<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Emmanuel Ene Rejoice Gideon </title>
    <description>The latest articles on DEV Community by Emmanuel Ene Rejoice Gideon  (@emmanuel_erogian).</description>
    <link>https://dev.to/emmanuel_erogian</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4071781%2F63e3b17e-540a-48aa-8c7c-9d366dc5a530.jpg</url>
      <title>DEV Community: Emmanuel Ene Rejoice Gideon </title>
      <link>https://dev.to/emmanuel_erogian</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/emmanuel_erogian"/>
    <language>en</language>
    <item>
      <title>The $0 Playbook: How We Train Real AI Models on Free GPUs (And Get Paid for Them)</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Tue, 27 Oct 2026 10:56:04 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/the-0-playbook-how-we-train-real-ai-models-on-free-gpus-and-get-paid-for-them-54lp</link>
      <guid>https://dev.to/emmanuel_erogian/the-0-playbook-how-we-train-real-ai-models-on-free-gpus-and-get-paid-for-them-54lp</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" alt="STEWMIND 1.7B" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Our team trained a 1.7B agentic model, a multilingual Nigerian speech AI, and benchmarked both — with &lt;strong&gt;zero dollars of compute budget&lt;/strong&gt;. Not a demo. Real models, real benchmarks, real downloads (1,000+ this month).&lt;/p&gt;

&lt;p&gt;This is the exact playbook. Steal it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The pipeline (how $0 AI training actually works)
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; ┌─────────────────────────────────────────────────────────────┐
 │                    THE $0 TRAINING PIPELINE                  │
 └─────────────────────────────────────────────────────────────┘

   FREE DATA                    FREE COMPUTE               SHIP
  ┌──────────────┐   ┌──────────────────┐   ┌────────────────────┐
  │ HF Datasets  │   │ Kaggle GPU 30h/wk│   │ Hugging Face Hub   │
  │ Common Voice │──▶│ Colab (free)     │──▶│ Replicate (sell API)│
  │ FLEURS/MMS   │   │ Lightning free   │   │ GGUF for laptops   │
  │ Your own     │   └────────┬─────────┘   │ Dev.to to market   │
  └──────────────┘            │            └────────────────────┘
        │                     │                     │
        ▼                     ▼                     ▼
   CLEAN + FILTER       LoRA r16 fine-tune      EVALUATE FIRST,
   (quality &amp;gt; size)      (1-7B model,           SHIP SECOND,
   1k good samples      not 100B!)              MARKET ALWAYS
   beat 10k noisy
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The secret nobody tweets: &lt;strong&gt;a 7B model with 1,000 perfect samples beats a 70B model with 100,000 noisy ones.&lt;/strong&gt; We cut our speech model's word errors by 43-70% with &lt;em&gt;400 clean samples per language&lt;/em&gt; and a free GPU.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: Gather data that makes models human-good
&lt;/h2&gt;

&lt;p&gt;A model is only as human as its data. Ranked by value per hour of your time:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Public datasets that exist for your niche&lt;/strong&gt; (HF Datasets, Common Voice, FLEURS, MMS) — hours of native speech and agentic workflows, free&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your own data&lt;/strong&gt; — voice notes, chat logs, work documents. Ours learned company knowledge and Pidgin from hours of conversation data&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Synthetic data from a bigger model&lt;/strong&gt; — ask GPT-6.1 to generate 500 examples of the behavior you want, then curate by hand&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Record real people&lt;/strong&gt; — for voice: 30 minutes per speaker covers one accent. Ten speakers beat ten hours of one speaker&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Rules: clean &amp;gt; big. Filter everything. Read 50 random samples personally — if 5 are garbage, fix your filter before training.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Free platforms that can train real models
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Kaggle notebooks&lt;/strong&gt;: 30 GPU hours/week, free forever. Our 1.7B agentic model trained here&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Google Colab&lt;/strong&gt;: free T4 sessions, perfect for LoRA fine-tunes under 7B&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lightning AI Studios&lt;/strong&gt;: free monthly GPU credits&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hugging Face&lt;/strong&gt;: free model hosting, free distribution, free credibility&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Replicate&lt;/strong&gt;: free model upload — you pay nothing until customers pay you&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The trick: &lt;strong&gt;LoRA, not full training.&lt;/strong&gt; Fine-tune 0.1% of a model's parameters (LoRA r16) and you train on free GPUs what used to need paid clusters. Chain sessions: our 100k-sample run saves checkpoints and resumes across free sessions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Evaluate or embarrass yourself
&lt;/h2&gt;

&lt;p&gt;Before publishing, we ran live tests: transcription on real audio, JSON tool calls on real prompts, WER benchmarks against the base model. Publish your numbers &lt;em&gt;and your failures&lt;/em&gt;. The honest benchmark table in my last post (GPT-6.1 vs our 1.7B) drove more traffic than any hype ever did. Proof beats promises.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 4: Monetize (the part they skip)
&lt;/h2&gt;

&lt;p&gt;How AI models make money, ranked by margin:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Your own API wrapper&lt;/strong&gt;: host your model, sell access per 1,000 calls via Paystack/Stripe. GPU cost becomes your COGS — margins are 70-90% when your model is small&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fine-tuning services&lt;/strong&gt;: companies pay you to build THEIR private model. You have the pipeline; they have the budget&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Product &amp;gt; model&lt;/strong&gt;: don't sell the model, sell what it does — transcription apps for journalists, voice agents for banks, lecture capture for universities&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Replicate marketplace&lt;/strong&gt;: publish, get discovered, charge for usage&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Licensing&lt;/strong&gt;: free for personal use, paid commercial license for call centers, courts, hospitals&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The real wealth formula is boring: &lt;em&gt;pick ONE niche that hurts&lt;/em&gt; (voice banking in local languages, exam prep for students, sermon transcription), build the tiny model that solves it perfectly, charge for the outcome.&lt;/p&gt;

&lt;h2&gt;
  
  
  The truth about getting "rich" from models
&lt;/h2&gt;

&lt;p&gt;Owning the weights is the moat. Companies renting GPT are renting their product's brain forever. Every improvement you ship compounds. Every download is free distribution. Every customer funds the next model.&lt;/p&gt;

&lt;p&gt;Big models cook for everyone. Your model cooks for you. 🍲&lt;/p&gt;

&lt;p&gt;Both our models are live: &lt;a href="https://huggingface.co/Erog0291/stewmind-1.7b" rel="noopener noreferrer"&gt;StewMind 1.7B&lt;/a&gt; (agentic brain) and &lt;a href="https://huggingface.co/Erog0291/stew-audio-0.1" rel="noopener noreferrer"&gt;Stew Yanga&lt;/a&gt; (Nigerian speech AI). Fork the recipe.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>opensource</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>OpenAI's GPT-6.1 Sol Is Brilliant. I Still Won't Put It Inside My Agents.</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Tue, 27 Oct 2026 10:55:54 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/openais-gpt-61-sol-is-brilliant-i-still-wont-put-it-inside-my-agents-i2o</link>
      <guid>https://dev.to/emmanuel_erogian/openais-gpt-61-sol-is-brilliant-i-still-wont-put-it-inside-my-agents-i2o</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" alt="STEWMIND 1.7B" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;OpenAI just shipped &lt;strong&gt;GPT-6.1 Sol&lt;/strong&gt; (Sept 29, 2026). It's genuinely brilliant.&lt;/p&gt;

&lt;p&gt;And I still won't use it inside my agents.&lt;/p&gt;

&lt;p&gt;Don't get me wrong — for deep reasoning, GPT-6.1 is untouchable by any 1.7B model, including mine. But here's what changed in 2026: &lt;strong&gt;API prices went UP on 10 of 14 model tiers&lt;/strong&gt;, rate limits got tighter, and every agentic loop I ship burns tokens on &lt;em&gt;every step&lt;/em&gt; — routing, tool calls, memory ops, retries.&lt;/p&gt;

&lt;p&gt;So I trained my own agent brain. &lt;strong&gt;StewMind 1.7B&lt;/strong&gt; — built in Nigeria 🇳🇬, fine-tuned for one job: running agents locally.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest benchmark
&lt;/h2&gt;

&lt;p&gt;Here's StewMind vs the latest GPT, measured on real agentic workloads. No marketing, measured and public-listed numbers only:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Capability&lt;/th&gt;
&lt;th&gt;StewMind 1.7B (local)&lt;/th&gt;
&lt;th&gt;GPT-6.1 Sol (API)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Cost per 1M tokens&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$0.00 — forever&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Paid, and 2026 prices rose on 10/14 tiers&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Agent routing call&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;~0.01s, on your CPU&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Network round-trip + queue&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Data privacy&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Never leaves the device&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Leaves your machine, every call&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Works offline&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Yes&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ownership&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Yours (Apache-2.0, weights included)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;OpenAI's&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Runs on a phone&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Yes — 1GB GGUF&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucination on tools&lt;/td&gt;
&lt;td&gt;Possible, but you can constrain it&lt;/td&gt;
&lt;td&gt;Possible, and you can't fix their model&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Deep reasoning&lt;/td&gt;
&lt;td&gt;Loses, clearly&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Wins, clearly&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Writes better poetry&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Yes&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Read that table again: for &lt;em&gt;agents&lt;/em&gt;, the rows that matter are the first six. Agents don't need a poet. They need reflexes.&lt;/p&gt;

&lt;h2&gt;
  
  
  What StewMind actually does (measured, live)
&lt;/h2&gt;

&lt;p&gt;I benchmarked her on my own machine — no GPU, plain CPU:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tool routing&lt;/strong&gt; — input: &lt;em&gt;"Send 5000 naira to my brother today"&lt;/em&gt;, reply with JSON only. She answered:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"tool"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"payment"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"amount"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;5000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"recipient"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"brother"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"currency"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"naira"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"date"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"today"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Correct tool, correct arguments, numbers typed as numbers, currency inferred. That's a real tool call a machine can execute.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;ReAct planning, memory ops, intent routing, Pidgin ("Wetin be your main work?")&lt;/strong&gt; — all coherent. She's being upgraded right now on 100,000 agentic workflow samples (checkpoints landing at &lt;a href="https://huggingface.co/Erog0291/stewmind-1.7b-ultra" rel="noopener noreferrer"&gt;stewmind-1.7b-ultra&lt;/a&gt; as we speak).&lt;/p&gt;

&lt;h2&gt;
  
  
  The uncomfortable truth about 2026 AI
&lt;/h2&gt;

&lt;p&gt;Everyone is building agents on rented brains. Every spike in GPT pricing, every rate limit, every outage — that's your product going down.&lt;/p&gt;

&lt;p&gt;The biggest AI insight this year came from NVIDIA's own researchers: &lt;strong&gt;"Small Language Models are the Future of Agentic AI."&lt;/strong&gt; Specialized, repetitive, latency-critical tasks — exactly what agents do all day — belong to small local models. Giants handle the once-a-week deep question.&lt;/p&gt;

&lt;p&gt;StewMind is the reflex layer. GPT-6.1 can be the brain you call when it's worth it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Get her
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Model + GGUF (runs in Ollama, llama.cpp, anything): &lt;a href="https://huggingface.co/Erog0291/stewmind-1.7b" rel="noopener noreferrer"&gt;Erog0291/stewmind-1.7b&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;3 lines of Python and she's running on your machine, zero API bills
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;transformers&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;AutoTokenizer&lt;/span&gt;
&lt;span class="n"&gt;tok&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;AutoTokenizer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Erog0291/stewmind-1.7b&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Erog0291/stewmind-1.7b&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;1,000+ people grabbed her this month. Big models cook for everyone. This one cooks for &lt;em&gt;you&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;If you're building agents in Africa, or anywhere the cloud bills bite — come cook with us. 🍲&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>opensource</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>Big Models Cook for Everyone. This One Cooks for You: StewMind 1.7B 🇳🇬</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Mon, 26 Oct 2026 08:36:36 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/big-models-cook-for-everyone-this-one-cooks-for-you-stewmind-17b-2b7o</link>
      <guid>https://dev.to/emmanuel_erogian/big-models-cook-for-everyone-this-one-cooks-for-you-stewmind-17b-2b7o</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fhuggingface.co%2FErog0291%2Fstewmind-1.7b%2Fresolve%2Fmain%2Fbanner.jpg" alt="STEWMIND 1.7B" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;StewMind 1.7B is live.&lt;/p&gt;

&lt;p&gt;A tiny agentic AI built in Nigeria 🇳🇬 that runs on YOUR machine.&lt;/p&gt;

&lt;p&gt;Big models cook for everyone. This one cooks for you.&lt;/p&gt;

&lt;p&gt;Most AI today is a vending machine: you pay per token, you rent someone else's brain, and when the API bills spike, your product dies. We built the opposite. A small, sharp, agentic brain you download once and run forever.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;→ &lt;strong&gt;Agent routing&lt;/strong&gt;: understands intent and picks the right skill&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;Tool-calling&lt;/strong&gt;: structured function calls, not vibes&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;ReAct reasoning&lt;/strong&gt;: thinks, acts, observes, repeats&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;Memory ops&lt;/strong&gt;: built for agent systems that remember&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;English + Nigerian Pidgin&lt;/strong&gt;: it understands how we actually talk&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;GGUF ready&lt;/strong&gt;: runs on your laptop, your mini-PC, even your phone&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;$0 API bills&lt;/strong&gt;: no subscriptions, no rate limits, no surprise invoices&lt;/p&gt;

&lt;h2&gt;
  
  
  Run it in 10 lines
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;transformers&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;AutoTokenizer&lt;/span&gt;

&lt;span class="n"&gt;tok&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;AutoTokenizer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Erog0291/stewmind-1.7b&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Erog0291/stewmind-1.7b&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;torch_dtype&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;auto&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;messages&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Plan a workflow: fetch weather, then notify the user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}]&lt;/span&gt;
&lt;span class="n"&gt;ids&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;tok&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;apply_chat_template&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;return_tensors&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;pt&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;add_generation_prompt&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;out&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;generate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ids&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;max_new_tokens&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;256&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;do_sample&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;False&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tok&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decode&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;out&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="n"&gt;ids&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;shape&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]:],&lt;/span&gt; &lt;span class="n"&gt;skip_special_tokens&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No GPU? No problem. The Q4_K_M GGUF is right in the repo — run it with llama.cpp, LM Studio, or Ollama. It sips RAM like jollof disappears at a party.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why 1.7B is a feature, not a bug
&lt;/h2&gt;

&lt;p&gt;Every agent developer I know has the same problem: the big models are brilliant but slow, expensive, and they burn your margin one tool call at a time.&lt;/p&gt;

&lt;p&gt;A small model fine-tuned specifically for agentic work does three things better:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;It routes in milliseconds&lt;/strong&gt; — no 900ms round trip just to decide which tool to call&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;It stays focused&lt;/strong&gt; — trained on agentic workflows, not on everything under the sun&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;It runs where your users are&lt;/strong&gt; — offline, on-device, private data never leaves the machine&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The future of agents isn't one giant brain in a datacenter. It's thousands of small brains living in your apps, your scripts, your hardware.&lt;/p&gt;

&lt;h2&gt;
  
  
  Built in Nigeria, for the world
&lt;/h2&gt;

&lt;p&gt;StewMind comes out of the S.T.E.W project — we're building Nigerian AI infrastructure: speech models that hear Igbo, Yoruba, Hausa and Pidgin, agentic models that think locally, and tools that work where connectivity is expensive and unreliable.&lt;/p&gt;

&lt;p&gt;If you're building agents in emerging markets, you know the pain: latency kills UX and API costs kill startups. Local-first AI is not a compromise here. It's the strategy.&lt;/p&gt;

&lt;h2&gt;
  
  
  Download it
&lt;/h2&gt;

&lt;p&gt;🥣 &lt;strong&gt;Model + GGUF&lt;/strong&gt;: &lt;a href="https://huggingface.co/Erog0291/stewmind-1.7b" rel="noopener noreferrer"&gt;https://huggingface.co/Erog0291/stewmind-1.7b&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Remix it, fine-tune it, ship it in your product. If you build something with StewMind, drop a comment — I want to see what you cook. 🍲&lt;/p&gt;

&lt;p&gt;And if you work with African languages, we also built Stew Yanga — Nigerian speech-to-text that actually understands Nigerian accents: &lt;a href="https://huggingface.co/Erog0291/stew-yanga" rel="noopener noreferrer"&gt;https://huggingface.co/Erog0291/stew-yanga&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>opensource</category>
      <category>python</category>
    </item>
    <item>
      <title>I Dogfooded My 49KB AI Coding Agent — It Fought Back 5 Times</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Sat, 05 Sep 2026 16:13:58 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/i-dogfooded-my-49kb-ai-coding-agent-it-fought-back-5-times-5hnc</link>
      <guid>https://dev.to/emmanuel_erogian/i-dogfooded-my-49kb-ai-coding-agent-it-fought-back-5-times-5hnc</guid>
      <description>&lt;p&gt;Yesterday I wrote about &lt;a href="https://dev.to/emmanuel_erogian/i-built-an-ai-coding-agent-that-ships-in-under-50kb-zero-dependencies-runs-on-a-phone-2c8c"&gt;building an AI coding agent that ships in under 50KB&lt;/a&gt;. A fair comment kept coming up: &lt;em&gt;"sounds cool, but does it actually work?"&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Fair question. So today I did the thing every builder dreads: I dogfooded the version sitting on npm — not my dev repo, the actual &lt;code&gt;npm i -g stew-ai&lt;/code&gt; package a stranger would get.&lt;/p&gt;

&lt;p&gt;It fought back. Five times.&lt;/p&gt;

&lt;h2&gt;
  
  
  Round 1: The auth flow was broken
&lt;/h2&gt;

&lt;p&gt;First command a new user runs after installing is either &lt;code&gt;stew register&lt;/code&gt; or &lt;code&gt;stew login &amp;lt;key&amp;gt;&lt;/code&gt;. Here's what 2.7.5 actually did:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ stew whoami
❌ Error: apiKey is not defined
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A &lt;code&gt;ReferenceError&lt;/code&gt;. In production. The &lt;code&gt;whoami&lt;/code&gt; command referenced an &lt;code&gt;apiKey&lt;/code&gt; variable that was scoped to a &lt;em&gt;different function&lt;/em&gt; — a classic closure bug that my offline tests never caught, because the broken path only triggers when a real key exists in &lt;code&gt;~/.stew/config.json&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;While fixing it, I found four more:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;whoami&lt;/code&gt; crash&lt;/strong&gt; — the scope bug above. &lt;code&gt;apiKey&lt;/code&gt; lived in &lt;code&gt;authCommand&lt;/code&gt;, not &lt;code&gt;whoamiCommand&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;stew login &amp;lt;key&amp;gt;&lt;/code&gt; ignored the key&lt;/strong&gt; — the CLI treated the first argument as the action name, so your key was parsed as a command and you got dropped into the interactive prompt anyway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;register()&lt;/code&gt; sent the wrong field name&lt;/strong&gt; — the SDK posted &lt;code&gt;full_name&lt;/code&gt; while the FastAPI backend expects &lt;code&gt;name&lt;/code&gt;. A silent 422.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;usage()&lt;/code&gt; authenticated wrong&lt;/strong&gt; — the SDK sent the key as a Bearer header while the endpoint expects an &lt;code&gt;api_key&lt;/code&gt; query param. &lt;code&gt;whoami&lt;/code&gt; could never show your plan.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;422 errors printed as &lt;code&gt;[object Object]&lt;/code&gt;&lt;/strong&gt; — FastAPI validation errors were stringified raw, so users saw garbage instead of "Missing/invalid field: name".&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Five bugs. All in the first 60 seconds of a new user's experience. All invisible in my test suite.&lt;/p&gt;

&lt;p&gt;This is why you dogfood the &lt;em&gt;published&lt;/em&gt; artifact. My tests ran against source files; the bug shipped anyway.&lt;/p&gt;

&lt;h2&gt;
  
  
  The fix: same day, with receipts
&lt;/h2&gt;

&lt;p&gt;All five fixes went into &lt;a href="https://www.npmjs.com/package/stew-ai/v/2.7.6" rel="noopener noreferrer"&gt;v2.7.6&lt;/a&gt;, plus five regression tests so each bug can never quietly return:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nf"&gt;check&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;auth: whoamiCommand no longer throws ReferenceError (apiKey scoped)&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...);&lt;/span&gt;
&lt;span class="nf"&gt;check&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;auth: direct alias routing — "stew login &amp;lt;key&amp;gt;" passes key, not action&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...);&lt;/span&gt;
&lt;span class="nf"&gt;check&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;SDK: register() sends "name" field (backend contract), not full_name&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...);&lt;/span&gt;
&lt;span class="nf"&gt;check&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;SDK: usage() passes api_key as query param (backend contract)&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...);&lt;/span&gt;
&lt;span class="nf"&gt;check&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;StewError: FastAPI 422 detail array becomes readable message&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;26/26 pass, package is 49.1KB — still under the 50KB budget, still zero dependencies.&lt;/p&gt;

&lt;p&gt;And the flow a user actually experiences now:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ stew login stew_O9X...
✅ Logged in! API key saved to ~/.stew/config.json

$ stew whoami
👤 S.T.E.W Account
  API Key: stew_O9X...P8b2
  Plan: free
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Round 2: Actually building something
&lt;/h2&gt;

&lt;p&gt;Auth works — fine. But the real question was whether the agent could code. So I opened the REPL (&lt;code&gt;stew code&lt;/code&gt;) and asked it to build a single-file landing page for a fictional Afrobeat festival in my hometown:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Build a stunning single-file index.html for a fictional Afrobeat concert in Enugu, Nigeria called 'Enugu Sound Fest 2026'. Dark theme, orange/green gradient accents, animated hero, artist lineup grid, live countdown timer, tickets section. Pure HTML/CSS/JS, no external dependencies.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It generated the whole page — hero, lineup grid, countdown, tickets. But when I opened the file, one line was broken:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nx"&gt;el&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;textContent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;$&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nx"&gt;d&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="nx"&gt;d&lt;/span&gt; &lt;span class="nx"&gt;$&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nx"&gt;h&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="nx"&gt;h&lt;/span&gt; &lt;span class="nx"&gt;$&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nx"&gt;m&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="nx"&gt;m&lt;/span&gt; &lt;span class="nx"&gt;$&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nx"&gt;s&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="nx"&gt;s&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c1"&gt;// ← template literal lost its backticks&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;So I did what any user would do: went back into the REPL and told it what was wrong. It rewrote the line with plain concatenation and applied the patch:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr2up0loacssv6ywyahwl.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr2up0loacssv6ywyahwl.png" alt="Stew Code fixing the broken countdown line in the terminal" width="800" height="498"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And the rendered result, screenshotted from the actual file it wrote:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm5je539y2v5zxqnpgysl.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm5je539y2v5zxqnpgysl.png" alt="The landing page Stew Code built — Enugu Sound Fest 2026" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Build → spot the bug → describe it → agent fixes it → verify. That loop worked, end to end, over the free API tier.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honest friction notes
&lt;/h2&gt;

&lt;p&gt;Dogfooding also surfaced things I'm &lt;em&gt;not&lt;/em&gt; proud of, so you get those too:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The file-apply format is too picky.&lt;/strong&gt; The CLI only writes files when the model formats code blocks in one exact way (a &lt;code&gt;// filepath:&lt;/code&gt; marker on the first line). The model often guessed a different format and the change silently didn't apply. It worked when I told it the exact format — but a user shouldn't have to know the magic words. That's next on the fix list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A queued message wedged the REPL&lt;/strong&gt; while a response was streaming — I had to restart the session. Input needs a lock during streaming.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Both of those are now tracked with the same rigor as the auth bugs: found by using the thing, not by staring at the code.&lt;/p&gt;

&lt;h2&gt;
  
  
  The takeaway
&lt;/h2&gt;

&lt;p&gt;If you build a tool for other people, the most valuable hour you can spend is being one of them — installing the published package, running the first three commands a stranger would run, and believing every error message you see.&lt;/p&gt;

&lt;p&gt;The five worst bugs in my auth flow survived 21 passing tests and two published versions. They died in one afternoon of pretending to be my own user.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Package: &lt;a href="https://www.npmjs.com/package/stew-ai" rel="noopener noreferrer"&gt;https://www.npmjs.com/package/stew-ai&lt;/a&gt; (v2.7.6, 49.1KB, zero deps)&lt;/li&gt;
&lt;li&gt;Source: &lt;a href="https://github.com/emmanuelerogian723-alt/stew-ai" rel="noopener noreferrer"&gt;https://github.com/emmanuelerogian723-alt/stew-ai&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Free key: &lt;a href="https://stew-agent.onrender.com" rel="noopener noreferrer"&gt;https://stew-agent.onrender.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Go break it. I'll fix what you find.&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>ai</category>
      <category>cli</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I Built an AI Coding Agent That Ships in Under 50KB — Zero Dependencies, Runs on a Phone</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Sat, 05 Sep 2026 15:13:50 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/i-built-an-ai-coding-agent-that-ships-in-under-50kb-zero-dependencies-runs-on-a-phone-2c8c</link>
      <guid>https://dev.to/emmanuel_erogian/i-built-an-ai-coding-agent-that-ships-in-under-50kb-zero-dependencies-runs-on-a-phone-2c8c</guid>
      <description>&lt;p&gt;I'm a developer from Enugu, Nigeria. Most devs I know here code on second-hand laptops — and a lot of them code straight from cheap Android phones using Termux.&lt;/p&gt;

&lt;p&gt;So when I set out to build an AI coding agent for the African market, "first, install 400 npm packages" was never going to work. Neither was "you need a $2,000 laptop."&lt;/p&gt;

&lt;p&gt;I set myself one hard rule: &lt;strong&gt;the entire package must stay under 50KB, with zero dependencies.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Meet Stew Code — a terminal AI agent that hits that rule and still does all of this:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;👀 &lt;strong&gt;Vision&lt;/strong&gt; — &lt;code&gt;/image screenshot.png&lt;/code&gt; and it actually &lt;em&gt;sees&lt;/em&gt; the screenshot&lt;/li&gt;
&lt;li&gt;🌐 &lt;strong&gt;Headless browsing&lt;/strong&gt; — cookies, forms, clicks, login flows. No Puppeteer&lt;/li&gt;
&lt;li&gt;📸 &lt;strong&gt;Screenshots &amp;amp; screen recording&lt;/strong&gt; — web pages or your whole screen&lt;/li&gt;
&lt;li&gt;🔒 &lt;strong&gt;Security scanning&lt;/strong&gt; — audits your codebase for real issues&lt;/li&gt;
&lt;li&gt;🐝 &lt;strong&gt;Swarm mode&lt;/strong&gt; — a planner splits your task, parallel agents execute it&lt;/li&gt;
&lt;li&gt;🗣️ &lt;strong&gt;Voice&lt;/strong&gt; — it reads responses out loud using OS text-to-speech&lt;/li&gt;
&lt;li&gt;🔌 &lt;strong&gt;MCP client&lt;/strong&gt; — connect any Model Context Protocol server&lt;/li&gt;
&lt;li&gt;🛠️ &lt;strong&gt;Self-healing installs&lt;/strong&gt; — missing Chromium? It auto-installs it, and warns you &lt;em&gt;first&lt;/em&gt;: "~250MB download, battery at 12% — plug in before we start"&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Quick start
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm i &lt;span class="nt"&gt;-g&lt;/span&gt; stew-ai
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's the whole install. 49KB. On a bad 3G connection it takes seconds, not minutes.&lt;/p&gt;

&lt;p&gt;Grab a free API key at &lt;a href="https://stew-agent.onrender.com" rel="noopener noreferrer"&gt;https://stew-agent.onrender.com&lt;/a&gt; (free tier: 1,500 calls/month), then:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;stew auth login   &lt;span class="c"&gt;# paste your key&lt;/span&gt;
stew code         &lt;span class="c"&gt;# launch the REPL&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And you're in.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the brain looks like
&lt;/h2&gt;

&lt;p&gt;Honest disclosure: the intelligence runs on my Stew Agent API — the CLI is the lightweight terminal client. The API is OpenAI-compatible (&lt;code&gt;/v1/chat/completions&lt;/code&gt;), so you can also point any OpenAI-shaped tool at it by changing the base URL.&lt;/p&gt;

&lt;p&gt;The API side packs 59 skills, a 100-agent swarm, 12 personas, and 6 AI providers (Groq, Mistral, NVIDIA NIM, OpenRouter, HuggingFace, OpenAI) with automatic failover — if one provider goes down, it switches mid-flight. And when you &lt;code&gt;npm i -g stew-ai&lt;/code&gt;, &lt;em&gt;none&lt;/em&gt; of that bloat lands on your machine. One package. Zero dependencies.&lt;/p&gt;

&lt;h2&gt;
  
  
  The zero-dependency receipts
&lt;/h2&gt;

&lt;p&gt;How do you fit vision, browsing, screenshots, recording, a REPL, and an MCP client into 49KB with zero deps?&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Node 18+ native &lt;code&gt;fetch&lt;/code&gt; — no axios, no node-fetch&lt;/li&gt;
&lt;li&gt;Raw ANSI escape codes for all terminal coloring&lt;/li&gt;
&lt;li&gt;Hand-written argument parsing&lt;/li&gt;
&lt;li&gt;Chrome headless &lt;code&gt;--screenshot&lt;/code&gt; for page shots; OS-native tools (&lt;code&gt;scrot&lt;/code&gt;, &lt;code&gt;screencapture&lt;/code&gt;, ffmpeg, PowerShell) for everything else&lt;/li&gt;
&lt;li&gt;The npm package ships minified; the GitHub repo stays fully readable&lt;/li&gt;
&lt;li&gt;A regression suite runs offline against mocks, so nothing breaks silently between releases&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it on your phone (yes, really)
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Install Termux from F-Droid, then:&lt;/span&gt;
pkg &lt;span class="nb"&gt;install &lt;/span&gt;nodejs
npm i &lt;span class="nt"&gt;-g&lt;/span&gt; stew-ai
stew code
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's an AI coding agent with vision and web access running on an Android phone. That's the entire point — AI tooling shouldn't require hardware most of the world can't afford.&lt;/p&gt;

&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;GitHub: &lt;a href="https://github.com/emmanuelerogian723-alt/stew-ai" rel="noopener noreferrer"&gt;https://github.com/emmanuelerogian723-alt/stew-ai&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;npm: &lt;a href="https://www.npmjs.com/package/stew-ai" rel="noopener noreferrer"&gt;https://www.npmjs.com/package/stew-ai&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Get a free API key: &lt;a href="https://stew-agent.onrender.com" rel="noopener noreferrer"&gt;https://stew-agent.onrender.com&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Docs &amp;amp; demo: &lt;a href="https://emmanuelerogian723-alt.github.io/stew-ai/" rel="noopener noreferrer"&gt;https://emmanuelerogian723-alt.github.io/stew-ai/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you're building for emerging markets, steal this constraint: &lt;strong&gt;under 50KB, zero deps.&lt;/strong&gt; It forces better engineering than any roadmap ever will.&lt;/p&gt;

&lt;p&gt;Star the repo if it's useful. And if something breaks on your machine, run &lt;code&gt;stew sysinfo&lt;/code&gt; and paste me the report — that command exists because debugging from a phone is hard. 🚀&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>ai</category>
      <category>opensource</category>
      <category>cli</category>
    </item>
    <item>
      <title>Stew Code 😋 Agent on terminal</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Thu, 03 Sep 2026 09:05:35 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/stew-code-agent-on-terminal-4864</link>
      <guid>https://dev.to/emmanuel_erogian/stew-code-agent-on-terminal-4864</guid>
      <description>&lt;p&gt;🚀 49.9KB. Zero dependencies. Built in Nigeria 🇳🇬&lt;br&gt;
&lt;strong&gt;Stew Code is the free AI coding agent rivaling Copilot &amp;amp; Cursor — without the $20/mo price tag.&lt;/strong&gt;&lt;br&gt;
Terminal-native. Built-in security scanner. Self-verifying code. Marathon mode. 6 AI providers with auto-failover. 16+ self-forging skills.&lt;br&gt;
One command:&lt;br&gt;
bash&lt;br&gt;
npm install -g stew-ai&lt;/p&gt;




&lt;p&gt;🔗 Blog: &lt;a href="https://erogian.vercel.app/blog/stew-code-free-ai-coding-agent-nigeria-2026" rel="noopener noreferrer"&gt;https://erogian.vercel.app/blog/stew-code-free-ai-coding-agent-nigeria-2026&lt;/a&gt;&lt;br&gt;
🔗 Site: &lt;a href="https://emmanuelerogian723-alt.github.io/stew-ai/" rel="noopener noreferrer"&gt;https://emmanuelerogian723-alt.github.io/stew-ai/&lt;/a&gt;&lt;br&gt;
📦 npm: pkg:npm/&lt;a href="mailto:stew-ai@2.5.0"&gt;stew-ai@2.5.0&lt;/a&gt;&lt;br&gt;
Built in Africa. For the world. 🌍&lt;/p&gt;

</description>
      <category>ai</category>
      <category>cli</category>
      <category>opensource</category>
      <category>programming</category>
    </item>
    <item>
      <title>Stew Code 😋 Agent on terminal</title>
      <dc:creator>Emmanuel Ene Rejoice Gideon </dc:creator>
      <pubDate>Thu, 03 Sep 2026 08:57:02 +0000</pubDate>
      <link>https://dev.to/emmanuel_erogian/stew-code-agent-on-terminal-7n4</link>
      <guid>https://dev.to/emmanuel_erogian/stew-code-agent-on-terminal-7n4</guid>
      <description>&lt;p&gt;🚀 49.9KB. Zero dependencies. Built in Nigeria 🇳🇬&lt;br&gt;
&lt;strong&gt;Stew Code is the free AI coding agent rivaling Copilot &amp;amp; Cursor — without the $20/mo price tag.&lt;/strong&gt;&lt;br&gt;
Terminal-native. Built-in security scanner. Self-verifying code. Marathon mode. 6 AI providers with auto-failover. 16+ self-forging skills.&lt;br&gt;
One command:&lt;br&gt;
bash&lt;br&gt;
npm install -g stew-ai&lt;/p&gt;




&lt;p&gt;🔗 Blog: &lt;a href="https://erogian.vercel.app/blog/stew-code-free-ai-coding-agent-nigeria-2026" rel="noopener noreferrer"&gt;https://erogian.vercel.app/blog/stew-code-free-ai-coding-agent-nigeria-2026&lt;/a&gt;&lt;br&gt;
🔗 Site: &lt;a href="https://emmanuelerogian723-alt.github.io/stew-ai/" rel="noopener noreferrer"&gt;https://emmanuelerogian723-alt.github.io/stew-ai/&lt;/a&gt;&lt;br&gt;
📦 npm: pkg:npm/&lt;a href="mailto:stew-ai@2.5.0"&gt;stew-ai@2.5.0&lt;/a&gt;&lt;br&gt;
Built in Africa. For the world. 🌍&lt;/p&gt;

</description>
    </item>
  </channel>
</rss>
