Rohit Kochar built the sensors, the controller, the runtime and the reinforcement-learning brain from scratch — then let the stack actuate valves and chillers across 30-plus Indian plants. The receipts: a 52 per cent energy cut at one Mumbai campus, roughly ₹150 crore in six years, and a thesis that India must build products, not just integrate them.
1:06:51
Most industrial AI stops at the dashboard. Bert Labs built everything between the chart and the valve: its own IoT sensors, an edge controller that speaks BACnet and OPC UA, a runtime that lands on any cloud, and a deep reinforcement-learning brain that reads 4,000-plus IO points every second and actuates the best combination of setpoints in real time. Rohit Kochar's proof is concrete — a 52 per cent energy cut at a million-square-foot Mumbai FMCG head office, a ₹7 crore power bill brought down to ₹4 crore, 30-plus deployments across pharma, cement, chemicals and automobile paint shops. Around the technology he wraps a larger argument: data infrastructure must precede physical plant, pricing should follow outcomes rather than headcount, and India's system-integration economy needs the audacity to build products from scratch. Six years in, Bert runs at roughly ₹150 crore with EBITDA above 20 per cent — profitable from day one, by design.
Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.
Dashboard-and-alert solutions plateau at 5–10% energy savings because a human operator still has to act on the prediction. Walmart India's team under Krish Iyer told Bert in 2017 that monitoring alone wouldn't go far and pushed it to close the loop. That demand birthed Bert Optimus and Bert Gemini — real-time AI models that actuate controls rather than recommend actions.
Incumbent HVAC vendors sold separate solutions for the air side and the water side, yet occupancy and ambient temperature on the air side set the load the chillers must serve. Integrating both is why Bert cut 52% off a 1-million-sq-ft Mumbai FMCG head office that already ran a building-management system plus a 2018-vintage optimisation layer. In rupees: a ₹7 crore annual energy bill brought down by ₹3 crore.
When a 40-year-old, ₹1,000-crore Chandigarh pulp-and-paper client wanted Bert paused until its European equipment was resized, Rohit refused the premise. A component swap doesn't change the demand-side ecosystem: parameters stay the same even if values differ, so the trained models keep their learning the way a brain doesn't reset when one input changes. Hence his standing advice to CXOs — digital-transformation infra ahead of physical plant infra.
PID and rule-based control are reactive and single-input-single-output, and no operator with 40 years on one line can correlate thousands of variables. Bert's deep RL reads 4,000-plus IO points every second at a Gujarat caustic-chlorine plant, infers unmeasurable 'soft sensors', and actuates the best combination in real time at 94–99.99% prediction accuracy. Nobody had risked unsupervised learning in plants where a 30-minute pharma deviation cancels an FDA licence — Bert spent five years making it run at 30-plus sites.
The Indian MDs of Analog Devices, Texas Instruments and Qualcomm spent Bert's first two years asking why it built its own controllers when plenty existed. Rohit's answer: only the rigmarole of building hardware and software from scratch earns the understanding on which ingenious work — like running, not just inferring, AI models on the edge — becomes possible. Skip that rigor and you remain a system-integration economy, which is exactly what he says India must outgrow.
Rohit abhorred IT services' input-based pricing; at Hayes, RR Donnelley and Macmillan he sold platform-based services and ran 25–30% EBITDA against competitors' 13–17%, and HOV did $600M of revenue with 11,000 people where peers needed 25–30,000. Bert carries the same fundamentals — SaaS fees plus a share of verified savings, every project self-sustaining — closing around ₹150 crore with 20-plus EBITDA in six years, two of them lost to pandemic access bans.
Accomplished professionals fail at starting up because the feeling of having done enough quietly slows their learning; Rohit's first rule in 2017 was 'I am a nobody', and his learnability quotient had to rise every year. The proof came when COVID emptied commercial buildings: rather than bleed, Bert used the crisis to build its manufacturing platform in 2021–22 — now 30-plus deployments with 20-plus more to go.
Every figure below was said on air — timestamps included, caveats kept.
Digital transformation infra needs to come ahead of your physical plant infra.
— Rohit Kochar ▶ 19:36
A human mind, in spite of 30, 40 years of expertise in one process line, can't fathom the correlations between these hundreds and thousands of variables.
— Rohit Kochar ▶ 16:44
This needs to change — where we have the audacity, and I'm using this word very responsibly, to build a product from scratch.
— Rohit Kochar ▶ 48:30
Put your ego in a trunk and have that trunk in the deepest end of the ocean. You're nobody.
— Rohit Kochar ▶ 1:04:03
A live wedge-finding case: monitoring's 5–10% ceiling, a customer demanding control, and the 52% proof that followed.
Why an equipment overhaul doesn't junk your models: parameters persist, values change — so lay the data rails first.
State space, rewards, soft sensors, 4,000 IO points a second, 94–99.99% accuracy — RL in production, plainly told.
Rohit ignored TI, Qualcomm and Analog Devices MDs and built his own hardware — the rigor that put AI models on the edge.
Operating rules for starting at zero after a C-suite career, and the crisis pivot that built Bert's manufacturing line.
Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.