Imagine you're in the world's most extraordinary library—vast, elegant, and meticulously curated. Books tower above you, each shelf organized by themes, disciplines, and wisdom. But there's a twist: the shelves extend far beyond journals and textbooks, weaving through startups, multinationals, blogs, conference videos, and insightful YouTube channels. This isn't your traditional literature review—it's a literature review 2.0, carefully engineered and augmented by generative AI.
This vision is precisely what my colleagues and I set out to achieve when we decided to explore how generative AI was reshaping engineering. Our goal wasn't merely to read the literature, but to orchestrate a grand symphony of knowledge, with us conducting and the machines playing every note precisely, tirelessly. Our human expertise—accumulated over 70 years of academic rigor, industry experience, and Silicon Valley innovation—provided the deep, guiding intuition. Generative AI, especially tools like Perplexity, ChatGPT, Mistral, and Gemini, brought breadth, speed, and precision, enabling us to process more than 900 rigorously vetted sources in just a couple of days.
We understood early on that the secret wasn't to let the machine roam freely, hallucinating with unchecked creativity. Instead, we created a taxonomy from our deep domain knowledge, rigorously reviewed and refined through adversarial competition between multiple LLMs—like intellectual gladiators sparring to reveal the best path forward. With that battle-tested taxonomy in hand, we carefully orchestrated a layered exploration: field by field, subcategory by subcategory, each step peeling back complexity like layers of an onion, guided by specific, targeted questions. Perplexity, tethered firmly to reality by real-time web searches, delivered us curated answers—answers that sparked new queries, building a virtuous cycle of discovery and insight.
But the magic truly accelerated when we turned toward writing. Writing can be excruciating—slow, solitary, and iterative. So we evolved it. We adopted an academic-inspired feedback loop, augmented by relentless, high-speed iterations powered by fine-tuned LLMs, trained to write precisely in our voice and style. Using few-shot prompting combined with maximal context, we transformed numerous, articulated, but rough bullet points and skeletal outlines into polished narratives indistinguishable from our human-crafted prose. Each draft was rigorously reviewed by competing LLMs, our virtual editorial board, ensuring clarity, resonance, and analytical precision.
The result? A 350-page detailed map of generative AI's impact on engineering, thoughtfully analyzed and distilled into a compelling white report, and elegantly coded into an interactive website using TypeScript, React, and Next.js—all completed in mere weeks. We even transcended language barriers, effortlessly translating and refining our work into French, broadening our impact.
In short, by fusing human intelligence, intuition, and expertise with AI's raw power and scalability, we reimagined how research can—and perhaps should—be done. Our methodology didn't merely augment our capabilities; it elevated the very nature of scholarly inquiry itself, producing insights at a velocity previously unimaginable.
1The literature review was conducted in October 2024, prior to the release of "deep research" functionalities by major LLM providers. Between December 2024 and February 2025, we tested these new tools (from ChatGPT, Gemini, and Perplexity) against our manually orchestrated method—based on a custom taxonomy, expert-driven decomposition, and iterative refinement. Despite deep research retrieving 200–300 sources per query, our approach consistently produced higher relevance and depth. Only when deep research was applied to well-crafted, expert-defined prompts did it approach comparable quality, suggesting it is a useful evolution but not a replacement for a human expert-led exploration.
