through CI\u002FCD every day, and steer sprints toward 70%+ milestone completion.","Own the outward-facing half of the job too: demo the agent to investors, run the weekly sync and sprint sessions, and open pilot conversations with partner organisations.","Shipped the production platform at rainscorp.id and the custom GPT apps that give the agent a natural-language front door.","Remote",{},"\u002Fexperience\u002Frains","2026-07 to present","Lead Developer",{"description":57},[76,77,78,79,80],"python","typescript","nuxt","postgresql","docker","2026-07-01","experience\u002Frains","An autonomous AI agent that runs alongside clinical and administrative staff and writes the activity reports they would otherwise write by hand.","Remote · Founding team","QbCI-h8fs1CsAhtR9eOYArUQ1pAQkGUyy38PSV8yJ6g",{"id":87,"title":88,"body":89,"company":88,"companyUrl":62,"current":42,"description":93,"end":62,"extension":25,"highlights":96,"location":69,"meta":102,"navigation":42,"path":103,"period":104,"role":105,"seo":106,"stack":107,"start":111,"stem":112,"summary":113,"type":114,"__hash__":115},"experience\u002Fexperience\u002Fmayfeyr.md","Mayfeyr",{"type":8,"value":90,"toc":94},[91],[11,92,93],{},"I also own the multi-server infrastructure and the technical SEO\u002FGEO work that got the full catalogue\nindexed and crawling cleanly.",{"title":21,"searchDepth":22,"depth":22,"links":95},[],[97,98,99,100,101],"Rebuilt catalogue ingestion as a distributed pipeline in Python and RabbitMQ, sized so that onboarding a large supplier catalogue and the daily price sync both run without losing work.","Integrated external supplier APIs behind strict retry queues, so a failed sync retries instead of surfacing as a stale price or a negative-margin order.","Refactored the pricing engine around DataLoader batching and cut per-product pricing from 100ms to 10ms, a 10x throughput gain.","Pulled the hot paths out into Go microservices (pricing, image caching, content generation, PDF rendering) and moved the rest of the background work onto Celery and n8n.","Built Mayfeyr Mirror, the virtual try-on engine, from zero to one, plus custom ChatGPT apps that take a conversational shopping query straight to a product.",{},"\u002Fexperience\u002Fmayfeyr","2024-12 to present","Full-Stack AI Engineer",{"description":93},[76,108,109,79,110,80],"go","rabbitmq","celery","2024-12-01","experience\u002Fmayfeyr","Catalogue ingestion, pricing, storefront content, and the AI features on top of them for a multi-brand luxury fashion retailer.","Remote · Full-time","l5UmmTr61aSGD7uD-hMoZHbWlIgK6Q1ZKpJKHpxhi0w",{"id":117,"title":118,"body":119,"company":123,"companyUrl":62,"current":124,"description":21,"end":62,"extension":25,"highlights":125,"location":130,"meta":131,"navigation":42,"path":132,"period":133,"role":134,"seo":135,"stack":136,"start":137,"stem":138,"summary":139,"type":118,"__hash__":140},"experience\u002Fexperience\u002Feducation.md","Education",{"type":8,"value":120,"toc":121},[],{"title":21,"searchDepth":22,"depth":22,"links":122},[],"Universitas Brawijaya",false,[126,127,128,129],"First author on a Q1 IEEE Access paper for GenBatik, with registered intellectual property and an open-source implementation.","Team lead on the Bangkit Academy capstone with Solafune Japan, ranked #1 in the company track and #1 overall across the academy, graduated with distinction.","Lab assistant for two years, teaching basic programming, data structures and algorithms, and computer networking at basic and advanced level.","Computer vision lab work on retinal blood vessel segmentation and historical manuscript text recognition.","Malang, Indonesia",{},"\u002Fexperience\u002Feducation","2021 to 2025","B.Sc. Computer Science",{"description":21},[],"2021-09-01","experience\u002Feducation","Faculty valedictorian, 3.95\u002F4.00 GPA, degree finished in 3.5 years.","dYxm3XGT3X-os_Cp5mZmHpHOaWBXTL9CVS9H9iNfwzk",[142,231,298,355,421,526,584,690],{"id":143,"title":144,"body":145,"cover":62,"description":21,"extension":25,"featured":42,"links":202,"meta":205,"metrics":206,"navigation":42,"order":213,"path":214,"pipelineStage":215,"role":216,"seo":217,"stack":218,"status":222,"stem":223,"tagline":224,"tags":225,"timeframe":62,"__hash__":230},"projects\u002Fprojects\u002Fgenbatik.md","GenBatik",{"type":8,"value":146,"toc":197},[147,152,155,159,162,185,189],[148,149,151],"h2",{"id":150},"the-problem","The problem",[11,153,154],{},"Batik motifs are usually recreated by hand or pulled from a fixed catalogue. GenBatik lets someone\ndescribe the pattern they want in plain language and get a batik-style image back, so a small\nworkshop can explore variants without a designer.",[148,156,158],{"id":157},"what-i-built","What I built",[11,160,161],{},"The whole path: the generation pipeline, the API that queues and streams results, and the interface\nthat shows progress to the user instead of a spinner that lies.",[163,164,165,173,179],"ul",{},[166,167,168,172],"li",{},[169,170,171],"strong",{},"Model",": fine-tuned diffusion on batik imagery, served behind a Python worker.",[166,174,175,178],{},[169,176,177],{},"Serving",": asynchronous job queue with WebSocket progress. Generation takes seconds, so the UI reports stages instead of blocking a request.",[166,180,181,184],{},[169,182,183],{},"Product",": Nuxt frontend with prompt presets, so a first-time user gets a usable result without knowing prompt engineering.",[148,186,188],{"id":187},"what-came-out-of-it","What came out of it",[11,190,191,192,196],{},"The platform has synthesized 60K+ patterns for public users with a 0% request drop rate, because the\nqueue absorbs load instead of refusing it. The work became my first-author paper in IEEE Access\n(Q1), the framework is registered for intellectual property, and both the interface and the model\nare open source at ",[193,194,195],"code",{},"izzudd\u002Fgen-batik",". I now review for the journal as well.",{"title":21,"searchDepth":22,"depth":22,"links":198},[199,200,201],{"id":150,"depth":22,"text":151},{"id":157,"depth":22,"text":158},{"id":187,"depth":22,"text":188},{"live":203,"repo":204},"https:\u002F\u002Fgenbatik.ub.ac.id","https:\u002F\u002Fgithub.com\u002Fizzudd\u002Fgen-batik",{},[207,210],{"label":208,"value":209},"Batik patterns generated","60K+",{"label":211,"value":212},"Dropped requests","0%",1,"\u002Fprojects\u002Fgenbatik",[171,177,183],"Model, backend, and frontend",{"title":144,"description":21},[76,219,220,78,221],"pytorch","huggingface","websocket","Live","projects\u002Fgenbatik","Text-to-image generation for batik patterns: a diffusion model wrapped in a real product, and a paper about it.",[226,227,228,229],"AI","Image Generation","Diffusion","Fullstack","a-TV4YAWjyiv5Tam9WS-KK1eafE_P2SoJBrwo0n0NPU",{"id":232,"title":233,"body":234,"cover":62,"description":21,"extension":25,"featured":42,"links":275,"meta":276,"metrics":277,"navigation":42,"order":22,"path":281,"pipelineStage":282,"role":283,"seo":284,"stack":285,"status":289,"stem":290,"tagline":291,"tags":292,"timeframe":296,"__hash__":297},"projects\u002Fprojects\u002Fsparse.md","SPARSE",{"type":8,"value":235,"toc":270},[236,238,241,243,263,267],[148,237,151],{"id":150},[11,239,240],{},"Mapping rooftop solar panels by hand across a region is slow. SPARSE detects them from satellite\ntiles so the result can be inspected on a map instead of a spreadsheet.",[148,242,158],{"id":157},[163,244,245,251,257],{},[166,246,247,250],{},[169,248,249],{},"Pipeline",": tile ingestion, model inference, and geo-referenced output, so each detection keeps its coordinates.",[166,252,253,256],{},[169,254,255],{},"API",": FastAPI service that serves predictions per tile.",[166,258,259,262],{},[169,260,261],{},"Interface",": map view where a reviewer can confirm or reject detections.",[148,264,266],{"id":265},"the-competition-half","The competition half",[11,268,269],{},"The project was the Company Capstone with Solafune Japan in Bangkit Academy, and I led the team. We\nbenchmarked four vision architectures (ResNet, EfficientNetV2, CoAtNet, MaxViT) on 12-band\nSentinel-2 imagery before settling on MaxViT Base, which gave us a 0.6459 test F1. We placed #1 in\nthe company capstone track and #1 overall across the academy.",{"title":21,"searchDepth":22,"depth":22,"links":271},[272,273,274],{"id":150,"depth":22,"text":151},{"id":157,"depth":22,"text":158},{"id":265,"depth":22,"text":266},{},{},[278],{"label":279,"value":280},"Test F1 (MaxViT-B)","0.6459","\u002Fprojects\u002Fsparse",[171,177,183],"Team lead, model",{"title":233,"description":21},[76,219,286,287,288,80],"fastapi","vue","leaflet","Awarded","projects\u002Fsparse","Solar panel detection on satellite imagery, from model to a GIS interface.",[226,293,294,295],"Computer Vision","Remote Sensing","GIS","2024","X_YOzzvJBCTGax6z7UNgkHL4brvCBvdoLu3OnLprnns",{"id":299,"title":300,"body":301,"cover":62,"description":21,"extension":25,"featured":42,"links":336,"meta":338,"metrics":339,"navigation":42,"order":340,"path":341,"pipelineStage":342,"role":343,"seo":344,"stack":345,"status":348,"stem":349,"tagline":350,"tags":351,"timeframe":62,"__hash__":354},"projects\u002Fprojects\u002Fidentifeye.md","IdentifEye",{"type":8,"value":302,"toc":332},[303,305,308,312],[148,304,151],{"id":150},[11,306,307],{},"Second-factor authentication usually means a code or a hardware key. IdentifEye explores a third\noption: prove identity with a face, without shipping a whole identity platform.",[148,309,311],{"id":310},"how-it-works","How it works",[163,313,314,320,326],{},[166,315,316,319],{},[169,317,318],{},"Embedding",": FaceNet produces a feature vector for each face.",[166,321,322,325],{},[169,323,324],{},"Verification",": similarity score against the enrolled vector, with a threshold instead of a binary classifier. It is a comparison problem, not a classification problem.",[166,327,328,331],{},[169,329,330],{},"Service",": Go HTTP API that wraps the inference step, so the model stays a small independent process rather than being embedded in the app.",{"title":21,"searchDepth":22,"depth":22,"links":333},[334,335],{"id":150,"depth":22,"text":151},{"id":310,"depth":22,"text":311},{"repo":337},"https:\u002F\u002Fgithub.com\u002Fizzudd\u002FidentifEye",{},[],3,"\u002Fprojects\u002Fidentifeye",[171,177],"Solo",{"title":300,"description":21},[108,76,219,346,347],"react","tailwindcss","Open source","projects\u002Fidentifeye","Face-embedding authentication exposed as a small, boring REST API.",[226,352,353,255],"Vision","Embeddings","fwDYLw_xWzfw4JKRhNQkmbtlWnUXyVz0lTsfzZkWXgY",{"id":356,"title":357,"body":358,"cover":62,"description":21,"extension":25,"featured":124,"links":401,"meta":403,"metrics":404,"navigation":42,"order":411,"path":412,"pipelineStage":413,"role":343,"seo":414,"stack":415,"status":222,"stem":416,"tagline":417,"tags":418,"timeframe":62,"__hash__":420},"projects\u002Fprojects\u002Ftatakata.md","TataKata",{"type":8,"value":359,"toc":396},[360,362,365,369,389,393],[148,361,151],{"id":150},[11,363,364],{},"Spell-checkers handle typos, not Indonesian grammar. TataKata is an attempt at the harder part:\nnon-standard words, affixes, and punctuation that EYD cares about.",[148,366,368],{"id":367},"approach","Approach",[163,370,371,377,383],{},[166,372,373,376],{},[169,374,375],{},"Rules first",": deterministic EYD\u002FKBBI corrections where the right answer is unambiguous.",[166,378,379,382],{},[169,380,381],{},"Model second",": only send the sentences that rules cannot resolve to a language model, because it is slower and less predictable.",[166,384,385,388],{},[169,386,387],{},"Feedback loop",": every accepted correction is stored so the rule set and the model improve.",[148,390,392],{"id":391},"where-it-is-now","Where it is now",[11,394,395],{},"Live and in use by 60+ active writers, who have run 100K+ document enhancements through it. The\nstored accept\u002Freject decisions are what I use to decide whether a correction belongs in the rule set\nor the model.",{"title":21,"searchDepth":22,"depth":22,"links":397},[398,399,400],{"id":150,"depth":22,"text":151},{"id":367,"depth":22,"text":368},{"id":391,"depth":22,"text":392},{"live":402},"https:\u002F\u002Ftata-kata.netlify.app",{},[405,408],{"label":406,"value":407},"Active users","60+",{"label":409,"value":410},"Documents enhanced","100K+",4,"\u002Fprojects\u002Ftatakata",[171],{"title":357,"description":21},[78,79,76],"projects\u002Ftatakata","Grammar checking for Bahasa Indonesia, built on EYD and KBBI rules, used by real writers.",[226,419,34],"NLP","DtZks7ZpmqxCJeqOM9Tjo8eTxse6mXC_EzWfaIA8aXw",{"id":422,"title":423,"body":424,"cover":62,"description":21,"extension":25,"featured":42,"links":503,"meta":506,"metrics":507,"navigation":42,"order":511,"path":512,"pipelineStage":513,"role":343,"seo":514,"stack":515,"status":222,"stem":518,"tagline":519,"tags":520,"timeframe":524,"__hash__":525},"projects\u002Fprojects\u002Fquoin.md","Quoin",{"type":8,"value":425,"toc":498},[426,430,433,435,477,481,487],[148,427,429],{"id":428},"why","Why",[11,431,432],{},"Most \"online JSON formatter\" pages are a textarea in front of a logging server, so pasting a token or\na customer payload into one is a small security incident nobody notices. Quoin takes the opposite\nposition: every tool is a pure function in the browser, and the data never leaves the tab.",[148,434,158],{"id":157},[163,436,437,447,453,467],{},[166,438,439,442,443,446],{},[169,440,441],{},"Grouped by task, not by vendor."," JSON beautifier and tree visualizer, JSON to TypeScript, Zod,\nPydantic and Go type generator, JSON Schema inference, encoders, slugifier, text and Git diffs,\ntimezone conversion, table and CSV editing with SQL ",[193,444,445],{},"INSERT"," output, Markdown conversion. The list\ngrows whenever I get annoyed.",[166,448,449,452],{},[169,450,451],{},"One registry, one shell."," A single SPA driven by a typed tool registry. The sidebar, the home\ngrid and the command palette all read from the same file, so shipping a tool is one page plus one\nentry, not a new app.",[166,454,455,458,459,462,463,466],{},[169,456,457],{},"Built for repetition."," Command palette on ",[193,460,461],{},"Ctrl+P",", pinnable tools, and drafts that survive a\nreload through ",[193,464,465],{},"localStorage",", because a utility you use daily should remember your work.",[166,468,469,472,473,476],{},[169,470,471],{},"Shipped as static assets"," on a Cloudflare Worker, built by Workers Builds on every push to\n",[193,474,475],{},"main",", gated by lint and typecheck in CI.",[148,478,480],{"id":479},"privacy-by-construction","Privacy by construction",[11,482,483,484,486],{},"No analytics, no cookies, no third-party requests. State lives in page memory or ",[193,485,465],{}," and\nnever on a server. Keeping the tools pure functions means privacy is a property of the architecture\nrather than a promise in a policy page.",[11,488,489,490,497],{},"I wrote the story behind it on dev.to: ",[491,492,496],"a",{"href":493,"rel":494},"https:\u002F\u002Fdev.to\u002Fizzudd\u002Fi-built-the-dev-tools-i-was-tired-of-looking-for-2ejh",[495],"nofollow","I built the dev tools I was tired of looking for",".",{"title":21,"searchDepth":22,"depth":22,"links":499},[500,501,502],{"id":428,"depth":22,"text":429},{"id":157,"depth":22,"text":158},{"id":479,"depth":22,"text":480},{"live":504,"repo":505},"https:\u002F\u002Fquoin.izzudd.my.id","https:\u002F\u002Fgithub.com\u002Fizzudd\u002Fquoin",{},[508],{"label":509,"value":510},"Payloads uploaded","0",5,"\u002Fprojects\u002Fquoin",[183],{"title":423,"description":21},[78,287,347,77,516,517],"bun","cloudflare","projects\u002Fquoin","Developer utilities that run entirely in the browser: no uploads, no accounts, no telemetry.",[521,522,523],"DevTools","Developer Experience","Privacy","2026","_FPVDJeksP7BOBQbqbItr1ROFH_viVHpqX7hiTNEX84",{"id":527,"title":528,"body":529,"cover":62,"description":21,"extension":25,"featured":124,"links":564,"meta":565,"metrics":566,"navigation":42,"order":570,"path":571,"pipelineStage":572,"role":343,"seo":573,"stack":574,"status":577,"stem":578,"tagline":579,"tags":580,"timeframe":582,"__hash__":583},"projects\u002Fprojects\u002Falpr.md","ALPR Pipeline",{"type":8,"value":530,"toc":560},[531,533,536,538,557],[148,532,151],{"id":150},[11,534,535],{},"Satria Data 2023 asked for an end-to-end pipeline that finds a license plate in a photo and reads\nthe characters on it. Two problems dressed as one: detection is a localization task, reading is a\nsequence task, and a mistake in the first one cannot be fixed by the second.",[148,537,158],{"id":157},[163,539,540,546,552],{},[166,541,542,545],{},[169,543,544],{},"Detection",": a CNN with custom activation layers to locate the plate in the frame.",[166,547,548,551],{},[169,549,550],{},"Recognition",": an ASTER-style network (CNN encoder, LSTM sequence model) that reads the crop character by character instead of segmenting it first.",[166,553,554,556],{},[169,555,249],{},": both stages chained into one service, with the intermediate crop kept so a wrong reading can be traced back to a wrong crop.",[11,558,559],{},"Placed 12th nationally.",{"title":21,"searchDepth":22,"depth":22,"links":561},[562,563],{"id":150,"depth":22,"text":151},{"id":157,"depth":22,"text":158},{},{},[567],{"label":568,"value":569},"National rank","12",6,"\u002Fprojects\u002Falpr",[171],{"title":528,"description":21},[76,219,575,576],"cnn","lstm","Competition","projects\u002Falpr","Automatic license plate recognition, built from scratch for a national data competition.",[226,293,581],"OCR","2023","VlFmzZ1Hzt_eBhZDoHi2LDXKh16ZXqmjQt6hCqMSAdA",{"id":585,"title":586,"body":587,"cover":62,"description":21,"extension":25,"featured":124,"links":665,"meta":667,"metrics":668,"navigation":42,"order":675,"path":676,"pipelineStage":677,"role":678,"seo":679,"stack":680,"status":222,"stem":683,"tagline":684,"tags":685,"timeframe":524,"__hash__":689},"projects\u002Fprojects\u002Fautomatrix.md","AutoMatrix",{"type":8,"value":588,"toc":660},[589,591,594,596,599,650,654,657],[148,590,151],{"id":150},[11,592,593],{},"Hosting prices are not comparable. The sticker omits egress, the IPv4 address, the renewal after\nthe promo window, and whether the meter runs per workflow, per task, or per credit. Every \"cheap VPS\"\nlist online compares the number that is easiest to scrape instead of the number that shows up on the\ninvoice.",[148,595,158],{"id":157},[11,597,598],{},"An audited price index, published as dated editions (currently 2026-Q4) rather than a live monitor,\nbecause a monitor across eleven vendors makes every figure unfalsifiable after the fact.",[163,600,601,611,624,634,644],{},[166,602,603,606,607,610],{},[169,604,605],{},"Traceability",": every plan record stores ",[193,608,609],{},"source_url",", a verbatim quote from the vendor page, the date it was read, and whether it was read directly or through a rendering proxy. The quote is expandable on the page and its sha256 is published.",[166,612,613,616,617,620,621,623],{},[169,614,615],{},"Null is not zero",": an unpublished rate is ",[193,618,619],{},"null",", never ",[193,622,510],{},". A plan that prices a term but not the rate after it refuses to compute a 36-month total.",[166,625,626,629,630,633],{},[169,627,628],{},"Arithmetic that fails the build",": converted prices must match ",[193,631,632],{},"native_price × fx_rate"," within two cents, the FX date must match the FX file, and every category needs exactly three featured plans. The build stops otherwise.",[166,635,636,639,640,643],{},[169,637,638],{},"The audits",": three live audits (compute, coding agents, automation), each provider scored on compute, egress, ipv4, storage and contract. ",[193,641,642],{},"not_published"," is treated as a finding, not a gap in the work.",[166,645,646,649],{},[169,647,648],{},"Output as data",": the site renders from the same records it publishes, so the prose cannot drift from the numbers. The dataset ships with a JSON Schema (2020-12) and a verification manifest.",[148,651,653],{"id":652},"how-it-is-held-honest","How it is held honest",[11,655,656],{},"A CI gate validates the schema, the five-vector coverage and the arithmetic; another checks the built\ndataset against its own published JSON Schema; a third re-checks weekly that every cited URL still\nresolves. The static output is generated by a script and deployed as a Cloudflare Worker with the\narchive embedded, since Pages direct-upload cannot be driven without custom headers.",[11,658,659],{},"The whole argument is one table: eleven compute plans, 36-month totals at 1 TB and 5 TB a month. The\ncheapest sticker is the cheapest bill at every traffic level I priced. The second-cheapest is not.",{"title":21,"searchDepth":22,"depth":22,"links":661},[662,663,664],{"id":150,"depth":22,"text":151},{"id":157,"depth":22,"text":158},{"id":652,"depth":22,"text":653},{"live":666},"https:\u002F\u002Fmatrix.invasikode.com",{},[669,672],{"label":670,"value":671},"Plans audited","18",{"label":673,"value":674},"Audits live","3",7,"\u002Fprojects\u002Fautomatrix",[183],"Concept, data model, build and deploy",{"title":586,"description":21},[77,681,517,682],"zod","github-actions","projects\u002Fautomatrix","A dated price index for developer infrastructure, where every number carries the sentence it came from.",[686,687,688],"Data","Publishing","Infrastructure","csYP_11GQJg75qOY-YiCkpbvYpwFQy2hvsS1u2BHQZU",{"id":691,"title":692,"body":693,"cover":62,"description":21,"extension":25,"featured":124,"links":767,"meta":769,"metrics":770,"navigation":42,"order":771,"path":772,"pipelineStage":773,"role":343,"seo":774,"stack":775,"status":348,"stem":779,"tagline":780,"tags":781,"timeframe":524,"__hash__":784},"projects\u002Fprojects\u002Flinkproxy.md","LinkProxy",{"type":8,"value":694,"toc":763},[695,697,700,702,756],[148,696,151],{"id":150},[11,698,699],{},"A direct S3 or CDN URL leaks where the file really lives, breaks the day the bucket moves, and\nloses range support the moment it is wrapped in a shortener that buffers the response first. I wanted\nstable links that still behave like the origin.",[148,701,158],{"id":157},[163,703,704,714,732,738,744,750],{},[166,705,706,709,710,713],{},[169,707,708],{},"Bounded-memory streaming",": pooled 64 KB buffers from ",[193,711,712],{},"sync.Pool",", so a multi-gigabyte file streams at O(1) memory with no temporary disk write.",[166,715,716,719,720,723,724,727,728,731],{},[169,717,718],{},"Range passthrough",": ",[193,721,722],{},"Range"," and ",[193,725,726],{},"Content-Range"," go through untouched, so video seeking and ",[193,729,730],{},"curl -C -"," resumption work as if you had hit the origin.",[166,733,734,737],{},[169,735,736],{},"SSRF protection",": every destination IP is checked against loopback, private IPv4 and IPv6, link-local and non-HTTP schemes, including across redirect chains.",[166,739,740,743],{},[169,741,742],{},"Deterministic links",": tracking parameters stripped, remaining query parameters sorted, and a 16-character Base62 id derived from a SHA-256 of the canonical URL, so the same file maps to the same link every time.",[166,745,746,749],{},[169,747,748],{},"Dashboard without a build step",": HTMX and Tailwind served from the Go binary, showing bandwidth served, access counts, searchable file records and live request metrics.",[166,751,752,755],{},[169,753,754],{},"Two storage modes",": pure-Go SQLite in WAL mode with zero CGO, plus an optional Redis cache that degrades to SQLite on its own when Redis is absent or unreachable.",[11,757,758,759,762],{},"The admin API sits behind basic auth while the public ",[193,760,761],{},"\u002Fp\u002F{id}"," links stay unauthenticated, because\nthe point of a mirror link is that the person you send it to does not need an account.",{"title":21,"searchDepth":22,"depth":22,"links":764},[765,766],{"id":150,"depth":22,"text":151},{"id":157,"depth":22,"text":158},{"repo":768},"https:\u002F\u002Fgithub.com\u002Fizzudd\u002Furl-proxy",{},[],8,"\u002Fprojects\u002Flinkproxy",[177,183],{"title":692,"description":21},[108,776,777,778,80],"sqlite","redis","htmx","projects\u002Flinkproxy","A streaming gateway that turns any file URL into a stable short link without buffering the file.",[782,783,688],"Go","Streaming","iBaCunhxhJBh75zZXceGQjukvarCco5jHCjFcmKrbCI",[786,825,868,907],{"id":787,"title":788,"body":789,"description":21,"extension":25,"items":793,"meta":805,"navigation":42,"order":213,"path":806,"seo":807,"stage":808,"stem":809,"summary":810,"tools":811,"__hash__":824},"skills\u002Fskills\u002F01-data-modeling.md","Data & Modeling",{"type":8,"value":790,"toc":791},[],{"title":21,"searchDepth":22,"depth":22,"links":792},[],[794,796,799,803],{"name":795},"Dataset preparation and label sanity checks",{"name":797,"evidence":144,"evidenceHref":798},"Fine-tuning diffusion models (LoRA \u002F SDXL)","#project-genbatik",{"name":800,"evidence":801,"evidenceHref":802},"CNN \u002F embedding models for vision","SPARSE, IdentifEye","#project-sparse",{"name":804,"evidence":233,"evidenceHref":802},"Evaluation beyond a single accuracy number",{},"\u002Fskills\u002F01-data-modeling",{"title":788,"description":21},"01","skills\u002F01-data-modeling","Framing the problem, cleaning the data, training and evaluating a model I can defend.",[812,815,818,821],{"name":813,"icon":814},"PyTorch","simple-icons:pytorch",{"name":816,"icon":817},"Hugging Face","simple-icons:huggingface",{"name":819,"icon":820},"Python","simple-icons:python",{"name":822,"icon":823},"NumPy","simple-icons:numpy","BAxygUscJAcYvKaI3uHLM0HcTlsXMMyrMTX-fBBu71U",{"id":826,"title":827,"body":828,"description":21,"extension":25,"items":832,"meta":849,"navigation":42,"order":22,"path":850,"seo":851,"stage":852,"stem":853,"summary":854,"tools":855,"__hash__":867},"skills\u002Fskills\u002F02-serving.md","Serving & Inference",{"type":8,"value":829,"toc":830},[],{"title":21,"searchDepth":22,"depth":22,"links":831},[],[833,836,838,840,843,845,847],{"name":834,"evidence":835,"evidenceHref":798},"REST \u002F async job APIs around model inference","GenBatik, IdentifEye",{"name":837,"evidence":144,"evidenceHref":798},"Streaming progress to the client (WebSocket)",{"name":839,"evidence":233,"evidenceHref":802},"Containerised deploy, GPU vs CPU trade-offs",{"name":841,"evidence":842,"evidenceHref":798},"Latency and cost per request","GenBatik, Mayfeyr",{"name":844},"Message queues and retry policies that survive supplier failures",{"name":846,"evidence":88},"Batch loading to cut per-product work (DataLoader, 100ms to 10ms)",{"name":848,"evidence":88},"Microservice extraction when one module gets hot",{},"\u002Fskills\u002F02-serving",{"title":827,"description":21},"02","skills\u002F02-serving","A model that only runs in a notebook is not a feature. I put it behind an endpoint and keep it honest.",[856,859,861,864],{"name":857,"icon":858},"FastAPI","simple-icons:fastapi",{"name":782,"icon":860},"simple-icons:go",{"name":862,"icon":863},"Docker","simple-icons:docker",{"name":865,"icon":866},"Cloudflare","simple-icons:cloudflare","JuI5lrbIcP_ilv9ivy9aN6fflkUUNxQw251Tnbus_pE",{"id":869,"title":870,"body":871,"description":21,"extension":25,"items":875,"meta":890,"navigation":42,"order":340,"path":891,"seo":892,"stage":893,"stem":894,"summary":895,"tools":896,"__hash__":906},"skills\u002Fskills\u002F03-monitoring.md","Monitoring & Iteration",{"type":8,"value":872,"toc":873},[],{"title":21,"searchDepth":22,"depth":22,"links":874},[],[876,878,881,883,885,888],{"name":877,"evidence":144,"evidenceHref":798},"Logging inputs\u002Foutputs to make failures reproducible",{"name":879,"evidence":357,"evidenceHref":880},"Feedback capture (accepted vs rejected output)","#project-tatakata",{"name":882},"Cost and usage tracking per feature",{"name":884},"Versioned experiments instead of one-off notebooks",{"name":886,"evidence":887},"Success rate and latency targets measured, not assumed","Rains, 92% query success",{"name":889,"evidence":52},"CI\u002FCD to staging as the daily default, not a release event",{},"\u002Fskills\u002F03-monitoring",{"title":870,"description":21},"03","skills\u002F03-monitoring","The part most portfolios skip. I care about what happens after deploy: what broke, what drifted, and what the next version should change.",[897,900,903],{"name":898,"icon":899},"Postgres","simple-icons:postgresql",{"name":901,"icon":902},"Grafana","simple-icons:grafana",{"name":904,"icon":905},"GitHub Actions","simple-icons:githubactions","DO27PErUcJIhQuGbYhLcuwRtJpzPNslZhvrdaX3RMLY",{"id":908,"title":909,"body":910,"description":21,"extension":25,"items":914,"meta":935,"navigation":42,"order":411,"path":936,"seo":937,"stage":938,"stem":939,"summary":940,"tools":941,"__hash__":952},"skills\u002Fskills\u002F04-product.md","Product Delivery",{"type":8,"value":911,"toc":912},[],{"title":21,"searchDepth":22,"depth":22,"links":913},[],[915,918,920,923,925,928,931,933],{"name":916,"evidence":917,"evidenceHref":798},"Nuxt \u002F Vue front ends for data-heavy products","GenBatik, invasikode.com",{"name":919},"API design and schema work in Postgres",{"name":921,"evidence":922},"AI features wired into an existing commerce stack","Mayfeyr Mirror, custom GPT apps",{"name":924,"evidence":88},"Multi-server infrastructure and technical SEO\u002FGEO",{"name":926,"evidence":927},"Static-first builds, prerender, and deploy pipelines","Portfolio (this site)",{"name":929,"evidence":930,"evidenceHref":35},"Technical writing that explains a system to non-authors","invasikode.com",{"name":932,"evidence":52},"Demoing a working system to investors, and partner conversations for pilots",{"name":934,"evidence":52},"Running sprint planning and team sync sessions",{},"\u002Fskills\u002F04-product",{"title":909,"description":21},"04","skills\u002F04-product","An AI feature is only as good as the interface around it. I build that part too.",[942,945,948,951],{"name":943,"icon":944},"Nuxt","simple-icons:nuxt",{"name":946,"icon":947},"TypeScript","simple-icons:typescript",{"name":949,"icon":950},"Tailwind","simple-icons:tailwindcss",{"name":898,"icon":899},"6x4RLhG4uNlp2G4Mjo0uNlup3ncYXZARQ_W_miNdB2Y",1793984327129]