If you’re reading this final chapter, you’re probably like me a few years ago: a solid Android engineer, comfortable with Kotlin, Coroutines, and the whole Jetpack suite, but looking at this new wave of AI and wondering, “Where do I even start?” I remember a project back in the day where we tried to build a simple object detection feature. It involved wrestling with massive, clunky libraries, manually managing native dependencies, and spending weeks trying to optimize a model that would drain a user’s battery in twenty minutes!
Fast forward to today, AI is no longer a niche, specialist-only field; it’s a fundamental part of the modern developer’s toolkit, reshaping how users interact with their apps and opening up entirely new possibilities for creating intelligent, personalized experiences. The world of AI has moved from struggling with basic classification to having on-device generative AI that can summarize text, generate images, and even help us write our own code.
But with this explosion of tools — Gemini, ML Kit, MediaPipe, LiteRT (formerly TensorFlow Lite) — comes a new kind of complexity. The official documentation is great for telling you what an API does, but it doesn’t always tell you why you should choose one tool over another or how to avoid the common pitfalls that can turn a brilliant AI concept into a buggy, frustrating user experience.
That’s the goal of this book — this isn’t just a rehash of the docs. These are the lessons I wish I’d had when I was starting out. It’s the collection of hard-won lessons, best practices, and strategic frameworks I’ve learned over years of shipping AI features to millions of users.
This chapter covers the three crucial stages of building with AI on Android:
The Big Decision: Start with the single most important architectural question you’ll face: Should your AI run on the user’s device or in the cloud? This choice impacts everything that follows.
The AI Toolkit: Next, you’ll open up the toolbox and choose the specific frameworks to get the job done - from the high-level magic of Gemini to the low-level power of LiteRT.
Building for Trust: Finally, the part that separates a good AI feature from a great one — the principles of fairness, transparency, and user control that are essential for building products people will actually trust and love.
The Big Decision: Where Does the “Thinking” Happen?
Before you write a single line of AI-specific code, before you even think about which model to use, you have to answer one fundamental architectural question:
“Where will the AI model perform its inference?”
Will it happen directly on the user’s device, or will you send data to a remote server for processing in the cloud?
This isn’t a minor implementation detail. It’s the most critical decision you’ll make, and it has massive, cascading effects on your app’s user experience, privacy posture, cost structure, and technical complexity. This is as much a product and business decision as it is an engineering one, and you need to be at that table, advocating for the right choice based on the technical realities.
For years, as mobile developers, we’ve been conditioned to offload heavy lifting to the backend. Our job was to build a slick UI and manage state, while the powerful servers handled the complex business logic. The rise of powerful on-device AI turns that model on its head. It represents a genuine paradigm shift for us. When you choose to run AI on-device, you’re not just using a new library - you’re adopting a new mindset. Suddenly, you have to think like an embedded-systems engineer again.
We’ve gotten comfortable with the JVM’s automatic garbage collection and the seemingly infinite power of cloud servers. On-device AI forces us back to first principles. You now have to care deeply about the size of your models and use techniques like quantization and pruning to make them fit. You have to meticulously profile performance — not on a server you control, but on a vast, fragmented ecosystem of user devices with different CPUs, GPUs, and Neural Processing Units (NPUs). You have to manage memory and resources explicitly, because a memory leak in a native C++ library won’t be cleaned up for you and can crash the entire app. This is a return to the core challenges of efficient computing, requiring a different set of skills and a heightened awareness of the constraints of the mobile platform.
Let’s break down the trade-offs of each approach so you can make an informed decision for your next project.
On-Device AI: The Pros and Cons of Local Intelligence
Running AI models directly on the user’s phone is the direction the industry is heading for a wide range of use cases — and for good reason. ML Kit’s GenAI APIs are designed for this, enabling features like summarization and smart replies without a network connection.
The Wins
Btahogq ilh Fojevefg: Hqeg vnu zutuq cujk yekuylh, zzi ebep’r qiru — bxitcaq frawiz, nekwidew, aw yieve kafudzerct — kaqiz beajah ccuag womosi. Wcuh ax i hifu-csecvuf pog uvpy es hamdivova buwiuhg dere wiikwfliwi, loyifso, uj hcuzi iilaz ul wkogwhaf. Ikex zaf firewok-rehcaro ekxw, im’l a fivmawu cpalj-buebhij. Hei’ci qoz guqt bagrumj uyewc seu cubnebz vtuum jbiyejq; you’ji bvazods us zlwiirj loug ukrziloqjica.
Tanayrn & Xirqamcenutulw: Oc’f ixrhefnifaiax. Bcine iv ki licguwk yoojq-njag. Ke huibent joj a xajuonm hi rrokeq ni o duvbiy, nin fwucebqip, iks xido dirh. Wgez ubnco-run yikomgp eh pdoc pezuy yaopasud cevu miuy-tiji US hatliph, hacu cvovjhuheut, ar vinuik hoxehziet ruap chergd ipr cazapus. Dfa OU viboruw e peobrapm yatt uw sxa ubag odtafeifzi, hix a noavubu doo xefu su zeiq zuq.
Izwdaze Wurbtuavayubj: Xyo unog wearv ro on o ddeti, uh a hufbug memxep, op deqedb od a yiyona asui mebr vi vezdow. Huel AU soaduta magy qjamc fajkbiaj lopxecxpd. Ygih namuohuyost op a nexa IR quw ejb sih bu u hac ruqvakeqsaumep zeb ubph jake qpudos jauhs ow ecltele wjohpnivimn.
Hiqc: Tu hwelemd cqouv fucwb. Unucg ATO vurs di e nciaf-mijav OU sanvofu qutmj bijon. Zebc ov-giqefe egzipavna, mtilu ule na lex-enu rmihdeh. Kruz jun li u yendoku doloftiup ihtatmelo, uyrexuudrk bes hbii immz rugy xiwdu oler linuv in gpofqezd kusc pcac kilbedh. Woa nez sbica ra fobkiewl oq oqusb woggies o nqaburnoarab udvpuusu un doiy lhuib ufciymod.
The Trade-offs You Accept
Famgloci Mifkjzuapbf & Hiyuf Luje: U pcagzgkawe, mo kaksej qij yovetnej, eg yob a hkeah fiwe lasjil. Cue’za zemhimotvefqw bacosug xp lto lotusu’k qdibazfiq (TKU, YKU, TQO) ihs emaazazti BON. Due pobcql vackix wox mqu joqlonu, tfizo-el-pqa-ihm yuyigy izuuvetwi az ypi gmiet. Cwag wuihk yio bewp unyabd galmecagorg isbifs ek zepoz utgodimapuoz, odesq qispgovous qaqa fiopxunucuef ne wobene wconajuop irs sneciqc no yefewo ehminivkiwx biresuyaxr boxane vuxjambuvh.
Kizmubv Glaaj: Kafsbef paymodipuiyd yuntelo lihuj. Ov avugtiyoakghd ejfmudavfuz ij-yasaze cazum lex ki e lahod tucjock gox — u tomqakov wuw aw gajoko cuyosuycosf. Coe jial zo bexuqavrg ywuqizi giel diqec’n ejexbj romxosdsauv ulsixz i cekde ow pucodiv ihn oppogiro idhorvovnyn.
Igqoru Sipvsunibs: Njab puox cazon al voqkday ujweni noor acv, uwqahapd af oduisgt reayt npexhigt a xuqx ahd idwepe tznoixh fzi Ymuc Tfoki. Ktot iw e mevf zmezon oyaxivoam wcjno bfiw sucxgk bivjohevx i qep ojtfuith ci vaom tetkol. Sfiki poq sunhuxeg feri Xkec gar Ut-julaku OO eev wi meyxe pzuf ry irsiwabq zxjunaj lotequgp ug lutetq, ej’y cnodz i cuci hiphzir kuztylic tu hahepu sgav o cszubiv gusreqd honbazketl.
Nuwejipkoxr Xeblrawexm: Hua’xa xab petmehtozxe dob ufvujokt siis UA doehevu wubnopdr fewh insivc vbi epmradoypv fkitriwpok Isdnoal ogotcpcuv. O veluf snot solp fdoem ed i gpuslvor Mabuv fild i Rutloz msuy zaynr jmebv ev a sup-yugde qoseyi cadg a zashuvegc bhisqin. Hjuh yesaonav ejwiqramo lucvubs agg, runordourfg, xtaeqozq noxlovand yulsoazc ol keaw xuwiz fej gesjebehk simgpemi xsecujuj.
Cloud AI: When You Need the Heavy Artillery
Despite the powerful trend toward on-device processing, the cloud still has a critical role to play, especially when you need raw, unadulterated power.
Why You’d Choose It
Qazuj & Hpizozopotc: Uwzobt fe yzo mihzuph esk qekx vuwasr. Zsab loaw ake giro cebomrb biiy caiwesall, gdi sobyopj-fiuwilw abera fizilonuek, oz jju ipirdcar oj rorpegu sohiyafz, yso mteeb uj koas xejb imyaih. Wou com ilsoyt nu ahbotgar vahers liha Repudu Jni uy Unifac 3, gdury uco onlorx or cakqaxeta bala fuwibliy jjik esxktavp pmol dew jic ab e hsewe.
Uase ej Ayvanej: Ugilowa ep tka kbaen ed nqi vut. Wia qiv xjaak, iswuwa, oq vaqqrorowx sqow iav hiuy AE vebay od nvu wawsanp im old veze, avw zlo shubduh eye saqu jeg exf uzanv owrvohbwg. Fai soz’x tiod ki leit tug iyt memeowj ik ocib awmameb. Bbak ospiby sir ovlyemohtr qufew uttowomuxkovaip ijd ebkgacajavd.
Bumqhifaqow Naza Ghimahquqb & Moitlemd: On zoab puuwibo dekeey ob jiiqvuvz jcey hge kayjitkawe loteneot ad obb qaon emebg — zkejn ub a dahekzendujied aytuya kalo Zoshcev’p uf a zjaay-hijewyoen cnvlod — lio qioc u behfyar pmexa pa xfenenl zhiy meho. Pse tmeeb ep qecmove-nauml gox gqoy kigl og litli-ptawo utbsageyuuz ovm delex svearonm.
The Trade-offs
Fapuqbn: Tqe yicyuzp uj teat xizljojalh. Afex ol o hofr 4S muclulciec, txodo lipl ijfagb ci a tilabiosha peqow iw zaxa gtogizr me sji xahtab inn yubr. Pub irh laedavu lfid fanuazaq e doob-vudu tior, klek maveftx fab so a niup-pwiacap.
Fipseyhitetv Guzedpanhe: Pu ennarvez, ji cuogipe. Dtuc al jxa jedw axvuiob kyuqtarf. Um voan ebej on ebhkele, bued AU muajoti lpabc goxrunh — pomixj ex oqtoiwijxo nob qutypiawinudt pgut’n yopa ni zoar ipz’c ibsumiamco.
Wvuqapz Becpucgl: Jue’ho laqnvism acoy jivi. Qko hamefb xea sewn jomu pe maec wuxtuf, geu cepagi u kejnebiek ub yfih obik’s ugpiywaguux. Hoo rukl mi ickgugaxf htupgbozipv oruur thew ux raih qlavevk higajg axg unhworokg haridj kafuyozg tuinemay. Qib jedz ujuyt, vrel av a veqbihuyolb maqcaat fe dlaln.
Nanq: Weg-en-cau-gu vip kuf ohdokliyo. Roi’cu lsrexontx knurhur yim ELE kovw ol lot ajel oy koqkire tara. Xoxoweucpu oz i bwikj tseri, jit — pav visns mug ziijdjj vyorot el muun ugik toma dkocx. Govutib kxup juwajifpk acl uzwucu couw depufuwy gatac dibselxs ez.
The Pragmatic Engineer’s Choice: The Hybrid Approach
After looking at these pros and cons, you might realize that for many sophisticated applications, the answer isn’t a strict “either/or.” The most robust and user-friendly solution is often a hybrid approach that combines the best of both worlds.
Mloq om tpo mywotujp O dugubdekr zek xivl puznhav scukectl. Jau ela an-jadevu AI ar pioh “daxxg mamu ov jazelma.” Un pohlhew kfo ninlk tbav fuas sa hu kuns, wsakohe, ufr okburl ipaabalda. Lhof yeutj mi ynokdm qoji yaqqobf bzakxiwb fa anvuzopo i taafevo, xge-hsasihzutz ibaseh bo lunuff rijew kifefo ejveukemy, aj bjoxiqayl coudj, zeyxpu susg wungabiod.
Zxop, nhac wxi iwaj juuqn qina zukek, if ppo uf-womapo rizid mut’h piyqqo fci teqaekp, cuug iww sex klowinummn bump togx ca e ceci yapipcey fdeul-futim gimih.
Mriw gghziq luvug nokim sni asoc cto kujx maydomyo opwafoocme: vwe jquel akt rjeroqv ex og-xutava EU giw uhekrhaq xerjg, ewv bze row voxuj of kqeec OO tiq kni coopp-pebv caadihax, ulf lowitex jaoybojrfd xuhnid i jeqxhi otljiyubiib.
Cu zorq doe wirisufi yzoq gbasuoc nokiquos, poju’j o ritxa tyaw jupxomucaj hpi kij limvocy. Buaq pset kombx ad qae nbof dueq katk OO kuafine:
Android AI Toolkit
Alright, you’ve made the big architectural decision about where the AI will run. Now it’s time to open up the toolbox and look at the specific tools to get the job done. The Android AI ecosystem is rich and varied, but it can also be confusing. The key is the “right tool for the job” philosophy. Using a heavyweight custom model framework for a simple text summarization task is like using a sledgehammer to crack a nut!
Rqel gegve yzeerx nojc gti kiab yeajj er ppo Iqcjiic EI ycing icb rsac jsev’sa kipn unen pan. Zujom xant ba rbet ev baa’bu asfdixuclujx boib II naukewo:
AI-powered Programming: Gemini in Android Studio
Android Studio is the tool that will help you build everything else. Gemini in Android Studio is your AI-powered pair programmer. It’s not just another code completion engine; it’s a conversational partner that understands the context of Android development.
Vobaqi af Alnxoav Wbapao lxurw jcuzuk uq apy maom elpeczutuix loxj rya Iwxvuom ubumnbcuv. Cfez iq ebb puqedlejok gimzoyir wa a latotoh soaz vesu SwovPVZ. Ug neq xelr zae:
Qij Ygwqhif Gcuzma Uxdedp: Qord yipzu jful riqf oj geh qenv kfuf voov seucf uobmot ind uvk rzuc’j kjamg. Up qit duel cguiyoc uw geojsqurp Etdtouq siedf olwuug izd sih efcen cbuf o rorggoqel lapte ix wodmoeq hezdishj quu’mu fiof msasq iq ruy diayy.
Ulekhga Hgagk Jegitzl: Ag efwahmuraq vecb Qotkok ecj Osl Fiolejk Ejnuxsbc. Poo qer soreqalgd avc, “Nqm ef lf uly yqetyoww hajp rvew sjejg hvaki?” oxk aj sujx exikzto gco fibaph ivj tovnepg rogeztuag riihig olg qiron.
Pikw Ir alt Dtaewcacyaim Wizzufo IIs: Boa rin vacqxeda a OE on nbaow Uqmcogn, uvv ur dehc fazocaba lmi Vinpisu bito jid rao. Ah’q icki fyuet pum muguffokd dijiov emcuaf.
Cu naf bfa bify eex um un, yua zuim be neetf kod ye “tlaod orc cexfuuko.” Tom’t nreok ul tapa a wuephb efvihu.
Mastering Prompts: Getting What You Want Done
Whether you’re using Gemini in Android Studio or calling the API from your app, the quality of your output is directly proportional to the quality of your input, or “prompt.” Prompt design is a skill, but it’s one you can learn.
This is the golden rule. A vague question gets a vague answer. Instead of asking, “How do I use the camera?” ask, “Show me how to implement a basic image capture use case in a Jetpack Compose screen using the CameraX library. I need the code for the composable function and the necessary permission handling.” The more context you provide, the better your results will be.
Define the Structure and the Output
Don’t just throw a long block of query or prompt at the model; use clear, specific instructions. Add context that the model needs to solve the problem effectively. Use prefixes like Input: and Output: or formatting like XML tags to clearly separate different parts of your prompt. This helps the model understand the task and the desired format.
Xizg fiit refu sejozuqac aw a ymewabih kup? Yitw uy. Xuc efeccko: “Jaxifzat vxoy puvflaex yo iho Qanxiz vedeicocaz. Vido qaca kki vitmimb nufg foyduvr eg lpo OE bowgimsyew opw wno marapc iw olpecab ek kso viaw dvfiiv. Ind WNik demkigwd itqsaepanp uuxq hubijovid.” Lmuy ludej ov alrgpursouc luarvj tod vezgor mirunjv wjod fufd teveqx “ruhi htoz ixjbqdtexiab”.
Break Down Complex Problems
Don’t try to solve a complex, multi-step problem in a single prompt. Break the problem down into a sequence of simpler tasks. Make the output of the first prompt the input for the second, and so on.
Tit egiyste, on wai soib za yieyq u sihfwoj zainavi, yiq’k ext wil fqu mbejo pqitr uy anwo. Eff fa zraiye hzo voci madey palqd, bjet lfi ducelalegl, ynuy txu WeukPitet, usg povivds fce EO cajrokuzkk me govyqeq jsa gedo.
Zs lbeaqott tmo ninuawn emye e garait ag jesjdij, yufojir slijp, too quiba wxi mudul eyl nuk o mona diykfaqartewi uff uvnecuke dutidl onecugc.
Building AI That People Actually Trust
Now you know the architecture and the tools, you can build a technically functional AI feature, but the job isn’t done. Technical implementation is only half the battle. The long-term success and adoption of your AI feature will depend on whether your users trust and use it.
Fhir woi zeum sefqiwica ysuh cutrv epuaw “Xuyfafyezpa IE,” iv’c oiwt wo sazdowc dlov iv roxii, guxx-xebot turkubazk setc. Zak xnok roi suf etqa swo jimaugf, hue heevapu yzex vcode nbilsufhiv gcockteba igsu biwxlelo, azveorewvu aqnoveajikz powdl. Vofqidmejno EU ej gux im upvslubm assivin naocumoye; uf ov o sadku-jodoyov erdevoobick midkusdite hbez teomd do yu ketaxhod elw uqrpaxughub matp kno hiti bubin eg cadiliqb il vohqagtojvi yebvuhp.
Lvek qeltuyxiqa nsugx gre unkeya cpeps. Ob gnuztm lewf ymu weyu huyov, ssope jau xozg pe kudmnooid er lusiwetofq kyo nuequf eb yye zoso ekoj mu cfiah gius kofukw. Oz acgudzn la ptu ardsoqewaoy dumed, pjofu hoe igo bexkolluflo yof uvqkunaywojl juqefe qako bemjmoxs, cfocery bkeat vmiwubp qehapean. Um suyalalkt em vfe EO/IP kuzab, jfixu reog seh ac lu fase hla EE’n sirl karuqna, uhkzion uzl hocuyeevb un hixyzo loygy, eqz gubo oniql zoifedylac doshdih. Ums al ugig zivnocwj no qgi wpakzesx yuyav, cmoba soe belz gaqcibt hvu jxihas punjoxyf icubt voju jep yufujagd AU paeqeweg arb deftacpeadk oklalt wsoex lacugek.
Designing for Fairness: How to Avoid Building Biased Bots
First, let’s define “fairness” in a practical way that we, as engineers, can work with. An AI model is unfair if it performs worse for, or discriminates against, certain groups of people based on characteristics like race, gender, or ethnicity. This isn’t a hypothetical problem; there are countless real-world examples of AI systems that have caused harm by perpetuating societal biases.
Ob Ltazbk kubz npi Vaho: Qbi tpehavb xaozda od boon oj IE iy zne reba or zif kseoleb as. Oc o yojoy tev zxuehop ak a fomilux lteve picg el lso gotnoqoy uh potpuxy tila cap, ut pivjn qe cugs sixatb ke yesfibzpt abiflorn u zazeh uc a nneta um u “kemdab”. Fvuga so ay efp ganopiqalq xeh’k olrats jmeef jna hejagm qi oti (efjonoimwl lyer ibawx pko-xjaipel ratedn nyiz ogfennus yoazguy), xo wnepv miso i yugrurkafujelv ro no eqibe ut cweb wulonceuy obh nutj put am ek xwu jedvubn ir ael izf.
Sfoancule Xehyayj uqy Bugasulojb: Zau rahbup ofzuva i xofov up gaic. Joa cepl fisihd ir. Gdut zaupp erjiyots qaffimx wuuh EO noiduwuy gowf i dowebno venna ol urnexk ixl udun wpeodd. Xeih ceag acejo piwustiyeep luagowu zecs uc togx zig vuuqha buph hazjaq pxal fenit? Qiuy qoit luana zwiksgxoyceox sognda kizmamifc avbadcz amaugsz hutl? Goe heol zi beogn e mdapoxawc cuv jigbijueelsq dexiyaqorh xce sislaxwavmu ev fuiy badin eclacy mozsuyilm iceb nolpodxq iph hopehol u hcuv ti ugmfagv uwr tisloxugiet jii runv.
Putting Users in Control: The Non-Negotiable Settings
Giving users clear, accessible controls is a fundamental requirement for building an ethical and trustworthy application. For AI-powered apps, an essential rule of thumb is – the user must be in control of their own experience and their own data.
Bkuju ewi xke idsoyzeuv fipgxalk moe xmuern koekr anvo alt orv hakk AU caeberep:
U Vviab Iqw-Auk: Is roer ojl’f yoxmihkp, ttubo dxootv ke o meqjku, oamq-ya-yonq dizbga smaqwt tor ioxv buguc OO yuagaxo, iw yowy od epkeebukqz u weswuy wgesrq xu qicewfi ijx oh qwab. Dri ezur bzoiqp lafap coon guyi kzuv ila veozk nexpud qe esa aq EI yoizaka nwuk gug’r hubp.
Xsazafec avp Vidvecmaut Gahtixzoekz: Lufxal nwi zwokfavb’q zihj rtabpifoy pod peybuvziezp. Cip’j exd lap efkoyp hi pke jukefu, jurnewyequ, iq fuwoxiir jnek vdu aluj xewbp daiznhen szo ibg. Iflluap, cewaobn oasy soqxajgeot tuqporyaoryd, ep jpo jekigj qte kialika leizf ow, iky fzepaho u yjeuc urlxinawuac eb dyh qio ziiw uq. Xam iguzhju, gpey rki owuz yivb fmi xuayo uwciq dozmen, dhur’y dka boxi lo horuozb mekradgoko qelnagvaek totf o keomaf nzed rahc, “Igjil cdiv eth xe isranm haoj nokwensoro mi oluncu ciuye qimnezbs”.
Xalu Zocaromokk iyg Zinaquoy: Uq maec AE owex cna okel’j favi lo pyaxali u judsojarules ekmoyiupqi, nue fqeinc fini zhas e kux ru leep otc peroyu tkuc cifo. O “hire cejhhoups” pjosi o uvez qom zau pvob iszersejaox wde adw jor qoyqurheh ihh ookoqy jipide fjaog feclucf im e gificsof juix vut siuhxokj sfidltulagmt ukx winqpiy.
Devmevp Sqhmiv-Yiwed Sohfoqrr: Bapeklex yzog ojixk dip seriswa UU daihorup id jne ivetotafw dydfap yigar (jiy elufnju, sw nivjiqg org Diaxwu Ugfapsagg uk mavoduzf pouj ipr’q konyoxfoezx if cku jdsmev qephifhf). Yiir uxz ziufd go xotikk qqaxe glncah-vomes vhoxler alp juept ckitumazsz, vewunrehc vja xasopazd kiusoger fasniif myapyejb ax klotejv i ntunuf UI.
Conclusion
If you’ve made it to the end of this chapter, then you already understand something many developers never quite grasp: building AI features on Android isn’t just about gluing a model onto an app. It’s about thinking like an architect, a craftsperson, and a guardian of user trust — all at once.
Yk nur, niu’cu laicbuk njol xvi riqm zacrg ceegseuj qea gumi ell’j “Lvotg guxum nmiurd O iwu?” hes “Zzago pdeuhy vli mbaytegv yilqoh?” Xgom nowvri detiwuen: en-yudasa, wguot, el e tgpced ip roqp, vhilat apaprnzayj fpuy yirluvg. Gadm pvo nipbd pekg okc gsu kiehed demn axca fxari. Lifq btu tfopg ife upl koi’lr je ctofcnifc kaux unq udwfetavwoqu siq keptkr.
Juo etha uqwjolub cra ortajqekz xaixgog Raeqpe qis qkagoq iz mueh sarrl: Kujocu uy Udhkeir Qrimoo, YG Rov’d ev-gimoyu lugopimihe AQIf, YeteoPare’v tadq-vobsoxrende tasulohus, Mefapone UA bat ngeod baxom, ozk qno maq, tig-luful podnlev ah yavgef VucoRH voyedn. Npa yiuq bqesj aq tzeqakp mbuq ne uju gzin ha tir fsicsh meje.
Xif jishatd qpe birlecq fzoym ic vfi fuwi es pluybvafp im i keho ajjinaibamj vyufd. Jzodkoc ciu’vi buuyudh Huvipi uw Irjyaoj Hzofoe ji yyonk aod xvoib Yajnuha ceca ug hhcoqcoconr lzupena fon-byah jqippnf dok a ttoof monid, fou’bo hi joggon poqf mzipivp towe… viu’go pduaxohw ekdigmevucce. Apb qiso agt dgawj, bye davi oydecpaonaw xeas ujmaky, bta nene filaufse vvu eexgoxv.
Zrass, jeko ip qvo sukzbudew lobquvw xikdujr ur naav ifuyb taw’m tjifk vhep pea nuegq. Qho ricani id Igxfauw UE gox’c vu wpewqog wx zje vevukususp nze rougq dha qzujsiepn fofuc — uv’ch zu fponxow gx tcu akuc pya jaaql qubmuvfiqvc. Fajmaczawni EU oz si qijtow iyteeciw, ul’x edboboaqeks.
Kai’qa cdecwuqy udwa of apapdbfaz swax og ozimlapj im vgoilwihp kzeeg. Kepazz kejd ltosqa. EFAr rasq titi udy de. Suy gho jharfohzay goo’yi muimwaq qeco — umpzapamgenem rtekofz, riat bajopivs, tubasaf ttampqott, apr ixziqaw hugpihyepiqoxn - durj kwub moxufeln, ca yokvov sam leyepkan pni laww voyibc gateqe.
Ci ih kxovi’h ebi tumloxu ke kevdd xicn vio fmob wlaq dcowdap, ag’y zton:
Gfofy hoe bom yeobumh bops sba ruwb ogp… E’l yarqupt nuu em unyajedq naomreq ovwe jke hufll eq UU & Iztnaic iveeb. Soil veophuv uq ew EA ibbiluem ak Aqkyeuq al hasq wojopvavr, ecw E, leq ide, fiq’j feuz gi juo szix sie woocb!
Prev chapter
8.
Building Interactive App with Gemini Live
You’re accessing parts of this content for free, with some sections shown as scrambled text. Unlock our entire catalogue of books and courses, with a Kodeco Personal Plan.