Podcast Summary: "Building the 'See Something, Say Something' AI for Every Camera"

Podcast: AI + a16z
Date: December 16, 2025
Host: Martin Casado (a16z)
Guest: Shikhar Shrestha (CEO & Founder, Ambient)

Overview:

This episode dives into the transformation of the $100 billion physical security industry as AI moves from passive camera surveillance to "agentic" systems that can not only detect security incidents, but reason about context, run forensics, and even launch automated, real-time responses. Shikhar Shrestha shares how his personal experiences inspired him to build Ambient, and how his company leverages AI vision-language models (VLMs) to create proactive, privacy-conscious security for both vast corporate campuses and high-profile individuals.

Key Discussion Points & Insights

1. Founding Story & Inspiration

• Personal Trauma as Catalyst:
  Shikhar recounts being a victim of an armed robbery at age 12, staring into a security camera and hoping for help that never came.

  "I'm just staring at the camera and I'm just hoping that someone's watching and will come and help us. Nobody's watching and, you know, nobody helped us." — Shikhar Shrestha (00:49)

• Early Vision:
  Partnered with CTO Vikesh in 2016, inspired by the rise of image-captioning AI, with a vision to make every camera a proactive witness that triggers real-time interventions, not just records footage.

2. The Value of AI in Physical Security

• Security is traditionally reactive: tons of cameras, but no one can monitor them all.
• AI turns every camera into a "see something, say something" node, flagging precursor behaviors—thus making security proactive:

  "There's always suspicious precursor behavior that's going on... if you notice something suspicious and you tell the security team... they can go respond... and prevent the bad thing from happening." — Shikhar (05:06–05:56)

• Human monitoring can't scale with today’s camera-dense environments; AI fills the gap.

3. Layers of Value: Prevention, Response, Forensics

• Prevention: Real-time detection of precursor actions (e.g., lockpicking, unauthorized entry).
• Immediate Response: AI can suggest or automate intervention steps (e.g., lock doors, call police).
• Forensics: AI dramatically reduces the time to investigate after an incident.

  "You can just tell the AI a laptop is stolen and it'll... build a whole trail of what actually happened during the incident for you." — Shikhar (08:17)

4. Technical Evolution: Vision Language Models (VLMs)

• Ambient’s system began with early deep captioning models, now leverages advanced VLMs.
• VLMs combine computer vision and large language models for deep scene understanding and context:

  "These reasoning vision language models... are better than humans. We’ve hit that milestone." — Shikhar (11:09)

• Shikhar distinguishes between old AI, which lacked context (e.g., couldn't differentiate a person falling ill vs. someone tying their shoe), and new models that reason about cause and effect:

  "Somebody drawing on a whiteboard versus somebody like tagging the wall... It looks exactly the same... The newer VLMs can do that [context reasoning]." — Shikhar (13:13)

5. Building Proprietary, Efficient AI

• Ambient recently launched their own VLM, Pulsar, trained on real security camera footage (not generic internet images), making it far more accurate and 50x more efficient than general purpose models:

  "These large models... can cost five to $10,000 a month for continuous inference... So, we've built our own VLM..." — Shikhar (14:03)

6. Addressing Privacy & Bias

• Ambient intentionally avoids facial recognition; focuses on context and behavior.
• The model is prompted with targeted queries (e.g., "Is someone brandishing a weapon?") rather than vague, holistic surveillance, reducing bias and privacy risks:

  "We don't actually do any facial recognition... We want to look for that suspicious thing happening... and respond to prevent the incident instead of identifying who they actually are." — Shikhar (15:30)

7. Operations & Human-in-the-Loop

• Ambient utilizes a 24/7 operations team to review low-confidence AI alerts. This human-in-the-loop approach ensures accuracy, provides valuable "hard negative" training data, and has become an industry standard.
• Efficiently scaling this team has become a competitive advantage (moat).

8. Vision for the Future: Agentic Security

• Shikhar outlines a roadmap from accurate detection, to contextual assessment, to automated real-time response (e.g., AI initiating lockdowns).

  "Almost like a real-time assistant where it tells the operator, 'Hey, I just saw a weapon brandished outside the building.' First, lockdown the building, then lock the elevator, then call law enforcement..." — Shikhar (19:03)

• Eventually, humans may override or trust the AI to autonomously run SOPs (standard operating procedures).

9. Market Segmentation & Customer Base

• Ambient serves large complex enterprises (corporate campuses, hospitals, critical infrastructure) as well as high net worth individuals. Interestingly, executive protection emerged organically.

  “High net worth individuals was more accidental… it’s actually become a seven figure... business for us.” — Shikhar (23:19)

• The same platform supports both enterprise and "prosumer" segments with minimal customization.

10. Real-World Incidents & Case Studies

• Battery Fire at Aerospace Client: AI detected an outdoor fire before it escalated, saving critical infrastructure. (25:06)
• Prevented Break-In: After deploying a "fence breach" detection, an intruder was caught within five minutes due to a real-time AI alert. (26:18)

  "Everybody... is on site, you know, the police department is on site and they actually ended up apprehending the perpetrator." — Shikhar (26:40)

11. Navigating Hardship: Covid & Industry Shifts

• Covid shut down Ambient's main vertical (corporate campuses); the company pivoted to verticals that remained active (museum security, data centers).
• Post-pandemic, security teams want remote, cloud-based operations—a trend accelerated by necessity:

  “There was a realization... having a cloud-based product where you can secure a site in a different part of the world... I think accelerated the trend from on-prem products.” — Shikhar (29:24)

12. Go-To-Market & Building a Category

• Selling subscription AI software into a hardware-oriented, risk-averse industry required visionary evangelism and education.
• The traditional sales channel (built for hardware distribution) is slowly adapting; now, 20–30% of Ambient’s pipeline comes via channel partners.
• Early fundraising was tough due to investor “scar tissue.” Now, with market validation, VCs recognize physical security as AI’s next “fertile ground.”

  "...if you think about where you can apply AI and unlock like billions in productivity value, physical security is just such a great space to do it." — Shikhar (35:36)

13. Personal Leadership Lessons: Building on "Hard Mode"

• Shikhar embraces the challenge, referencing The Hard Thing About Hard Things as inspiration:

  "There's like a couple of these levers you turn on. If you turn on all of them... you're just building the company in hard mode." — Shikhar (37:27)

• Advice: Choose a problem worth solving, expect difficulty, and let the potential impact keep you motivated through hardships.

Notable Quotes & Memorable Moments

• “If you notice something suspicious and you tell the security team... they can go respond... and prevent the bad thing from happening.” — Shikhar Shrestha (05:18)
• “These reasoning vision language models... are better than humans. We’ve hit that milestone.” — Shikhar (11:09)
• "The biggest value driver for us is because we can detect so many things... almost agentically detect threats which you just can't do with one siloed detection alone." — Shikhar (21:28)
• "If you pick a problem that is... valuable enough to solve where the summit... is meaningful... just expect it to be hard all the way." — Shikhar (38:04)

Key Timestamps for Reference

• 00:00–01:30: Shikhar’s robbery experience; inspiration for Ambient
• 04:54–06:23: Why AI improves security over traditional systems
• 07:56–09:14: Prevention, response, and forensics with AI
• 09:46–11:51: Early tech journey and transition to advanced VLMs
• 13:13–15:18: Contextual awareness, proprietary model (Pulsar), and efficiency
• 15:18–16:42: Privacy and bias management strategies
• 17:09–18:45: Human-in-the-loop operations for robust detection
• 18:58–20:24: Future: Autonomous, agentic physical security workflows
• 21:43–23:39: Customer segments: enterprises, high net worth individuals
• 25:06–27:46: Detailed incident case studies
• 28:17–30:22: Covid pivot and acceleration towards cloud-based security
• 32:29–34:55: Go-to-market lessons, channel partnerships, sales process
• 36:46–38:41: Leadership advice: Resilience, embracing difficulty

Conclusion

This episode is a compelling, technical, and thoughtful look at how AI is fundamentally reshaping physical security—one camera and one incident at a time. Shikhar Shrestha’s personal motivation, technical depth, and entrepreneurial tenacity shine through as he and Martin Casado discuss not just building advanced technology, but also creating a new market and category entirely.