Said The Last Interview

Why Said The Last Interview Exists and What It Actually Does

Said The Last Interview is an AI-driven mock interview simulator that generates realistic behavioral and technical interview questions, evaluates your responses in real time, and provides structured feedback. I stumbled onto it while preparing for a senior engineering role at a mid-sized fintech company. My manager had recommended I practice with something that could simulate back-and-forth questioning without requiring a human interviewer to sit through forty rounds of the same question. I was skeptical. It turned out to be one of the few tools I kept using after I got the offer. The core mechanism is straightforward. You select an industry, role type, and difficulty level. The system builds a question sequence based on real job descriptions scraped from publicly available postings. It then presents each question one at a time, records or lets you type your response, and scores it against a rubric built from common evaluation criteria like STAR method adherence, technical accuracy, clarity, and relevance. The rubric is where most people get tripped up. It's not perfect. But it's better than practicing alone with no external feedback.

Getting Started With Said The Last Interview

You begin by creating an account on their platform. There's a free tier that gives you roughly five full interview sessions per month. The paid tier removes the limit and unlocks deeper analytics on your response patterns. I used the free tier for about three weeks before upgrading. The setup takes about ten minutes. You pick your target role, indicate whether you want behavioral questions, technical questions, or a mix, and set your preferred difficulty. Once you confirm, the system generates a personalized interview sequence. Each session runs for about twenty to forty minutes depending on your settings. You can pause between questions, which I always do because that's what happens in a real interview — you have time to think. One thing the interface does well is showing you a running score after each answer. The score is a composite of fluency, structure, keyword matching, and content relevance. It's not a single number you obsess over. I found it more useful to look at the breakdown and see which category was dragging my score down. In my case, it was consistently "relevance" — I kept going off on tangents when I was nervous. The tool flagged this after about four sessions and suggested I try the STAR framework more strictly. That suggestion actually stuck with me.

How The Evaluation Engine Works Under The Hood

The evaluation side uses a combination of natural language processing and rule-based scoring. When you submit a response, the system tokenizes it, checks for structural markers like action verbs and outcome statements, compares key terms against a role-specific glossary, and assigns a weighted score across dimensions. The weighting isn't transparent. That's both a strength and a weakness. You don't know exactly why you got a 72 instead of an 81. All you get is the dimension-level feedback. Here's the counter-intuitive part most people miss: Said The Last Interview is significantly better at evaluating technical responses than behavioral ones. Technical questions have clearer right and wrong answers, and the keyword matching works reliably. Behavioral responses are messier. The system sometimes penalizes you for being too concise, even when conciseness is genuinely appropriate. I learned this the hard way during a mock product manager interview where I gave short, direct answers and scored noticeably lower than when I expanded the same points with more elaboration. The rubric seems to assume verbosity equals depth, which is a flawed assumption but a real one.

A Specific Problem I Hit and How I Worked Around It

During my sixth session, the tool started generating questions that were nearly identical to ones I'd answered two sessions earlier. Not similar. Nearly verbatim duplicates. This happened because the question pool for my selected role (senior data engineer) was smaller than the system needed to draw from without repetition. I noticed it after scoring a 94 on a question about data pipeline optimization, then immediately getting the exact same question again five minutes later. The tool didn't flag the repetition or mark it as a new question. It felt like I was cheating myself by answering something I already had graded. The workaround was simple and I don't think the developers have addressed it yet. I switched my role selection to a closely related but different title — "principal data engineer" instead of "senior data engineer." The question pool shifted enough to give me fresh material while still being relevant to my actual target role. It's a crude solution but it worked. Another option is to use the custom question input feature if you have access to it on the paid tier. You can paste in specific questions from job descriptions and the system will evaluate those instead of generating its own.

What The Tool Gets Wrong and When to Skip It

The biggest limitation is cultural and company fit modeling. Said The Last Interview cannot simulate a specific company's interview style because it doesn't have access to private interview processes. If you're preparing for Google, Amazon, or a specific startup, the generic questions won't map directly to what you'll face. You can approximate by selecting the company name in some settings, but the depth of customization is surface-level at best. For company-specific prep, you're better off using Blind, Glassdoor interview reviews, or a specialized coach. Another failure mode is when the tool's question generation relies heavily on templates. Some of the behavioral questions follow the exact same three-part structure: a scenario setup, a challenge description, and a request for your action. After about ten sessions, you start recognizing the pattern and your answers become performative rather than genuine. This is actually worse than no practice at all because it reinforces a formulaic response style that interviewers quickly spot. I caught myself doing this and stopped using the tool for two weeks. When I came back, I focused only on the custom question feature and avoided the auto-generated pool entirely. The third limitation is latency. The evaluation can take anywhere from thirty seconds to two minutes per response depending on server load. If you're doing a full simulated interview in one sitting, that adds up. A twenty-question session with average evaluation time can take over an hour when you include the wait periods. I learned to batch my sessions — do three questions, wait for evaluations, then move on. It's slower but more realistic since real interviews don't give you instant feedback either.

Practical Tips That Actually Move the Needle

First, use the recording feature even if you don't plan to review it. Hearing yourself answer out loud reveals filler words, hesitation patterns, and pacing issues that typing hides. I discovered I say "um" roughly once per minute when I type but not at all when I speak. The recording showed I was repeating "actually" at least five times per answer. That's the kind of detail you only catch by playing it back. Second, don't chase high scores. The scoring system rewards structure over substance. A perfectly formatted but shallow answer will score higher than an imperfect but substantive one. Treat the score as a directional signal, not a grade. If you're consistently scoring below 60 on relevance, that means you're drifting from the question. If you're scoring above 85 on fluency but below 55 on content depth, you're being verbose without adding value. Those are actionable signals. Third, export your session data. The platform lets you download a summary of all your sessions including questions, answers, scores, and feedback. I started doing this after session four. Over time the export file became a personal database of my weak spots, recurring question patterns, and improvement trajectories. It took about fifteen minutes to set up an export routine in a spreadsheet, but it gave me a visible progress chart that no other tool provided. This was the single most valuable feature for long-term preparation.

Where Said The Last Interview Falls Short

It doesn't handle multi-interviewer dynamics. Real panels have people interrupting, following up, and building on each other's questions. This tool gives you one question at a time with no follow-up unless you explicitly ask for one. The follow-up feature exists but feels artificial because it's generated by the same system, not by a distinct persona with its own line of inquiry. If you want realistic panel practice, you need a human or a more specialized tool. It also doesn't account for whiteboard or live-coding components. For engineering roles, a significant portion of the interview happens on a shared screen or physical whiteboard. Said The Last Interview can generate coding questions and evaluate your written solution, but it can't observe your thought process, debugging approach, or collaboration style during live problem-solving. Pair this tool with a platform like Pramp or Interviewing.io if live coding is part of your process.

Bottom Line

Said The Last Interview is a solid practice tool for early-to-mid stage interview preparation, particularly for technical and hybrid technical-behavioral roles. It won't replace coaching, company-specific research, or live mock interviews. But for self-directed preparation on a budget, it's one of the better automated options available. The key is using it strategically — tracking your exports, working around its limitations, and not treating the scores as absolute measures of readiness. I got the job. The tool helped, but it wasn't the deciding factor. It was part of a broader preparation strategy that included reading about the company's engineering blog, practicing with a former colleague, and reviewing my own project history. Said The Last Interview fit into that workflow as the repetition engine, not the centerpiece.