<?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 Waseem Syed on Medium]]></title>
        <description><![CDATA[Stories by Waseem Syed on Medium]]></description>
        <link>https://medium.com/@waseems0605?source=rss-543f42d7bfd3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*YuiG94aHAP-FC7OKGs4r1g@2x.jpeg</url>
            <title>Stories by Waseem Syed on Medium</title>
            <link>https://medium.com/@waseems0605?source=rss-543f42d7bfd3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 18:56:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@waseems0605/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[Building StockSense: Wiring an LLM to Live Market News for Traceable Sentiment]]></title>
            <link>https://medium.com/@waseems0605/building-stocksense-wiring-an-llm-to-live-market-news-for-traceable-sentiment-03ec3f7d87a6?source=rss-543f42d7bfd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/03ec3f7d87a6</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[web-scraping]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[software-architecture]]></category>
            <dc:creator><![CDATA[Waseem Syed]]></dc:creator>
            <pubDate>Thu, 20 Aug 2026 02:33:46 GMT</pubDate>
            <atom:updated>2026-08-20T02:33:46.731Z</atom:updated>
            <content:encoded><![CDATA[<h4>Concurrent I/O, structured LLM outputs, and grounding a model in real data. A learning project, not financial advice.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qcie69EeMURiZFUboxMMJg.png" /><figcaption>StockSense — Recommendation view</figcaption></figure><p>A couple of years ago a friend and I kept hitting the same wall: there’s no shortage of stock commentary, but no way to say <em>“only read the sources I trust, and tell me what they add up to.”</em> We wanted to define the trusted set once, and have something else do the reading.</p><p>StockSense is that idea picked back up. It takes a watchlist, reads the recent news for each ticker, and returns a sentiment score with a rationale — grounded in the actual reporting, and traceable back to the article it came from.</p><p>The model scores and explains. It doesn’t pick the label, doesn’t supply the links, and doesn’t get to imply the data is fresher than it is.</p><h3>What it actually does</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*vcifvvKD_aXH3HOsIuYSag.gif" /><figcaption>Analyzing a four-ticker watchlist</figcaption></figure><p>Two screens.</p><p><strong>Portfolio</strong> lists holdings with share counts, cost basis, current value, and gain/loss, plus rolled-up totals. Prices are live; the holdings themselves are sample data.</p><p><strong>Recommendations</strong> is the real feature. Type a watchlist — AAPL, MSFT, NVDA — or send your whole portfolio over in one click. Each ticker comes back as a card with:</p><ul><li>a <strong>sentiment score</strong> from 1–100, with a colored bar</li><li>a <strong>recommendation badge</strong> on a seven-point scale, from Strongly Bearish to Strongly Bullish</li><li>a <strong>few-sentence rationale</strong> citing the specific developments behind the score</li><li>the <strong>current price</strong> and day change</li><li><strong>links to the actual articles</strong> the judgment was based on, with headlines and publishers</li><li>a <strong>freshness indicator</strong> — whether this was computed live or served from cache, and when</li></ul><p>Tickers with no recent news still come back, marked “No Data,” so the response is always one row per ticker you asked for.</p><h3>How it works</h3><p>Five layers, each doing one job.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Yk5O6G4m_JMu1gurGQF9yA.png" /><figcaption>StockSense request pipeline diagram</figcaption></figure><h4>News fetch</h4><p>The whole watchlist arrives as one request and leaves as one JSON response, however many tickers were in it. Inside, the feed is queried per ticker for recent coverage — headline, publisher, timestamp, and article link.</p><p>Notably, not the article body. Most news APIs return metadata and a one-line teaser, which is nowhere near enough for a model to reason from. Closing that gap is the next layer’s job.</p><p>The interface here is deliberately narrow: give it a ticker, get back a list of articles. Nothing downstream knows or cares which provider answered, so swapping the feed is a configuration change.</p><h4>Cache layer</h4><p>Sits <em>beside</em> the pipeline, not inside it — every fetch and every scrape checks it first. A hit skips the external call entirely; a miss falls through, fetches live, and stores the result. Two kinds of entries live here on a short TTL: news keyed by ticker, and article text keyed by URL. The second matters more than it looks — two tickers citing the same article share one entry instead of scraping it twice.</p><h4>Scraper</h4><p>A feed’s article description is usually one sentence. To ground the model in the actual reporting, StockSense fetches and parses the source pages for every article it retrieves.</p><p>This is the one layer that runs <strong>concurrently</strong> — it’s I/O-bound on a dozen unrelated web servers, each taking as long as it takes. Every ticker’s articles get flattened into one job list and dispatched together through a thread pool:</p><pre>with ThreadPoolExecutor(max_workers=SCRAPE_MAX_WORKERS) as executor:<br>    futures = {executor.submit(scrape_article, a[&quot;article_url&quot;]): a for a in jobs}<br>    for future in as_completed(futures):<br>        text, was_cached = future.result()<br>        if text:<br>            futures[future][&quot;full_text&quot;] = text</pre><p>Threads, not asyncio — the underlying extraction library and its HTTP fallback are blocking, synchronous calls with no async-native API, so there&#39;s no coroutine to await. Threads are what actually let several slow downloads run at once.</p><p>How much text this layer recovers turned out to matter more than I expected. Scraping only the first couple of articles per ticker left the model with two full articles and three one-line summaries — thin material for a multi-sentence rationale, and it filled the gaps from its own pretrained knowledge of these companies. Scraping everything raised source text from roughly 8,000 to 19,500 characters per ticker and largely stopped that.</p><p>Fragility is assumed, not handled as an edge case. Any scrape can fail — paywall, timeout, bot detection, an unfamiliar page layout — so scrape_article() returns None rather than raising, and the model falls back to the feed&#39;s short description for that article. Iterating with as_completed() means one hung request never holds up results that already finished.</p><h4>LLM reasoning layer</h4><p>One call scores the entire watchlist — not one call per ticker. The model is asked for exactly two things: a 1–100 score, and a multi-sentence rationale grounded in the news it was shown. It’s never asked to invent the recommendation label or to cite its own sources; both come from the layer below.</p><p>Output is schema-constrained — structured outputs or function calling, same schema either way, so a malformed response isn’t a failure mode you have to write defensive parsing for.</p><p>temperature=0 and a fixed seed handle a subtler problem: the same news scored twice shouldn&#39;t produce two different numbers. That&#39;s sampling noise, not signal. Removing it doesn&#39;t freeze the judgment — new articles still move the score, as they should.</p><h4>Deterministic decision layer</h4><p>Two things happen here that the model never touches:</p><pre>_RECOMMENDATION_THRESHOLDS = [<br>    (90, &quot;Strongly Bullish&quot;), (75, &quot;Bullish&quot;), (60, &quot;Slightly Bullish&quot;),<br>    (41, &quot;Neutral&quot;),<br>    (26, &quot;Slightly Bearish&quot;), (11, &quot;Bearish&quot;), (0, &quot;Strongly Bearish&quot;),<br>]</pre><p>The label is a lookup against the score, so the badge can never contradict the bar next to it. And the sources list is rebuilt from the article metadata the scraper already fetched, not parsed out of the model&#39;s reply — early on, the model would occasionally paraphrase a citation into a URL that didn&#39;t resolve. The fix wasn&#39;t a better prompt. It was removing the decision from the model entirely.</p><p>The same layer verifies the rationale. Reading my own output, I found figures that appeared in no article — a “$678 billion backlog,” an “84%” growth rate — sitting in prose otherwise grounded in real reporting. The result is indistinguishable from grounded analysis unless you check every number. So now the code checks: each figure in a rationale must appear in the source text, and any sentence citing one that doesn’t is dropped before the response is returned. Instructing the model not to do it took untraceable figures from 38% to 10%. Verifying in code took the rest.</p><h3>Known limitations</h3><ul><li><strong>Scraping is fragile by nature</strong> — paywalls, bot detection, and layout changes mean some articles never yield full text. Production means a licensed provider with an SLA.</li><li><strong>Figures are verified, attribution isn’t.</strong> Every number is checked against the source text and dropped if absent. A real number attached to the wrong subject still passes — catching that needs claim-level verification, not string matching.</li><li><strong>No publisher allowlist yet.</strong> The scraper takes any URL the feed returns. This one bothers me most, since curating trusted sources was the original point.</li><li><strong>No persistence</strong> — the cache is in-process memory, cleared on restart.</li><li><strong>Not a trading signal</strong> — just a model’s read of recent public news.</li></ul><h3>Built with</h3><p>Python and Flask on the backend. Vanilla HTML, CSS, and JavaScript on the frontend — no framework, no build step. newspaper3k handles article extraction, with a plain requests fallback.</p><p>This build uses Polygon.io for market news and OpenAI’s gpt-4o-mini for scoring, but both sit behind narrow interfaces — the pipeline treats them as a news source and a reasoning model, not as specific vendors.</p><p>The frontend came together fast with AI-assisted tooling, which closed the gap between “it works” and “I can show someone this.”</p><p>StockSense is on <a href="https://github.com/wsyed1/stock-sense-ai">GitHub</a> — MIT licensed, not a trading tool, not financial advice. A learning project that happens to be a real one.</p><p>If you’re building something similar, I’d love to hear how you approached it. Fork it, break it, tell me what you’d have done differently — that’s the best thing that could come of writing this down.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03ec3f7d87a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Voice AI Pipeline: On-Device and Cloud STT in Practice]]></title>
            <link>https://medium.com/@waseems0605/building-a-voice-ai-pipeline-on-device-and-cloud-stt-in-practice-f6eab68c41eb?source=rss-543f42d7bfd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6eab68c41eb</guid>
            <category><![CDATA[multimodal-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[speech-recognition]]></category>
            <dc:creator><![CDATA[Waseem Syed]]></dc:creator>
            <pubDate>Sat, 01 Aug 2026 04:27:22 GMT</pubDate>
            <atom:updated>2026-08-01T04:27:22.140Z</atom:updated>
            <content:encoded><![CDATA[<h4>A technical walkthrough of on-device vs cloud STT, multi-turn conversation, multilingual support, and TTS voice selection — built end to end.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qv0t0phYoPd2hxw7BFzKEw.png" /><figcaption>AI generated</figcaption></figure><p>Voice is a better interface than typing for a surprising number of use cases. The challenge isn’t the idea — it’s the pipeline. Getting audio in, getting intelligence out, and making the response feel like a conversation rather than a form submission is where the real engineering lives.</p><p>This post walks through two working approaches to that pipeline. Not a comparison of vendors — a comparison of architectural tradeoffs. Where each approach shines, where it doesn’t, and what the scenarios actually revealed when we tested them.</p><blockquote>⚠️ Note: The medical scenario is purely a demo to make the conversation realistic and multi-turn. The patterns explored here may be relevant to other domains where voice is a natural interface.</blockquote><h3>What’s in Scope</h3><h4>Built and working:</h4><ul><li>Two STT approaches — on-device (Live Mode) and cloud via Whisper</li><li>Real-time transcription and pause detection</li><li>Multilingual support and code-switching</li><li>Multi-turn conversation with GPT-4o</li><li>Voice selection with OpenAI TTS — friendly, warm, soft</li><li>Mock patient context with allergy and medication awareness</li></ul><h4>Not built — shown in architecture only:</h4><ul><li>RAG, structured database lookup, or function calling</li><li>Security layer — auth, rate limiting, audit logs, data storage policies, retention</li></ul><h3>System Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8A98wcMn-yKMWYvMX6WofA.png" /><figcaption>Two approaches · One shared reasoning and response pipeline</figcaption></figure><p>The diagram shows two modes converging into one shared pipeline. Each layer is independently replaceable — that’s the architectural principle worth carrying forward.</p><p><strong>Audio capture</strong> — Precision Mode records audio and passes it as a file or bytes to the backend. Live Mode transcribes in real time on-device — no audio file, no network call for STT. Two different paths to the same destination: text.</p><p><strong>STT component</strong> — where the modes diverge. Cloud STT offers broader language support and higher accuracy on diverse speech. On-device STT is real time with no audio leaving the device. Once text is produced, the pipeline is identical.</p><p><strong>Reasoning layer</strong> — both modes converge here. Text, conversation history, and context are assembled into a prompt and passed to the LLM. This is also where RAG and a knowledge database would slot in — at the prompt assembly step, without touching any other layer.</p><p><strong>TTS and playback</strong> — the response is converted to audio and streamed back. No file stored. Both modes share the same interrupt handler — user speaks, audio stops, new input takes over. The loop repeats.</p><h4>Live Mode — On-Device STT</h4><p>Client (ex: iOS) transcribes speech in real time as the user speaks — no audio file, no network call for transcription. Text arrives instantly:</p><pre>recognitionTask = recognizer?.recognitionTask(with: request) { result, _ in<br>    if let text = result?.bestTranscription.formattedString {<br>        self.transcript = text<br>    }<br>}</pre><p>Pause detection fires the backend call after 1.5 seconds of silence:</p><pre>silenceTimer?.invalidate()<br>silenceTimer = Timer.scheduledTimer(withTimeInterval: 1.5, repeats: false) { _ in<br>    self.sendToBackend(text: self.transcript)<br>}</pre><p>The experience is immediate — text appears as you speak, the conversation flows naturally. For standard English in a quiet environment, this feels surprisingly close to a real conversation.</p><h4>Precision Mode — Cloud STT (Whisper)</h4><p>Client records and passes the audio file or bytes to the backend. The tradeoff is visible — no live transcription during recording. The transcript appears after processing. The added latency is the cost of accuracy and broader language range.</p><pre>transcript = client.audio.transcriptions.create(<br>    file=audio_buffer,<br>    model=&quot;whisper-1&quot;,<br>    response_format=&quot;verbose_json&quot;<br>)</pre><p>The tradeoff is visible — no live transcription during recording. The transcript appears after processing. The added latency is the cost of accuracy and broader language range.</p><blockquote>The transcription landscape is moving quickly — OpenAI released gpt-transcribe in July 2026 around the time this article was published, claiming significant improvements. For this POC we kept whisper-1 as it’s validated and stable. Personally, I’m excited about where this space is heading. The core pipeline remains relevant regardless of which models emerge.</blockquote><h3>In Practice — Four Scenarios</h3><h4>Scenario 1 — Live Mode: Natural Conversation</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FOpey_mo6xXI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FOpey_mo6xXI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FOpey_mo6xXI%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/91873a6bcaa7757859f24d415a00eace/href">https://medium.com/media/91873a6bcaa7757859f24d415a00eace/href</a></iframe><p>The conversation flows naturally — identity verification, symptom discussion, medication guidance — each turn building on the last without the user having to repeat context. The user mentions ibuprofen for a headache and cough. The AI flags it — given the Metformin history, ibuprofen can affect kidney function. It then references the penicillin allergy unprompted when asked.</p><p>The context comes from a mock patient profile in the system prompt. That’s where a knowledge retrieval layer — RAG, structured database lookup, or function calling — would make this genuinely useful in production.</p><p><strong>Multi-Turn Conversation</strong></p><p>Session history kept in memory, full context rebuilt on every turn:</p><pre>completion = client.chat.completions.create(<br>    model=&quot;gpt-4o&quot;,<br>    temperature=0,<br>    messages=[<br>        {&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: SYSTEM_PROMPT},<br>        *get_history(session_id),<br>        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: user_text}<br>    ]<br>)</pre><p>The system prompt is where patient context lives — and where RAG slots in for production without touching anything else:</p><pre>SYSTEM_PROMPT = f&quot;&quot;&quot;<br>You are a voice AI assistant.<br>Patient context: {MOCK_PATIENT}<br>- Ask clarifying questions before advising<br>- Never diagnose or prescribe controlled substances<br>- Always recommend urgent care when appropriate<br>- Keep responses concise — this is a voice interface<br>- &lt;Add here&gt;<br>&quot;&quot;&quot;</pre><p><strong>TTS — Voice Selection</strong></p><p>OpenAI offers several built-in voices — refer to <a href="https://platform.openai.com/docs/guides/text-to-speech">OpenAI’s TTS documentation</a> for samples. For this POC, shimmer was used — soft and gentle, appropriate for a sensitive context. Voice is a product decision, not an afterthought.</p><pre>tts_response = client.audio.speech.create(<br>    model=&quot;tts-1&quot;,<br>    voice=&quot;shimmer&quot;,<br>    input=response_text,<br>    response_format=&quot;mp3&quot;<br>)</pre><h4>Scenario 2 — Live Mode: English + Spanish Code-Switching</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FqRQtwgauFmQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FqRQtwgauFmQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FqRQtwgauFmQ%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/58664bd42b958522bbe6aec91099c56d/href">https://medium.com/media/58664bd42b958522bbe6aec91099c56d/href</a></iframe><p>The input was:<br> <em>👤“I have a bad cough and headache since yesterday. También tengo fiebre y me siento muy cansado. No sé qué medicina tomar.”</em></p><p>What Live Mode transcribed:<br> <em>👤“I have a bad cough and headache since yesterday the think of me Siento Muy Cando no”</em></p><p>The English portion came through cleanly. The moment the user switched to Spanish, the on-device transcription lost the thread entirely. The reasoning layer will often try to compensate — inferring from partial context — but it’s reasoning from a broken input. In a high-stakes context, that gap matters.</p><h4>Scenario 3 — Precision Mode: English + Spanish Code-Switching</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fm8ha06hdK_E%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2Fm8ha06hdK_E&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fm8ha06hdK_E%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e3b3965da369e4d6f068f96d0bc77bdc/href">https://medium.com/media/e3b3965da369e4d6f068f96d0bc77bdc/href</a></iframe><p>👤 <em>“I have a bad cough and headache since yesterday. También tengo fiebre y me siento muy cansado. No sé qué medicina tomar.”</em></p><p>What Precision Mode transcribed:<br> 👤 <em>“I have a bad cough and headache since yesterday. También tengo fiebre y me siento muy cansado. No sé qué medicina tomar.”</em></p><p>Same phrase. Different result. The transcript is accurate — exactly what was spoken, English and Spanish both captured cleanly. The text arrives after recording rather than in real time, but the input the reasoning layer receives is complete and correct. For multilingual users this isn’t a marginal improvement — it’s the difference between a system that works reliably and one that works most of the time.</p><h4>Scenario 4 — Precision Mode: Extended Language Range</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FhyYCWf-e-sI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FhyYCWf-e-sI%3Fsi%3DJ3Y5JqfJfbcDP9jX&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FhyYCWf-e-sI%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/36bfd03e35dbf8043976f2f66a74297b/href">https://medium.com/media/36bfd03e35dbf8043976f2f66a74297b/href</a></iframe><p><em>👤 “</em><strong><em>Swahili:</em></strong><em> Nina hōmanāni nagisikia vibāya sana. Nina kikohosina maumivu jakichua. Sīdu’i dawā pāni nitu mie.”<br> * (</em><strong><em>English translation:</em></strong><em> I have a bad cold and feel terrible. I have a cough and a sore throat. I don’t know what medicine to take.)<br> 🤖 “</em><strong><em>Swahili:</em></strong><em> Pole kusikia hivyo, John. Kwa kuwa unatumia Metformin kwa ajili ya kisukari, ni muhimu kuwa makini na dawa unazochukua. Je, umewahi kujaribu kupumzika au kunywa vinywaji vya moto?”<br>* (</em><strong><em>English translation:</em></strong><em>I’m sorry to hear that, John. Since you are taking Metformin for diabetes, it is important to be careful about the medications you take. Have you tried resting or drinking hot beverages?)</em></p><p>The transcription was accurate and the assistant responded directly to what was said in Swahili — the conversation felt natural on both ends. For a user most comfortable expressing symptoms in their native language, that matters enormously.</p><p>Swahili is spoken by an estimated 100–200 million people across the world. The cloud STT layer handles it natively — entirely in Swahili, no switching required.</p><p>The choice of STT layer has real implications for who your product actually serves. A pipeline that works beautifully for some users and silently fails for others is worth thinking about carefully from the start.</p><h3>Precision Mode vs Live Mode</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*g8_ApvjLRaXvAp0thfPcFw.png" /><figcaption>Precision Mode Vs Live Mode</figcaption></figure><p>Observations from building and testing this POC. Independently verify claims against latest provider documentation for your specific use case.</p><h4>Evals</h4><p>Testing was manual. The code-switching demos provided a natural WER comparison — same phrase, two modes, visibly different accuracy.</p><p>For production, worth thinking about:</p><ul><li><strong>STT accuracy</strong> — WER across your user demographics, languages, and accents</li><li><strong>LLM quality</strong> — guardrail adherence, context usage, escalation judgment</li><li><strong>Latency</strong> — across multi-turn sessions</li></ul><p>Each layer can be tested independently</p><h4>Security Considerations</h4><ul><li>On-device audio never leaves the device — a meaningful privacy advantage</li><li>Patient data has regulatory implications — worth understanding before you build, not after</li><li>LLM guardrails belong in the system prompt from day one</li><li>Auth, rate limiting, and data retention are production concerns — design for them early</li></ul><h3>Closing Thoughts</h3><p>Can a voice interface feel like a real conversation? That question drove this POC — and the answer is closer than expected. The two modes aren’t competing — they’re complementary: Live Mode for standard English, real-time feel, and privacy; Precision Mode for multilingual reach and accuracy on diverse speech. The voice AI space is evolving fast, and the composed, layered approach is designed to absorb that change rather than resist it. Voice is one of the most human ways to communicate — and we’re just getting started.</p><p>Thanks for reading — would love to learn more about what you’re building. 🙏</p><p><a href="https://github.com/wsyed1/ai-voice-kit">GitHub - wsyed1/ai-voice-kit</a></p><p><em>Built with Python, Flask, OpenAI Whisper, GPT-4o, OpenAI TTS, Swift, AVFoundation.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6eab68c41eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AppFabric Mobile: Intuit’s Unified Mobile Platform]]></title>
            <link>https://medium.com/intuit-engineering/appfabric-mobile-intuits-unified-mobile-platform-fef2db3910b2?source=rss-543f42d7bfd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/fef2db3910b2</guid>
            <category><![CDATA[intuit]]></category>
            <category><![CDATA[developer-productivity]]></category>
            <category><![CDATA[mobile-engineering]]></category>
            <category><![CDATA[mobile-platforms]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <dc:creator><![CDATA[Waseem Syed]]></dc:creator>
            <pubDate>Fri, 01 Oct 2021 20:44:11 GMT</pubDate>
            <atom:updated>2022-02-02T00:28:52.079Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*55A94rGEUIpy-3OZ" /></figure><p><em>This blog post is co-authored by </em><a href="https://www.linkedin.com/in/waseem-syed-731a39142/"><em>Waseem Syed</em></a><em>, Staff Software Engineer, </em><a href="https://medium.com/u/b89f955671ac"><em>Muralidhar</em></a><em>, Group Development Manager, and</em><a href="https://medium.com/u/2539c27e8aee"><em> Kerry Kartchner</em></a><em>, Principal Engineer in the AppFabric Mobile team at Intuit.</em></p><p>Intuit’s mission is to power prosperity around the world. Our mobile applications — including <a href="https://turbotax.intuit.com/">TurboTax, </a><a href="https://quickbooks.intuit.com/">QuickBooks</a>,<a href="https://www.mint.com/"> Mint</a>, and<a href="https://www.creditkarma.com/"> Credit Karma</a> — serve approximately 100M customers. To deliver awesome mobile experiences for consumers, small businesses, and self-employed individuals, Intuit has built a robust, unified developer platform for the seamless sharing of common services and user experiences across apps.</p><p>Over many years, the release of new capabilities and features across product lines led to exponential application growth, prompting us to explore ways to scale them. With a goal to accelerate developer velocity and streamline the delivery of customer benefits, we embarked on a journey<a href="https://medium.com/intuit-engineering/accelerating-developer-velocity-intuits-appfabric-platform-41562af87e65"> to create a unified AppFabric platform for all web applications at Intuit.</a> As a foundational element of that journey, we set out to tackle the most pressing challenges facing mobile teams across Intuit.</p><p>In this blog, we’ll take a deep dive into how we achieved the following mobile app development outcomes:</p><ul><li>Reducing duplication of work</li><li>Delivering personalized developer experiences with apps and widgets</li><li>Achieving scalable design and seamless integration across products and services</li><li>Increasing developer productivity</li></ul><p>To illustrate our approach, let’s take a look at a couple of development team scenarios:</p><p><strong>Use Case 1: </strong>Team A has built a “<em>choose credit card payment”</em> function. They brainstormed ideas, created the design, developed the module and screens, and integrated it within their app during this process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2R9PjjyQyKUvH6FI" /></figure><p>When Team B needs to build a similar functionality, they would have to duplicate the process, resulting in disjointed customer experiences for the<em> “choose credit card payment”</em> feature across multiple apps. Now, imagine for a moment that five other teams need the same functionality in their apps!</p><p><strong>Use Case 2: </strong>A team has created all the tooling and capabilities for a new module [e.g., logging, analytics, CI/CD (continuous integration/continuous delivery), real-time user monitoring, data layer, authentication, experimentation framework].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KoWmsclBjjnKIbWa" /></figure><p>In this scenario, would another team want to recreate these capabilities for a new module? Obviously, the answer to that question is, “NO!”</p><p>These examples are typical of the challenges faced by Intuit customers and developers, and the impetus for building a unified mobile platform for all Intuit applications, across product lines.</p><h3><strong>AppFabric Mobile Platform</strong></h3><p>With our AppFabric Mobile platform, we have transformed our approach to mobile development here at Intuit, enabling developers and development teams to deliver innovative new products and services faster to customers. Foundational to our approach was decomposing mobile applications into smaller, reusable units of work called widgets. <strong>Mobile widgets </strong>are shared components with UI logic and associated services that can be quickly stitched together to create user experiences in existing products and/or brand-new offerings. <strong>Web widgets</strong> are developed for web applications using web technologies and can be easily configured and loaded in the embedded web browsers in mobile applications:</p><h4><strong>Reducing duplication of work</strong></h4><p><strong>Shared widgets: </strong>A widget is now shared across various apps and teams at Intuit. A good example of this is the “choose credit card” widget, which allows customers to choose preferred banks and store credit cards. We’ve seen incredible growth in the sheer number of shared widgets at Intuit over the past two years, which has been helpful for both existing and new apps. We now have over 60 shared widgets that provide a wide range of functionalities to the apps and this number is growing.</p><p><strong>Built-in capabilities:</strong> AppFabric Mobile’s built-in capabilities, such as<em> logging, analytics, real-time user monitoring, A/B tests</em>, etc., have streamlined the development process for a new app or a widget by providing a wide range of ready-made capabilities to choose from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2MEV_wR8IeMKEtcb" /></figure><h4><strong>Delivering personalized developer experiences with apps and widgets</strong></h4><p>Personalized dashboards for <em>logging, performance, real-time user monitoring, instrumentation,</em> etc. eliminate the need to search through logs and help developers to identify issues faster.</p><h4><strong>Achieving Scalable Design and Seamless Integration</strong></h4><p>AppFabric Mobile facilitates faster and more independent development across Intuit. Widgets can integrate seamlessly into any app. This helps customers see features faster and enables teams to develop, test, and deploy experiences rapidly and independently. Consequently, our developers build experiences that are used across apps, rather than a feature that is limited to one app.</p><h4><strong>Increasing developer productivity</strong></h4><p>AppFabric Mobile makes the development and release process quick and easy, reducing the time required to create projects with common components from a few hours/days to a few seconds/minutes.</p><p>To create an app or widget, Intuit developers can simply go to the developer platform portal at Intuit and add a few simple details to auto-generate a project. The project comes loaded with built-in capabilities, tooling, CI/CD integration, and set up for an Amazon Device Farm (ADF) and Virtual Device Farm (VDF), etc. These capabilities dramatically increase developer productivity, freeing them up to deliver awesome user experiences faster to our customers.</p><h3><strong>Fostering a Culture of Collaboration, Unified Standards &amp; Best Practices</strong></h3><p>Ultimately, our success with AppFabric Mobile can be attributed to Intuit’s company-wide commitment to collaboration across engineering teams — throughout the product development cycle — enabled by a set of unified AppFabric Mobile standards, best practices, and feedback loops.</p><p>For example, our request for comments (RFC) process — modeled after <a href="https://github.com/reactjs/rfcs">React RFC </a>— enables developers to <a href="https://en.wikipedia.org/wiki/Inner_source">inner source</a> their solutions to AppFabric Mobile. And, developers can participate in code jam sessions and easily view sample apps, documentation, etc. to stay on top of AppFabric Mobile capabilities. This has led to contributions from more than 100 developers across various teams.</p><h4><strong>AppFabric Mobile Platform: Key Takeaways</strong></h4><p>Over the past two years, we have reduced duplication by 50 percent with over 60 shared widgets and more than 25 ready-made capabilities and delivered personalized developer experiences quickly. The tremendous success of our <a href="https://medium.com/intuit-engineering/accelerating-app-development-at-scale-intuits-micro-frontend-architecture-ea87d565f0d">micro frontend architecture for web applications at Intuit</a> was a major source of inspiration and a contributor to these results. We’re proud of our accomplishments to date and committed to continuous innovation, as part of our overarching <a href="https://medium.com/intuit-engineering/accelerating-developer-velocity-intuits-appfabric-platform-41562af87e65">AppFabric Platform strateg</a>y to:</p><ul><li>Promote reuse of experiences to share across the products and reduce the build of clones.</li><li>Stitch seams between products so a customer navigating from one product to the other does not have starkly different experiences.</li><li>Deliver complete platforms with out-of-the-box capabilities for operational excellence, developer productivity, and elimination of duplicate work.</li><li>Shorten release cycles to deliver delight to users faster.</li></ul><p>Our journey continues — so stay tuned!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fef2db3910b2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/intuit-engineering/appfabric-mobile-intuits-unified-mobile-platform-fef2db3910b2">AppFabric Mobile: Intuit’s Unified Mobile Platform</a> was originally published in <a href="https://medium.com/intuit-engineering">Intuit Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mobile Architecture — What design pattern I should follow?]]></title>
            <link>https://medium.com/macoclock/mvvm-with-modularization-decoupling-scalable-easier-to-unit-test-design-pattern-5d05c247d0aa?source=rss-543f42d7bfd3------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d05c247d0aa</guid>
            <category><![CDATA[design-patterns]]></category>
            <category><![CDATA[mobile-architecture]]></category>
            <category><![CDATA[swift]]></category>
            <category><![CDATA[modularization]]></category>
            <category><![CDATA[mobile-design]]></category>
            <dc:creator><![CDATA[Waseem Syed]]></dc:creator>
            <pubDate>Wed, 15 Apr 2020 01:09:19 GMT</pubDate>
            <atom:updated>2021-10-17T02:05:15.759Z</atom:updated>
            <content:encoded><![CDATA[<h3>Mobile Architecture — What design pattern I should follow?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*erVqlTKQ7b1wnZhdBRMNKQ.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@ffstop?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Fotis Fotopoulos</a> on <a href="https://unsplash.com/s/photos/software?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>All the developers at some point in time would have asked themselves — What is the best design pattern to build my application? Do I go with any of the design patterns like MVC, MVP, MVVM, VIPER, etc.? While all of the design patterns mentioned above have their pros and cons, any good design should follow the basic principles:</p><ul><li>Modularity</li><li>Single Responsibility</li><li>Scalability</li><li>Testability</li></ul><p>This blog talks about one such design that satisfies the four principles listed above and the benefits it offers to the developers with a simple example.</p><h3>Modularity</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sVTW-6ghKdQ4ifmcM1E1cw.png" /><figcaption>Design</figcaption></figure><p>Any system can be broadly divided into four logical units:</p><ul><li><strong>View Controller</strong> is the core of the entire design and it has modules like <em>dataSource, controller,</em> and <em>delegate</em>. The <em>controller</em><strong> </strong>is responsible for the basic setup. <em>data source</em><strong> </strong>handles the business logic and the <em>delegate</em> handles the navigation logic.</li><li><strong>Network Layer</strong> has two modules namely n<em>etwork manager</em> and s<em>ervice client </em>that handle all the network operations.</li><li><strong>Presentation Layer</strong> provides the data to the screens and has modules like <em>models</em>, <em>view models</em>, and <em>cell models</em>. The <em>view model</em> is an observer in the <em>controller</em> which usually triggers a data reload based on the <em>model </em>injected into it. The view model consists of functions much à la the <em>data source</em> functions like <em>numberOfSections</em> ,<em>numberOfRows</em> , <em>cellForRowAtIndexPath.</em></li><li><strong>UI Layer: </strong>has the v<em>iews</em> and c<em>ells</em> that render content.</li></ul><p>Some of the benefits modularity offers to developers include:</p><ul><li><strong>Independent development</strong>: Modules can be developed independently and can be easily shared between different systems.</li><li><strong>Easier to test</strong>: Since the system is broken down into smaller modules. Testing each module becomes a lot easier.</li></ul><h3><strong>Single Responsibility</strong></h3><p>All the classes have a single responsibility. For example, the cells, whose only responsibility is to display data do not need to know where the data came from. The <em>data source</em> which handles business logic does not have to worry about what goes on in the <em>delegate</em>.</p><h4><strong>Example</strong></h4><p>Let’s take a look at a real example of how all this translates into code with a simple app that fetches data from the URL and displays it in a tableView.</p><ul><li>The <em>ViewController</em> should only have a simple <em>setUp</em> and <em>fetch</em> data function. It owns the viewModel which in turn has a <em>tableView</em> reload function to reload the tableView every time there is a change in the <em>ViewModel</em>. The other components of ViewController, that are separated out and owned by the ViewController are DataSource and Delegate.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QDE1H9O42zDiZdzCfxt5jw.png" /><figcaption><strong>Controller</strong></figcaption></figure><ul><li>The <em>Service</em> and <em>Network Clients </em>are simply used to fetch data.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NWkCj7XEWx9S22xz1WLo-A.png" /><figcaption><strong>Service Layer and Network Client</strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gG3YXo5LJvk5k4koAqTAMQ.png" /><figcaption><strong>Endpoint</strong></figcaption></figure><ul><li>The service call returns a data model to be injected as a dependency to the ViewModel.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/634/1*JP_lW82A47QzgveZq0pjFg.png" /><figcaption><strong>Model</strong></figcaption></figure><ul><li>The <em>ViewModel</em> has functions that are exact replicas of the DataSource. It uses the model to create data to be supplied for data source functions. The data needed by the rows/cells in the function cellForRowAtIndexpath are supplied in the form of a cell model. PhotoCellModel in the below example.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q0J-RYNK24EdRJmkPr6hxg.png" /></figure><p><em>Delegate</em> just handles the navigation logic and assigns the responsibility of push or present DetailViewController to the ViewController.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rBufpe0QcU2aHKao2fqmFQ.png" /></figure><p><strong>Benefits:</strong></p><ul><li><strong>Separation of concerns</strong>: A class is not impacted by what happens in another class.</li><li><strong>Better readability and error detection: </strong>Since each class is responsible for only one task, reading the code and detecting errors becomes a lot easier.</li></ul><h3><strong>Scalability</strong></h3><p>Since the modules are decoupled from the core logic and developed independently, most of them could be passed around different screens with slight changes.</p><h4><strong>Benefits:</strong></h4><ul><li>Becomes much easier to extend functionality without having to change the entire system.</li><li>Shared modules limit the amount of duplicate code written and thereby make maintenance easy.</li></ul><h3><strong>Testing</strong></h3><p>Since we divide the system into smaller components and use dependency injection to inject objects, it becomes a lot easier to add tests.</p><p><strong>Testing EndPoint, Service, and Data Model: </strong>Using the mock class and mock JSON, we can test the entire Service layer, Endpoint, and Data models. The <strong>System Under Test </strong>(sut)in this case, is the <strong>Service</strong> PhotosService.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pFz3s0yUCSR73b6Z19agWA.png" /><figcaption><strong>EndPoint, Service, and Data Model tests</strong></figcaption></figure><ul><li><strong>Testing ViewModel and Row/Cell Models</strong></li></ul><p>The ViewModel holds clearly separates the presentation logic from the Controller and hence testing the ViewModel thoroughly is as good as testing the Views. The <strong>System Under Test </strong>(sut)in this case is the <strong>ViewModel </strong>PhotoViewModel.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ta0-EHU4y0yavw3r7JE7vQ.png" /><figcaption><strong>View Model and Row/Cell Model Tests</strong></figcaption></figure><h4><strong>Benefits:</strong></h4><ul><li>Thanks to dependency injection, each class can be fully tested.</li><li>Classes can be tested independently without integrating into a system.</li><li>Makes it easier to mock classes and data.</li></ul><h3><strong>Conclusion</strong></h3><p>The design we discussed so far can be applied to any app with slight modifications based on the requirements and complexity of the app. To summarize, a good design is one that:</p><ul><li>Enables the developers to work independently.</li><li>Allows you to modify parts of your app easily.</li><li>Scales as the system grows.</li><li>Is easy to test.</li></ul><p>Cheers!</p><p><a href="https://github.com/wsyed1/photos-practice-app">wsyed1/photos-practice-app</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d05c247d0aa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/macoclock/mvvm-with-modularization-decoupling-scalable-easier-to-unit-test-design-pattern-5d05c247d0aa">Mobile Architecture — What design pattern I should follow?</a> was originally published in <a href="https://medium.com/macoclock">Mac O’Clock</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>