<- all tokdocs
2026.10.04@bashifuirkashi

The $89/month community sold in this RAG interview skit links three free public courses, and its own calendar screenshot shows 12 cancelled sessions

Career
Watch on TikTok

View on TikTok ->

A 107-second clip (1:47) at 720x1280, H.265 video and AAC audio at 356 kbps, 4,805,305 bytes, posted 2026-10-04 at 21:16 UTC by @bashifuirkashi under the channel nickname "Bashi | Software Engineer", scored to a track the metadata labels "original sound" by the same account, and captured 2026-10-05 at 05:02 UTC with 4,086 views, 221 likes, 16 comments, 21 reposts and 159 saves. I read 35 of the 54 extracted frames and all 374 transcript words. The clip is a one-actor sketch: the same person plays an interviewer in a black bomber jacket behind a silver MacBook, a "Theory" candidate in a white short-sleeve shirt, and a "Builder" candidate in a white long-sleeve oxford, all against the same beige wall with a plant at right. The opening card reads "🧠 Theory vs 🖥️ Builder" on a white pill above a black pill reading "Ai Engineer Interview (RAG Edition)". After the title, a persistent label sits at the top of every candidate shot, either "🧠 Theory" or "🖥️ Builder". Three cutaway graphics appear. The first, over the Theory candidate, is a five-step box diagram: "Query" to "Embedding model" (1), to "Vector DB" (2), to "Retrieved contexts" (3), to "LLM" (4), to "Response" (5). The second, over the Builder, is titled "ENTERPRISE KNOWLEDGE MANAGEMENT WITH RAG: CONCEPT & DATA FLOW" with a subhead "SIMPLIFIED REAL-TIME RAG DATA FLOW" and a sidebar reading "DEFINITION: ENTERPRISE RAG (RETRIEVAL-AUGMENTED GENERATION) / Retrieval-Augmented Generation (RAG) combined with enterprise data sources, real-time streaming ingestion, and context-aware AI generation for reliable, up-to-the-second responses"; its six numbered stages are "KNOWLEDGE SOURCES / Document Updates", "REAL-TIME INGESTION / CDC, Chunking, Transformation", "EMBEDDING GENERATION / Generate Vectors", "VECTOR DATABASE / Real-Time Vector Upserts", "USER QUERY / SRE Asks Question", "RETRIEVAL & GENERATION / Semantic Search, Context Assembly, Answer Synthesis", ending at "GROUNDED ANSWER". The third is a dark ten-stage chart headed "RAG Pipeline: How it works in a production environment" listing "Data ingestion / Gathering data from various sources", "Data streaming / Continuous data flow into the pipeline", "Data chunking / Dividing data into manageable segments", "Data cleaning / Removing inconsistencies and errors", "Embedding / Converting text into numerical vectors", "Indexing / Organizing embeddings for efficient retrieval", "Prompt assembly / Crafting prompts with retrieved context", "Retrieval / Finding relevant context for a query", "Orchestration / Coordinating pipeline components" and "Monitoring / Tracking pipeline performance". The last 8 seconds cut to three screen recordings of a Skool community headed "BASWE.Ai Engineer" with the tab row "Community Classroom Calendar Members Map Leaderboards About". One shows the lesson "Basic Retrieval Augmented Generation (RAG)" with a "RAG from Scratch" video thumbnail stamped 2:33:11, the goal line "Build a system that answers questions from your own documents, the single most in-demand AI engineering skill on the market right now", a step 2 linking to https://www.deeplearning.ai/short-courses/langchain-chat-with-your-data/, and two posts by "Kitu Komya" dated Jun 17 and Jun 18 titled "(KK 1) AI: High-level RAG concepts" and "(KK 2) AI: RAG architectural concepts & breakdown". A second shows "scikit-learn: Pipelines and Model Training" linking https://inria.github.io/scikit-learn-mooc/, under a sidebar reading "BASWE.AI Learning Roadmap" at "0%" with sections "Program Structure (READ FIRST)", "LLMs & RAG", "Integration, Agents & Orchestration", "Ops & Evaluation", "Safety & Ethics" and "ML Foundations". The third shows a month grid headed "August 2026 / 1:20pm Los Angeles time".

The "Theory" answer the skit mocks is the definition from the original RAG paper

The Theory candidate says: "first you ingest your documents chunk them create embeddings store them in a vector database retrieve the relevant chunks then pass them into an llm to generate the answer". The on-screen diagram over his shoulder draws exactly that, five boxes from Query to Response.

That description matches the architecture named in Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, submitted 22 May 2020 by Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel and Douwe Kiela. The paper pairs a "parametric memory" (a pre-trained seq2seq model) with a "non-parametric memory" that is "a dense vector index of Wikipedia, accessed with a pre-trained neural retriever". Retriever plus generator over a dense index is the definition, and the Theory candidate gave it.

The skit's framing puts a brain emoji on that answer and a monitor emoji on the Builder, implying the first candidate is wrong. He is incomplete on operations, evaluation and failure isolation, which is what the Builder adds. The Builder himself opens with "same foundation", which concedes the point. The audio ends with the interviewer saying "you're hired" to the Builder and the Theory candidate saying "i hate you", so the sketch resolves as a hiring outcome rather than a correctness verdict.

Every cost and latency fix the Builder names has a documented API behind it

Under the "traffic just went up by 10x" prompt, the Builder lists five moves. Each one checks out against vendor documentation, with specific numbers the clip does not give.

"Batch embeddings during ingestion" maps to the OpenAI Batch API, which documents a 50% cost discount versus synchronous requests, a 24-hour completion window, support for the /v1/embeddings endpoint, and a ceiling of 50,000 embedding inputs across all requests in a batch.

"Cache repeated queries or retrieval results" maps to Anthropic's prompt caching, which prices a 5-minute cache write at 1.25x base input tokens, a 1-hour cache write at 2.0x, and cache reads at 0.1x or lower depending on model. Minimum cacheable prefix lengths run from 512 to 4,096 tokens by model, so short retrieval contexts may not cache at all.

"Re-ranking" maps to a dedicated model class. Cohere's Rerank documentation describes models that "sort text inputs by semantic relevance to a specified query" and lists rerank-v4.0-pro, rerank-v4.0-fast, rerank-v3.5, rerank-english-v3.0 and rerank-multilingual-v3.0. That page carries no benchmark claim against embedding-only retrieval, so any relevance-lift figure attached to reranking is not coming from Cohere's own docs.

"Hybrid search", named in the hallucination segment, has a concrete implementation in Elasticsearch's reciprocal rank fusion. The rrf retriever combines a BM25 query and a kNN query with a default rank constant of 60 and no manual weighting between the two.

"Force the model to abstain when the context doesn't support an answer" and "citations" map to a shipped feature. Anthropic's Citations is marked generally available and returns "the exact passages that support each claim", supported across the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

The one piece the clip asserts without a mechanism is measurement. "Evaluation as separate components so i can measure and improve each one independently" has an off-the-shelf metric: Ragas Faithfulness scores a response from 0 to 1 as supported claims divided by total claims against the retrieved context. That is the number you would use to detect the exact failure the interviewer describes.

The community calendar in the B-roll shows 12 cancelled sessions, which works against the "daily calls" pitch

The closing pitch, verbatim from the transcript: "they have a full learning roadmap projects to put on my resume and daily calls with working ai and machine learning engineers".

The screen recording he chose to illustrate that is the Calendar tab for "August 2026", timestamped "1:20pm Los Angeles time", with the 26th circled in red as the current day. The grid spans Mon 27 July through Sun 6 September 2026. Twelve entries in that view carry a cancellation in the title text: 27 July "3:30pm - CANCELL...", 28 July "5pm - canceled sor...", 30 July "5pm - [CANCELLE...", 1 August "8:30am - Cancelle...", 3 August "3:30pm - CANCELL...", 5 August "5:30pm - Canceled...", 8 August "8:30am - Cancelle...", 10 August "5pm - Canceled so...", 17 August "5pm - Canceled at ...", 18 August "5pm - Canceled: H...", 19 August "5:30pm - Canceled..." and 20 August "5pm - Canceled: H...". Nine of the twelve fall inside August itself, and the week of 17 to 20 August shows a cancellation on four consecutive days.

Six dates in the view carry no events at all: 7, 15, 16, 22, 23 and 29 August. Five of those are weekend days and one is Friday 7 August.

The live cadence that does hold up in the grid is a weekday rhythm of named recurring blocks: "10am - AI/ML - Kitu" on most Tuesdays and Thursdays, "5pm - SWE:Intervie..." most weekdays, "3pm - SWE Office ..." and "5:30pm - System d..." on Wednesdays, "2:30pm - Data Scie..." on Fridays, and "12pm - Resume Re..." on Sundays. Calls do run. The screenshot does not support them running daily without cancellation.

The same post disagrees with itself on frequency. The spoken line says "daily calls". The TikTok description under the video says "weekly coaching from ai engineers & recruiters". The BASWE.AI Engineer Skool about page says "Daily calls with AI/ML Engineers, Recruiters & Hiring Managers". Two of those three claims cannot both be the offer.

Timing note: the calendar capture is dated 26 August 2026 and the video posted 4 October 2026, so the footage used to sell current access is 39 days old. The roadmap sidebar in the other two screen recordings reads "0%", meaning the account used for the demo has completed none of the program.

The roadmap's own lessons link to three free public courses, and the paid tier is $89 a month

The two Classroom screens expose the actual lesson content, and the links in them resolve to free material.

The LLMs & RAG lesson "Basic Retrieval Augmented Generation (RAG)" embeds a video titled "RAG from Scratch" with a 2:33:11 runtime badge. That is the freeCodeCamp release of LangChain's RAG From Scratch course, taught by Lance Martin of LangChain and published free on YouTube. LangChain announced the freeCodeCamp release on 17 April 2024. Step 2 of the same lesson links to DeepLearning.AI's "LangChain: Chat with Your Data", a 1 hour 18 minute course taught by LangChain co-founder and CEO Harrison Chase, listed as free.

The ML Foundations lesson "scikit-learn: Pipelines and Model Training" opens with "Start here (free course, with videos): scikit-learn MOOC, built by the scikit-learn core team" and links https://inria.github.io/scikit-learn-mooc/. That MOOC is developed by Inria Learning Lab, the scikit-learn team at La Fondation Inria, Inria Academy and probabl, and is published under a CC-BY licence at no cost. It does contain the two modules the lesson tells you to work through, "The predictive modeling pipeline" and "Selecting the best model".

The cost the video never states: the BASWE.AI Engineer community is $89 per month, listed at 304 members, with a 7-day money-back guarantee and a stated price increase at 500 members. There is also a free tier, Become an AI Engineer, with 6.7k members, offering starter roadmaps and the member network. Both are run by Bashiri Smith, the creator in the video. The sales page carries self-reported earnings claims ($0 to $265k in 2 years for the founder, $3M+ for members over 12 months) that have no independent source I could find, so treat them as marketing copy rather than verified figures.

The one named instructor visible in the Classroom is real and credentialed. The call recordings are posted by Kitu Komya, whose MentorCruise profile headlines "AI / ML Engineer & Senior Data Scientist. Shopify, UCLA, O'Reilly alum". Her name appears correctly spelled on the Skool posts and on the recurring "10am - AI/ML - Kitu" calendar blocks.

Two text errors: the calendar misspells FAANG, and the transcript garbles "query rewriting"

A recurring block on the community calendar is truncated to "FANNG AIE ..." at 2pm on 28 July, 12pm on 13 August, 2pm on 27 August and 10am on 3 September, with a matching "11:30am - FANNG ..." on 5 August. FAANG is the acronym for Facebook, Amazon, Apple, Netflix and Google. "FANNG" transposes a letter. This is the community's own calendar in its own promotional B-roll, so it is the operator's typo rather than a transcription artifact.

The transcript line "add metadata filters hybrid search query writing or re-ranking" almost certainly captures "query rewriting", the standard term for reformulating a user query before retrieval. "Query writing" is not a retrieval technique. I flag this as a Whisper artifact rather than an error by the speaker, since the surrounding list is otherwise a correct inventory of retrieval fixes.

Key Takeaways

  • Partial correction: the "Theory" answer the skit treats as inadequate is the canonical RAG architecture from Lewis et al., 2020. It is thin on operations and evaluation, and the Builder himself opens by saying "same foundation".
  • Verified: the Builder's cost and latency moves all have documented backing. Batch embeddings get a 50% discount with a 24-hour window on the OpenAI Batch API; prompt caching costs 1.25x on a 5-minute write and 0.1x or lower on reads; RRF hybrid search uses a default rank constant of 60.
  • Correction: the clip claims "daily calls", the video description says "weekly coaching", and the Skool page says "Daily calls". Three statements in one funnel, two of them incompatible.
  • Correction: the calendar B-roll used to sell those calls shows 12 entries titled as cancelled across the 27 July to 6 September 2026 view, 9 of them inside August, including four consecutive days from 17 to 20 August.
  • Correction: the community's own calendar spells FAANG as "FANNG" on its recurring interview-prep block, which appears five times in the visible month.
  • Unstated cost: $89 per month for BASWE.AI Engineer, 304 members at check, price rising at 500 members, 7-day money-back guarantee. The video says only "comment interview, i'll send it to you".
  • Unstated: the three courses visible inside the paid roadmap are free and public. RAG from Scratch on freeCodeCamp, LangChain: Chat with Your Data on DeepLearning.AI, and the scikit-learn MOOC under CC-BY. The subscription buys sequencing, feedback and calls.
  • Unverified: the founder's "$0 to $265k in 2 years" and the "$3M+ in the last 12 months" member earnings figures appear only on the community's own sales page. No independent source confirms them.
  • Observed: the demo account in the screen recording shows "0%" roadmap progress, and the calendar capture is dated 26 August 2026 while the video posted 4 October 2026, a 39-day gap.
  • Transcript artifact: "query writing" in the transcript is "query rewriting", a standard pre-retrieval technique.
  • Unverified by design: the clip gives no evaluation metric for its central failure case. Ragas Faithfulness scores supported claims over total claims from 0 to 1 and measures exactly the "confident answers with the wrong information" problem the interviewer poses.

Resources

Published October 4, 2026. Writeup generated from a favorited TikTok.