I Replaced My “Dumb” Ask NHL With AI… and It Actually Slaps

A developer replaced a hardcoded NHL stats app with an AI-powered assistant using Google's Gemini API, allowing users to ask natural-language questions about hockey. The project demonstrates how to structure prompts and implement retry logic for reliable AI responses.

“Who are the top scorers this season?” Old app: ❌ hardcoded endpointsjust knows New app: ✅ AI that Yeah… we’re doing that. I turned my Ask NHL screen into a real AI-powered assistant using Google’s Gemini API — and it took way less code than you think 👀 Let’s build it together 🛠️ Most sports apps do this: But users don’t think in endpoints… they think in questions: 👉 “Who’s the best goalie right now?” 👉 “Who leads in penalty minutes?” 👉 “What happened in today’s games?” So I said… screw it. Let’s let AI handle it. We send a user’s question → to Gemini → get a clean NHL answer back. That’s it. But the 🔑 is how you structure it so it’s: If your prompt sucks… your AI sucks 🤷♂️ Here’s mine: private fun buildPrompt question: String : String = """ You are the "Ask NHL" assistant inside an NHL fan app. Answer the user's hockey question directly and concisely 2-4 sentences max . If the question isn't about hockey or the NHL, politely say you can only help with NHL topics. Question: $question""".trimIndent Here’s the full request flow with Retry/Backoff strategy — this is a must : / Talks to Gemini to answer free-text NHL questions from the "Ask NHL" screen. /class AskNhlRepository @Inject constructor private val geminiApi: GeminiApi, { / Sends question to Gemini and returns the generated answer text. Wrapped in a Result so the ViewModel can cleanly show an error state instead of crashing or leaking a raw exception into the UI layer. Retries transient failures rate limits, 5xx, network blips with exponential backoff before giving up — Gemini free-tier flash throws 503 "overloaded" fairly often, so one failed attempt isn't treated as a real failure yet. / suspend fun askQuestion question: String : Result<String { return try { val answerText = withRetry { val request = GeminiRequest contents = listOf GeminiContent parts = listOf GeminiPart text = buildPrompt question val response = geminiApi.generateContent apiKey = BuildConfig.GEMINI API KEY, request = request, // Gemini nests the answer text pretty deep: candidates 0 .content.parts 0 .text response.candidates ?.firstOrNull ?.content ?.parts ?.firstOrNull ?.text } if answerText.isNullOrBlank { Result.failure IllegalStateException "Gemini returned an empty response" } else { Result.success answerText.trim } } catch e: Exception { // Anything that escapes withRetry a non-retryable HTTP error, // or retries exhausted lands here as a Result.failure instead // of propagating up and crashing the caller. Result.failure e } } / Runs block up to maxAttempts times with exponential backoff between failures. maxAttempts is 2 not 3 : with a 20s per-attempt timeout see NetworkModule , 2 attempts caps the worst-case user wait around ~40s plus a small backoff delay. A 3rd attempt rarely rescues a request that's failed twice and just makes the user wait longer for the same likely outcome — not worth it for a chat-style UI on MVP. Only retries error types that are likely transient — a 401 or a malformed request will fail identically every time, so those are rethrown immediately instead of wasting an attempt on them. / private suspend fun <T withRetry maxAttempts: Int = 2, initialDelayMs: Long = 1 000, maxDelayMs: Long = 8 000, block: suspend - T, : T { var lastError: Throwable? = null repeat maxAttempts { attempt - try { // Success — return immediately, skipping any remaining attempts. return block } catch e: HttpException { // A 429 with a daily quota ID means retrying is pointless — // the limit doesn't reset for hours, not for a backoff delay. // Read the error body once here and bail immediately instead // of burning an attempt and the user's time on a retry that // cannot succeed. if e.code == 429 && e.isDailyQuotaExhausted { throw e } val retryable = e.code in setOf 429, 500, 502, 503, 504 if retryable throw e lastError = e } catch e: IOException { // No connection / timeout / DNS failure — always transient, // always worth a retry. lastError = e } // Exponential backoff: 1s, 2s, ... capped at maxDelayMs. // Jitter random extra delay up to 25% of backoff spreads out // retries from multiple concurrent requests so they don't all // retry in lockstep and cause another spike. val backoff = initialDelayMs 2.0.pow attempt .toLong .coerceAtMost maxDelayMs val jitter = Random.nextLong 0, backoff / 4 + 1 delay backoff + jitter } throw lastError ?: IllegalStateException "Gemini request failed with no captured error" } / Wraps the raw user question with instructions so Gemini stays on-topic and keeps answers short enough to fit the UI. / private fun buildPrompt question: String : String = """ You are the "Ask NHL" assistant inside an NHL fan app. Answer the user's hockey question directly and concisely 2-4 sentences max . If the question isn't about hockey or the NHL, politely say you can only help with NHL topics. Question: $question """.trimIndent }/ Checks whether a 429's error body indicates the daily quota RPD request per minute was exhausted, as opposed to a per-minute rate limit RPM , which IS worth retrying. Google's error body includes a "quotaId" string that contains "PerDay" for daily quota violations — that's the signal we key off of. errorBody can only be read once per HttpException, so this reads and discards it; callers shouldn't try to read the body again afterward. /private fun HttpException.isDailyQuotaExhausted : Boolean { val body = response ?.errorBody ?.string ?: return false return body.contains "PerDay", ignoreCase = true }/ Converts a failure from askQuestion into copy that's safe to show a user. Lives near the repository rather than in the ViewModel, since "what does this error mean to a user" is closer to a networking concern than a UI-state concern. /fun Throwable.toAskNhlUserMessage : String = when this { is HttpException - when { code == 429 && isDailyQuotaExhausted - "Ask NHL has hit its free daily limit — check back tomorrow " code == 503 || code == 429 - "NHL brain is a little overloaded right now — give it a sec and try again." code == 401 || code == 403 - "Something's wrong on our end, not you. We're on it." else - "Couldn't get an answer for that one. Try rephrasing?" } is IOException - "Looks like you're offline — check your connection and try again." else - "Something went sideways. Try again?"} / Retrofit interface describing the Gemini REST endpoint. Retrofit generates the actual HTTP implementation of this interface at runtime — you never implement it yourself, just declare the shape of the request/response. Docs: https://ai.google.dev/api/generate-content /interface GeminiApi { / @POST marks this as a POST request. The {model} in the URL is filled in by @Path "model" below — Retrofit substitutes it before sending. @Header x-goog-api-key sends the API key in the request header for secure authentication, as required by the Gemini REST API. / @POST "v1beta/models/{model}:generateContent" suspend fun generateContent // Switched from gemini-3.5-flash 20 RPD on free tier to // gemini-3.1-flash-lite, which carries a much more generous free-tier // quota on this project — 500 RPD / 15 RPM vs. 20 RPD / 5 RPM. // Confirmed via live AI Studio rate-limit dashboard, not assumed. @Header "x-goog-api-key" apiKey: String, @Path "model" model: String = "gemini-3.1-flash-lite", @Body request: GeminiRequest, : GeminiResponse}/ Top-level request payload sent to Gemini. @JsonClass generateAdapter = true tells Moshi to generate JSON serialization code for this class at compile time — fast, no reflection. Make sure to -keep your models in proguard for R8 Minification purposes. /@JsonClass generateAdapter = true data class GeminiRequest // Gemini expects the prompt wrapped in a list of "turns" to support // multi-turn conversations. We only ever send a single turn. val contents: List<GeminiContent , val generationConfig: GeminiGenerationConfig = GeminiGenerationConfig , / A single turn in the conversation. /@JsonClass generateAdapter = true data class GeminiContent // A turn is split into "parts" text, image, etc . We only ever send // one text part. val parts: List<GeminiPart , // "user" = our prompt. Gemini also accepts "model" for prior AI // replies in real multi-turn chat — we're not doing that, so this is // always "user". val role: String = "user", / A single chunk of content within a turn — just the raw prompt text. /@JsonClass generateAdapter = true data class GeminiPart val text: String, / Controls HOW Gemini generates its response not what it's asked . - temperature: randomness/creativity 0.0 = deterministic, 1.0 = very random . Kept low since we want factual, consistent hockey answers. - maxOutputTokens: hard cap on response length. With thinking disabled see thinkingConfig below , this budget is now spent entirely on the visible answer, so 400 comfortably covers a 2-4 sentence response. - thinkingConfig: disables Gemini's internal reasoning pass. Without this, gemini-3.5-flash a "thinking" model spends a chunk of maxOutputTokens on invisible reasoning before writing the actual answer — in testing this ate 382 of our 400 token budget, leaving only 14 tokens for the real response and cutting it off mid-sentence finishReason: MAX TOKENS . Simple factual Q&A like "who won X in Y" doesn't need deep reasoning, so turning it off is free speed + a meaningful token/cost savings — worth it for MVP on free tier. /@JsonClass generateAdapter = true data class GeminiGenerationConfig val temperature: Double = 0.4, @Json name = "maxOutputTokens" val maxOutputTokens: Int = 400, val responseMimeType: String = "text/plain", val thinkingConfig: GeminiThinkingConfig = GeminiThinkingConfig , / thinkingBudget = max tokens the model may spend on internal reasoning before producing the visible answer. 0 disables thinking entirely. /@JsonClass generateAdapter = true data class GeminiThinkingConfig @Json name = "thinkingBudget" val thinkingBudget: Int = 0, / Top-level response from Gemini. /@JsonClass generateAdapter = true data class GeminiResponse // Nullable + defaulted because a failed/blocked request can come back // with no candidates at all e.g. safety filter tripped . "candidates" // is a list because Gemini can return multiple response variants — we // only ever look at the first one. val candidates: List<GeminiCandidate ? = null, / A single generated response option. /@JsonClass generateAdapter = true data class GeminiCandidate // Mirrors the request's Content/Part shape. val content: GeminiContent? = null, // Why generation stopped: "STOP" normal , "MAX TOKENS" cut off , // "SAFETY" blocked , etc. Not branched on yet, but useful later if // you want different handling for truncated vs. blocked vs. success. @Json name = "finishReason" val finishReason: String? = null, fun onQuestionSubmitted question: String { if question.isBlank return // Start cycling "thinking" messages immediately, before the network // call even starts — this also sets the initial Loading state. startThinkingTicker viewModelScope.launch { geminiRepository.askQuestion question .fold onSuccess = { answer - // Stop the ticker now that we have a real result — // otherwise it'd keep overwriting uiState every 500ms // even after we've already set Success. thinkingTickerJob?.cancel uiState.value = AskNhlUiState.Success question, answer }, onFailure = { throwable - thinkingTickerJob?.cancel // toAskNhlUserMessage converts the raw exception HTTP // code, IOException, etc into copy that's safe to show // a user — e.g. a 503 becomes "overloaded, try again" // instead of a stack trace or raw error string. uiState.value = AskNhlUiState.Error throwable.toAskNhlUserMessage }, } } We’re not just building AI… we’re building experience . Simple. Clean. No overengineering. @Composablefun AskNhlScreen navController: NavController, viewModel: AskNhlViewModel = hiltViewModel { val uiState by viewModel.uiState.collectAsState var inputText by remember { mutableStateOf "" } val keyboardController = LocalSoftwareKeyboardController.current val focusManager = LocalFocusManager.current fun submit { if inputText.isNotBlank { viewModel.onQuestionSubmitted inputText keyboardController?.hide focusManager.clearFocus } } fun onPromptSelected prompt: String { inputText = prompt viewModel.onQuestionSubmitted prompt } NhlMainScaffold topBar = { MultiWidgetOnClickToolBar handleBackNavigation = { navController.popBackStack }, title = "ASK NHL Beta ", actions = {}, }, navController = navController, bottomBarType = BottomBarType.AD BANNER, { padding - Column modifier = Modifier .fillMaxSize .padding padding .imePadding , { LazyColumn modifier = Modifier .weight 1f .fillMaxWidth .padding horizontal = 16.dp , verticalArrangement = Arrangement.spacedBy 8.dp , { when val state = uiState { is AskNhlUiState.Idle - { item { Text text = "WHERE DO YOU\nWANT TO START?", fontFamily = FontFamily Font R.font.nhlitalic , fontWeight = FontWeight.Bold, fontSize = 36.sp, lineHeight = 32.sp, modifier = Modifier.padding top = 26.dp , Spacer Modifier.height 16.dp } items viewModel.suggestedPrompts { prompt - PromptChip prompt = prompt, onClick = { onPromptSelected prompt } } } is AskNhlUiState.Loading - { item { CircularProgressIndicator } } is AskNhlUiState.Success - { item { Spacer Modifier.height 16.dp Text text = state.question, style = MaterialTheme.typography.bodyLarge Spacer Modifier.height 8.dp Text text = state.answer, style = MaterialTheme.typography.titleLarge } } is AskNhlUiState.Error - { item { Text text = "Something went wrong: ${state.message}" } } } } NhlOutlinedText value = inputText, onValueChange = { inputText = it }, label = "Ask about the NHL", trailingIcon = { if inputText.isNotBlank { TextButton onClick = { submit } { Text text = "Go", color = Color.Black } } }, modifier = Modifier.padding all = 16.dp , keyboardOptions = KeyboardOptions imeAction = ImeAction.Done , keyboardActions = KeyboardActions onDone = { submit } , shape = RoundedCornerShape 28.dp , } }}@Composableprivate fun PromptChip prompt: String, onClick: - Unit { SuggestionChip onClick = onClick, label = { Text text = prompt, style = MaterialTheme.typography.titleLarge, modifier = Modifier.padding all = 8.dp , }, shape = CircleShape, border = SuggestionChipDefaults.suggestionChipBorder enabled = true, borderWidth = 1.dp, , colors = SuggestionChipDefaults.suggestionChipColors containerColor = Color.Transparent, , Spacer Modifier.height 12.dp } Let the AI shine ✨ This isn’t just a feature… it’s a shift in how we build apps . Instead of: “What endpoints do I need?” Start asking: “What questions do my users want answered?” If you’re building Android apps and NOT experimenting with AI yet… You’re already behind 👀🔥 Grab the beta/MVP on GooglePlay https://play.google.com/store/apps/dev?id=7614313297301862853&hl=en US . A sleek Android experience crafted with: 👨👨👦👦 Join the fun Android group on ➜ LinkedIn https://www.linkedin.com/groups/14584550/ Best, RAFA ROZAY 🍾 🧠🔥 I Replaced My “Dumb” Ask NHL With AI… and It Actually Slaps 🤯🏒 https://blog.stackademic.com/i-replaced-my-dumb-ask-nhl-with-ai-and-it-actually-slaps-85e83a011dfa was originally published in Stackademic https://blog.stackademic.com on Medium, where people are continuing the conversation by highlighting and responding to this story.