The models are good enough for a large class of enterprise work. What stops systems reaching production is the inability to answer, in a review, six months later, which source supported which claim and what was knowingly let through. That is an architecture problem, and it has patterns.
Harshit Bhatia
Chief Technology Officer, Assert AI · Delhi
The problem
It is “how do you know?”
A pilot clears its evaluation suite and stalls in legal review. A summary is accurate and cannot be defended, because nothing retained the paragraph it came from. A pipeline is audited a year later and the record shows only that it ran, not what it decided or what it let through.
Guardrails inspect the output. Evaluations score the model. Neither produces the artifact a reviewer actually needs: a per-claim account of evidence, and an honest list of the imperfections the system knowingly shipped.
Verification-first design inverts the usual order. The evidence path, the refusal path, and the audit record are designed first. Generation becomes a stage inside that structure rather than the thing the structure decorates.
Verified Intelligence
Confidence is easy to manufacture — a fluent paragraph feels convincing. Trust comes from evidence, from transparency, and from being able to answer the hard question when someone asks why they should believe this.
Verified Intelligence is the approach I use to build enterprise AI systems that can defend their own answers. Not a model, not an agent framework, not a prompting technique — a way of designing the system around the evidence rather than around the generation.
Products
An ingestion-to-publication pipeline that scores sources, drafts briefings, fact-checks every claim against the source text, and refuses to publish when a deterministic gate fails.
Verified
Every claim is checked against the retained source text before it ships, and a model-free gate blocks publication when the check fails.
Other systems
Running in public
Week 33, 202610 Aug - 16 Aug 2026
OpenAI releases GPT-5.6 agent guide; Apple cuts retraining costs via low-influence points; Meta launches local Muse.
Architecture
Product architecture notes for turning visual and operational events into enterprise decision intelligence.
A reference architecture for retrieval, ranking, security, evaluation, and observability in enterprise RAG systems.
A production-oriented architecture for routing, memory, tool access, approvals, and observability in enterprise agent systems.
Writing
What building through 2024 and 2025 taught us about infrastructure, product discipline, and the shift toward agentic interfaces in 2026.
The recurring production failure modes behind enterprise multi-agent systems, from unclear ownership to missing evaluation loops.
A practical playbook for moving LLM applications from prototypes to production systems.
Profile
Harshit Bhatia is the Chief Technology Officer at Assert AI, based in Delhi, India. His work connects AI platform engineering, computer vision, edge AI, LLM applications, and AI security into systems that teams can operate, observe, govern, and scale.
The operating focus is practical: define the right enterprise use case, design the product boundary, build reliable data and model pipelines, add guardrails, and measure whether the system improves business outcomes.
