An answer has
to hold up.
Reconstructed interface study for this exhibition. It illustrates the product idea, not a screenshot or a live system.
01 / AI PRODUCTS + AGENT-LED ENGINEERING
Colligence AI
A bounded assistant agent loop, evidence-grounded retrieval, and agent-led engineering across Aikyat Mail and Chat at Colligence.
Current source: bounded agent loop
The current assistant code uses an iterative tool-calling loop, and the committed staging and production configurations enable it. This case describes repository implementation and documented evaluation, not a fresh verification of live deployment revisions or public release.
What if the highest-scoring AI answer is still the wrong product?
My part in it
My work connected product decisions with implementation across Mail and Chat. I directed coding and evaluation agents, set review gates, challenged assistant behavior, and worked through the engineering needed to make those decisions concrete. Agents did implementation work; my contribution included direction, constraints, review, and judgment.
What stayed with me
The important question was not whether an answer sounded complete. It was whether the answer was current, supported by evidence, and appropriate for the person asking. That changed what we built and what we chose to measure.
Assistant architecture: a bounded agent loop
The current implementation lets the model choose approved tools, read their results, and decide what to do next. It can take another tool round rather than execute a single precomputed plan. Precise deterministic guards handle recognized simple requests; other turns enter the loop. Exact mailbox questions use structured database tools, while topical questions use hybrid retrieval. The server owns permissions, tool validation, execution limits, and stopping conditions.
Throughline: challenge the apparent winner
The Chat search work included a functioning retrieval-and-synthesis prototype and a backend exercised against Postgres and pgvector. We paused an apparent evaluation winner when completeness rewarded stale or unsupported answers. The next work tested currentness, permissions, attribution, provenance, and when to abstain.
Working with agents, not outsourcing judgment
I used distinct product, architecture, engineering, and cross-model review perspectives to challenge the implementation and the evaluation itself. The model under test did not grade its own answers. Builder agents could propose and implement; findings still had to survive review. This describes the development and evaluation method, not a claim that the customer-facing product used a multi-agent runtime.