Put it against a real system
Applied AI architecture provides the production boundary: task routing, model choice, document workflows, controls, and handoff to operators.
See the practice← Research agenda
03 · Intelligence
Model assembly
Research question
Time to assemble the right specialised intelligence for a task.
Why the problem is difficult
The mechanism and its failure modes.
Frontier models, open-weight models, and narrow specialists win different work. A single-model default spends capability, cost, latency, or control where the task does not require it.
The practical problem is assembly: identify the task, choose the right intelligence, carry context across the boundary, and know when the result is good enough to proceed.
Current state
What Anant is doing now.
Document vision and perception is the bounded specialism. The assembly question is defined, but Anant has not published a benchmark, model suite, or performance result. The programme must produce those artifacts before it can claim progress.
Proposed method, output, and evaluation
Proposed method: route a documented task across frontier, open-weight, and narrow specialists while measuring task quality, provenance, control, cost, and latency. Proposed output: a documented routing method and task evaluation harness focused first on document vision and perception. Evaluation: compare task quality, provenance, control, cost, and latency against a documented baseline and task set.
From question to output
Two ways this work becomes useful.
Produce an inspectable result
The result should apply beyond Anant’s commercial scope. Fund the measurement, specification, implementation, and public artifact—not a consulting subsidy.
Discuss funding