Ashish Dhiman

Founder & CEO, Cerebro — your company brain for AI agents. Every agent reads from one shared vault. MIT-licensed.

[Ashish Dhiman @ashishdhiman23]

We made the same architecture decision THREE times.
Slack. Notion. PR. Each time the next person re-litigated it.
Then I noticed: it's even worse with AI agents.
So we built Cerebro — your company brain for AI.
🎥 1-min demo + thread 👇

[Ashish Dhiman @ashishdhiman23]

@mweinbach If you auto-convert, ops stay on CPU. Same thing we see with QNN: graph says HTP, a few ops silently don't. We assert NPU residency on real Snapdragon in CI. https://t.co/Z3OHGis08T

[Ashish Dhiman @ashishdhiman23]

@cactuscompute @coolgadgetru If you wrap QNN for glasses, watch the silent path. HTP is advertised, unsupported ops land on CPU, battery and thermals go. We gate that on real Snapdragon. https://t.co/Z3OHGis08T

[Ashish Dhiman @ashishdhiman23]

Edit a prompt in n8n and you've deployed to production.

No review gate. No failing test. The workflow still returns 200.

It just agrees to things it used to refuse.

[Ashish Dhiman @ashishdhiman23]

Saw research this week showing a phone benchmarked fine on a single forward pass, then throttled 44% two inferences into a real loop. That's exactly why Behavioral-Gate doesn't sign off a single number — it signs a verdict after a real run on real Jetson/Qualcomm silicon. Building EdgeGate around workloads, not forward passes.

https://t.co/R6FpNmKpXl

[Ashish Dhiman @ashishdhiman23]

The model you red-teamed in FP16 cloud form is not necessarily the model you shipped after quantization.

Safety behavior can move with quant format, runtime, prompt template, driver, and context length.

For device LLMs, run the frozen eval set on the actual artifact.

#OnDeviceAI #AISafety #ModelOptimization

[Ashish Dhiman @ashishdhiman23]

EU AI Act logging is easy to say and hard to prove.

A normal app log can tell you an AI system produced an output. It cannot prove the record was not edited after the fact.

For on-device AI, the evidence needs a model hash, device/runtime path, gate verdict, and a signature over the bundle.

#AIGovernance #EdgeAI #AuditTrail
https://t.co/WaPxL9OzLh

[Ashish Dhiman @ashishdhiman23]

Dynatrace is paying $915M for Arize. AI observability is a real category with real multiples now. Worth noticing what that category actually covers: it tells you what your agent did — traces, evals, hallucination rates, latency. It doesn't tell your next agent what the last one decided, or why. Those are different problems filed under the same word. Observability is a record for humans. Memory is a record other agents can act on

[Ashish Dhiman @ashishdhiman23]

There's a paper on exactly this — arXiv 2608.05223, out this month. They synthesized 2,826 malicious skill files and ran them against real agents: Gemini CLI was compromised in ~96% of runs, Qwen Code ~72%. The number that should worry people: the agents raised a safety concern in only 1.99% of runs. Susceptibility barely changed based on which model authored the attack. Reading a skill's README is not a trust decision an agent can make alone

[Ashish Dhiman @ashishdhiman23]

Agreed that the leverage moved from prompts to the system around the model. The gap I keep hitting: the loop learns something real on iteration 4 — this API paginates weirdly, this test is flaky — and all of it evaporates when the loop exits. The next run rediscovers it from scratch. Feedback that only survives inside one loop isn't feedback, it's throughput. Where do you put state that outlives the run?

[Ashish Dhiman @ashishdhiman23]

The EU AI Act's high-risk deadline moved: Annex III to Dec 2, 2027. Annex I (vehicles, medical devices) to Aug 2, 2028.

It's now enacted law (in force July 27), not a proposal.

The delay is runway, not a reprieve. The evidence requirement didn't move with it.

https://t.co/KhpR6xy11e

[Ashish Dhiman @ashishdhiman23]

NVIDIA's Jetson Thor case studies: moved from a 64GB to a 32GB Orin without compromising performance.

Performance measured how? A benchmark proves it's fast. It doesn't prove the robot still behaves the same.

Compute density solves cost. It doesn't prove behavior didn't change.

https://t.co/WP0uAEgYuL

[Ashish Dhiman @ashishdhiman23]

Jetson is now a real EdgeGate path, precisely stated: Behavioral Gate is live on Jetson via llama.cpp/CUDA. Perf/profiling gates are still Qualcomm today.

That distinction matters. Evidence beats broad multi-silicon claims. #OnDeviceAI #AIGovernance #EdgeAI

[Ashish Dhiman @ashishdhiman23]

An NPU efficiency result is not just a benchmark win. It is a gate-design warning.

If one unsupported op pushes a workload back to CPU, heat and power change before accuracy does.

Gate compute split, layer placement, and median-of-N behavior on the device. #EdgeAI #NPU #ModelOptimization

[Ashish Dhiman @ashishdhiman23]

FleetML deploys models to your edge fleet. EdgeGate decides whether they're allowed to.

Every deploy through FleetML can now carry a signed evidence bundle — proof the model was actually validated on real hardware (Snapdragon, Jetson), not simulated.

→ FleetML pulls the bundle and re-verifies the Ed25519 signature itself — it doesn't trust the network, it trusts the bytes
→ Confirms the model checksum matches the registry
→ Refuses the deploy if the signature, status, or checksum don't line up
→ One click from an EdgeGate run opens FleetML's deploy modal, pre-filled

No signature, no pass, no deploy. The gate is the default, not an afterthought.

Proven end-to-end: real device run → signed bundle → FleetML pulls + verifies → gate opens → live on a Jetson Orin in the fleet.

#EdgeAI #MLOps #FleetML #EdgeGate #DeviceFleet #Snapdragon #Jetson

[Ashish Dhiman @ashishdhiman23]

@FireworksAI_HQ The tooling around training has come a long way. It feels like deployment tooling is finally starting to catch up as well. Once models are fine-tuned, the next challenge is making sure they behave consistently across different runtimes and hardware.

[Ashish Dhiman @ashishdhiman23]

@Qualcomm_Dev @Qualcomm Nice to see more capable edge AI hardware becoming available. The hardware story is getting strong—now the challenge is making sure optimized models behave the same across deployment targets.

[Ashish Dhiman @ashishdhiman23]

@NVIDIAAIInfra Benchmarks answer how fast. Deployment also needs to answer did the optimized model still make the same decisions? Performance and behavioral validation solve different problems, and production systems need both.

[Ashish Dhiman @ashishdhiman23]

@FireworksAI_HQ Training infrastructure has improved dramatically. I think deployment validation is the next bottleneck.

We can fine-tune and optimize models quickly now, but proving an optimized model still behaves like the reference model in production is becoming just as important.

[Ashish Dhiman @ashishdhiman23]

@launch_llama We're building EdgeGate because deploying AI at the edge shouldn't mean guessing whether model optimization changed behavior. We make that measurable.

https://t.co/Z3OHGis08T

[Ashish Dhiman @ashishdhiman23]

@Joi2James We're building EdgeGate because deploying AI at the edge shouldn't mean guessing whether model optimization changed behavior. We make that measurable.

https://t.co/Z3OHGis08T