Trinetra

Trinetra -> the third eye for your production systems.

Inspiration

Every SRE team has the same incident story: the first 15-20 minutes are spent figuring out what is actually wrong while alerts pile up, Slack gets noisy, and teams rediscover the same root causes again and again.

We built Trinetra because incident response should not lose institutional memory at the exact moment it matters most. Instead of making another alert dashboard, we wanted an agentic system that can reason like a senior on-call engineer: checking metrics, logs, history, runbooks, and confidence before deciding whether to repair automatically or ask a human.

The name Trinetra means “third eye,” which fits the goal: giving engineering teams deeper visibility across system signals before humans have to manually search for the answer.

What it does

Trinetra is a Qwen-powered autonomous incident response system. When an incident occurs, it uses a commander agent, specialist investigation agents, adjudication logic, triage, and a deterministic remediation gate to decide the safest next action.

It can diagnose a broken demo storefront, identify likely root cause, match the issue to approved runbooks, request Slack approval for risky fixes, execute allowlisted remediation tools, verify recovery, and keep a full audit trail of every decision.

How we built it

We built Trinetra with Node.js, Qwen Cloud Model Studio through DashScope-compatible APIs, Slack interactivity, Docker, Render, and a structured multi-agent backend.

The backend includes Qwen agent calls, Alibaba Cloud/Qwen configuration proof, incident orchestration, safe tool-calling remediation, Slack approval buttons, audit logging, and a live demo storefront that can intentionally fail and be repaired.

Code proof: https://github.com/sairam782/trinetra/blob/main/backend/cloud/qwen-client.mjs

Alibaba/Qwen integration proof: https://github.com/sairam782/trinetra/blob/main/backend/cloud/alibaba-client.mjs

Challenges we ran into

The hardest part was making the system safe instead of just impressive. Early agents were too confident, so we had to improve confidence calibration and keep the final remediation gate deterministic instead of letting the model decide everything.

Slack approval also took work because interactive buttons require signed request handling, correct public URLs, reinstalling the Slack app after scope changes, and a backend that can process Slack payloads securely.

We also had to balance demo reliability with honesty: the app is hosted on Render for the public demo, while the Alibaba/Qwen proof is shown through the Qwen Model Studio integration and Alibaba Cloud configuration/code path.

Accomplishments that we're proud of

We are proud that Trinetra is not just a chatbot or dashboard. It actually follows an incident pipeline: investigate, reason, choose a runbook, request approval when needed, execute a safe tool, verify the fix, and preserve an audit trail.

We also got Slack approval buttons working end-to-end, so human-in-the-loop remediation feels real and usable.

What we learned

We learned that autonomous remediation needs strict boundaries. Models are useful for reasoning and planning, but production actions should go through deterministic gates, allowlisted tools, verification checks, and audit logs.

We also learned that multi-agent systems need confidence calibration and historical weighting. Simply asking multiple agents and picking the loudest answer is not enough.

What's next for Trinetra

Next, we want to connect Trinetra to real production observability sources, persist memory in a managed database, expand the runbook registry, support more cloud deployment targets, and make the Slack approval and audit workflow production-ready for real SRE teams.