Joshua Fox — AI Product / GTM Leader | PM/PMM | AI Agents / Voice AI | Ex-Deepgram, Avathon (formerly SparkCognition), AT&T
AI Product / GTM Leader | PM/PMM | AI Agents / Voice AI | Ex-Deepgram, Avathon (formerly SparkCognition), AT&T
Joshua Fox ranks #312 of 14,983 LinkedIn creators in Computer Software, and is a standout voice in United States. They have 1.8K followers and published 6 posts in the last 30 days at a 48.2% average engagement rate.
- 1.8K followers
- 6 posts / 30d
- 48.2% avg engagement
- — follower growth / 30d
The roast
Joshua, you claim to speak fluent AI because you spent two decades making machines sound human, yet you’re the only person on this platform whose personality has the exact same tone, cadence, and soulless lack of conviction as an automated customer service chatbot.
About Joshua
TL;DR: I'm a product guy who speaks fluent AI — and makes it sound human. The Longer Version: I am a technical product executive with nearly 20 years of experience in B2B SaaS and subscription-based service platforms, specializing in product management and marketing for AI/ML solutions. Since earning my M.S. in Computer Science with an AI/ML focus from Georgia Tech, I have dedicated my career to transforming complex technologies into market-leading products that drive growth and innovation across sectors—from telecommunications and media to industrial AI. My journey began in engineering roles at companies like National Instruments and ClearCube Technology, where I mastered the fundamentals of building and optimizing systems. Early on, I discovered my passion for bridging the gap between technology and business, which led me to launch my first startup and later transition into product management. These formative experiences taught me the power of combining deep technical expertise with strategic vision and compelling storytelling to turn ideas into reality and drive product success. Over the years, I have refined my leadership skills at major organizations such as AT&T and at forward-thinking startups like SparkCognition and Deepgram. At Deepgram, I executed innovative go-to-market strategies for generative AI models that power many of today's leading conversational AI products, resulting in significant ARR growth. My ability to distill complex technical concepts into clear, impactful narratives has been a key driver of market differentiation and business success. Looking ahead, I am passionate about advancing the AI space at innovative, high-growth companies. I am eager to leverage my technical acumen, strategic product leadership, and data-driven insights to create solutions that push technological boundaries while enhancing customer experiences and driving meaningful business outcomes. I invite you to join me on this journey as I continue to explore new frontiers in AI and share my findings and perspectives here on LinkedIn, as well as play my part in shaping the future of technology and its manifestations in the products I am lucky to help build and bring to market.
Highlights
- Top Engager — 48.19% rate · top 1%
- Top 5% in Computer Software — Ranked #60 of 4267 creators
- Top 5% in United States — Ranked #117 of 5205 creators
- High Impact — 851 avg engagements per post · top 5%
Recent posts
The biggest mistake in agent engineering today is treating skills as prompts. Teams write a better instruction, add a few examples, maybe run a prompt optimizer, and hope performance improves. When it does, nobody knows exactly why. When it regresses, there is no rollback mechanism, no validation boundary, and no reliable notion of progress. The authors of “SkillOpt” from Microsoft and Shanghai Jiao Tong University start from a different premise. If skills are the main adaptation layer for frontier agents, then skills should be trained the same way we train models. That sounds subtle, but it
303 reactions · 5 comments · 26 reposts
Fine-tuning LLMs costs millions in compute and destroys what models already know. We just proved there's a closed-form alternative that fixes both, runs on a forward pass, and doubles as persistent memory. Open-sourced today. To fine-tune Qwen2.5-7B on GSM8K, LoRA inserts 4.4M trainable parameters per task, adds heavy inference latency, and catastrophically forgets unrelated capabilities (we measured a 16.3pp drop on held-out code generation after GSM8K adaptation). We proved one SGD step has an exact dual: a small signed controller on the forward pass that reproduces the supervised update w
2.0K reactions · 74 comments · 165 reposts
For decades, enterprise technology strategy had one answer. Centralize. Centralize the team. Centralize the tools. Centralize the development. That worked when the scaling challenge was technical complexity. AI changes that model completely. Because agents are increasingly becoming the execution layer sitting on top of operational systems. Technical complexity was expensive, specialized, and difficult to maintain. Centralization made sense because organizations couldn't distribute that level of staffing and technical capability across the business. SaaS accelerated the model even further.
711 reactions · 193 comments · 54 reposts