Pinchit · live at pinchit.in
A shared workspace where people and AI agents work on projects together: assign a task to either, track ownership, and give agents enough context to act like real team members instead of one-off tools.
The problem
Most AI agents work outside a team's normal workflow. They get a one-off prompt, lose context between tasks, and have no clear ownership, so someone still has to move information by hand between chat, GitHub, and the agent. Giving an agent a real place in the workflow means it has to pick the right specialist for a task, hold credentials safely, and act directly on GitHub instead of just describing what to do next.
What I built
- Autonomous pitchdeck pipeline. A 4-stage agent pipeline (market research → deck architecture → narrative → build) that generates an investor pitch deck end-to-end, streaming live progress into the project chat. Used live in 2 client demos, and one closed the deal.
- Agentic GitHub integration. Ran a full design-spec → implementation-plan → build → test cycle to give agents 13 GitHub tools (PR, issue, and branch lifecycle) inside an LLM tool-calling loop, auto-detecting GitHub intent and switching from single-shot completion to a 10-iteration tool-calling loop. ~15 real PRs and issues processed in production.
- LLM-based task routing. Replaced keyword-overlap matching with an LLM judge that picks the right specialist agent from live candidates, falling back to the original scorer automatically if the LLM call fails.
- Agent catalog at scale. Adopted an open-source agent-agency framework and fine-tuned it into 4 bundles of 5 agents each, spanning engineering, marketing, UX, and research, so the roster covers the business's full range of operational needs instead of one narrow use case.
- Encrypted secrets, checked before use. AES-256 secrets storage with per-secret nonces and a pre-flight check that blocks a task and tells the user exactly which credential is missing and how to add it, instead of failing silently.
- Slack integration. Built HMAC-SHA256 webhook signature verification from scratch, with a replay-window rejection and constant-time comparison, and fixed a cross-service ID mismatch so the Slack backend and the agent-orchestration backend share one project ID.
Where it's headed
Pinchit is turning into infrastructure for teams running several AI agents at once, not a single assistant. That means giving each agent clear ownership of its task, project context that persists across sessions, a visible place inside the team's workflow, and progress a person can actually trace. The goal is a team that assigns work, follows it, and reviews the output from one shared context, instead of re-prompting a separate tool at every step.
Stack
Go · PostgreSQL · Docker · go-chi · OpenRouter · Ollama · NestJS/TypeScript