Before Convilyn was a workflow platform, it was a keyboard. This is how a simple belief — that AI should be frictionless for everyone, not just power users — led us to rethink everything twice, and why that conviction still drives every decision we make.
The two-sided platform vision — where users with problems and developers who can solve them find each other on a shared, stable architecture.
Two products, one vision. How the workflow foundation we build today becomes the intelligence layer for a much simpler AI experience tomorrow.
Real scenarios: upstream trigger → workflow → downstream result
Your boss forwarded a thread and said "handle this." Upload the context documents — get a ready-to-send email draft with the right tone.
Upload a stack of invoice PDFs — line items, tax, and totals extracted into an accounting-ready spreadsheet. No re-entry.
Upload your images or brief — get 3–5 platform-ready post variations for Instagram, LinkedIn, and X with captions and hashtags.
Upload any contract — get a clause-by-clause analysis with risk scoring and obligation summaries before the signing deadline.
Saw something inspiring? Got told to "make it social"? Upload the source doc — get platform-ready content for every channel.
Upload any notes file — the agent extracts key concepts into Q&A pairs organized by topic. Study smarter, not longer.
Upload a lecture or tutorial video — get timestamped study notes with key concepts, summaries, and review questions automatically.
Resume + job description in → role-specific question bank, model answer frameworks, and skills gap analysis out.
Upload your resume and the JD — the agent extracts requirements and writes a cover letter that speaks directly to the role.
Upload your resume. The AI analyzes it, suggests job directions, and generates tailored resumes + an application tracker for each.
Step-by-step guides for users and developers
One install, no API key, no network. What survives the conversion, which images get dropped, why PDF tables need ruling lines, and how to catch scans that exit 0 with an empty file.
Official TypeScript and Python SDKs with typed methods for every API resource. Auto-generated from our OpenAPI spec.
Open docs
From uploading a file and selecting a workflow, through interactive checkpoints, to downloading your finished output. Walkthrough with AI Job Search Assistant.
A step-by-step guide to fast-track format conversion — image, document, PDF, OCR, video transcription, and more.
Every Convilyn workflow starts and ends with a file. That's not a limitation — it's the design principle that keeps AI focused, measurable, and composable.
Two products, one strategy. How proven AI workflows become the intelligence layer for a much simpler, more embedded everyday experience.
Our vision for a two-sided platform where users with problems and developers who can solve them find each other on a shared, stable architecture.
Before Convilyn was a workflow platform, it was a keyboard — a vision for frictionless AI for everyday people, not just power users.
Every deploy workflow was disabled, and one job could still have shipped production on the next merge. A failing test was all that held it back. Two paths I thought were off, and the tests that now keep them gone.
A no-LLM gate that rejects any amount, large number or clause reference not found verbatim in the source, with an LLM judge that only watches. What it stopped on its first day, and how each case was resolved.
Seven per-marketplace spreadsheet tools replaced by generic inspect and fill tools, with the agent mapping columns. A real template broke the tools in three layers, and the agent ignored an overwrite warning. The fixes, and what the trade costs.
Running each turn without the chat history keeps the agent's input small and readable. The price was two memory bugs: asking for clarification twice, and forgetting the request after a clarification. Both fixed inside the design.
Eighteen model calls, five turns, nothing saved, and a slower loop around the whole project where every wall got one more layer. What we hit building an agent builder, and what to decide before you build a harness.
An agent guessed block names a tool description never listed. That led to five places where my own text handed it tool ids that didn't exist, and two fixes: generate the names the agent reads, check the ones only tests read.
Six agent tools moved behind an MCP server with a fail-open fallback. A field-path parity check caught every mismatch before the switch; after it, five of six calls were silently served by stand-ins until the logs gave it away.
Resuming after interrupt() re-ran a node and submitted a paid job twice, async durability lost a checkpoint in a crash, an approval never got used up, and a retry replayed its own failure. Measured offline, with the design rules that follow.
Any node can write any channel, a retry on the same thread goes through the reducers, and a subgraph checkpoints into the parent's store. Measured offline on langgraph 1.2.11, with the guards that follow.
LangChain's input_tokens includes the prompt cache; Anthropic's pricing assumes it doesn't. Feeding one into the other priced every cache hit twice and turned a token limit into turns × prefix size.
An unanswered tool call that Bedrock rejects, and a langchain-aws upgrade that removed null from tool schemas while every test stayed green. What the adapter sends, how to pin it, and where each rule belongs.
99 tools bound per turn, a median of two legal. Letting the model search for tools cut schema tokens by two thirds and did not make runs cheaper, because the tool list was already prompt-cached.
Emptying one list turned 177 passed into 9 skipped, exit 0. Why an empty input makes assertions true, the suite-level versions of the same failure, and the probe that finds them.
Five production loops that threw nothing and stayed under every timeout. What each one varied, why LangGraph's idle timeout misses the commonest one, and the detectors that catch them.
Two prompt fixes and three runs on identical inputs left the grounding score flat. The model had been handed "parsed 6 sections" instead of the résumé. How to tell noise from effect, and how to print what a LangChain model actually received.
Prompt chaining as a typed three-stage pipeline — Planner picks the tool, Composer writes the artifact, Verifier scores it. Phase-FSM middleware governs the hops; AgentDecision is the contract between links; a pre-store quality gate is the line of last defense before S3.
The two-layer prompt architecture for the Workflow Builder — Layer A platform-fixed, Layer B as untrusted data — drawing on Anthropic's prompt-injection research, OpenAI's Model Spec, and the OWASP LLM Top 10.
Single-shot rubrics catch wrong answers; trajectory rubrics catch agents that take a 12-step detour to get there. With a nightly auto-issue loop that turns drift into actionable GitHub issues.
Manual evaluation is the silent productivity tax of every LLM team. The baseline-and-gate axis system that runs in CI, posts pass/fail comments on PRs, and made eyeballing obsolete.
EMF for tool latency, retry classification across four explicit categories, multi-agent HandoffRecord, and a Sentry DSN URL guard we should have had in v8.
After three months of parallel observability paths, we collapsed everything to one OTel seam, one event taxonomy, and one redaction policy. The seven-phase charter that took us from alpha to beta.
LangGraph state machines, MCP capability servers, interactive checkpoints, and SSE streaming — plus lessons from studying Claude Code and Codex CLI open source.
A 2-tier testing framework: lifecycle compliance (does the workflow complete correctly?) and goal-oriented content validation (does the output achieve its purpose?).
Six event types, milestone-based progress calculation vs iteration fallback, connection recovery via event IDs, and the progress reporter protocol.
The formula max(100K, max_iterations × 10K) — why a fixed budget fails at both ends, and how it interacts with the iteration limit.
How the agent manages 120-iteration conversation history — the trimming strategy, what gets preserved vs discarded, and the history loss bug we fixed.
The threshold formula that switches the LLM from mandatory tool invocation to autonomous reasoning — and why control-flow tools are excluded.
One template drives every workflow in the catalog. The decision framework for when to write a 351-line specialized prompt instead — and why.
Capability-based MCP servers, the six file argument strategies, reference IDs for binary tool outputs, and the three control-flow tools.
How JSON spec files define nearly 100 distinct AI workflows without any new agent code — schema, preflight validation, and spec versioning.
How Goal Lane models agent execution as a 5-node graph with conditional routing — reason, call_tool, request_input, nudge, and finish.
Why Convilyn splits into two completely separate execution paths — and why a unified pipeline would have made both worse.
10 real-world scenarios — each one walks from the upstream trigger through Convilyn tools to the downstream deliverable, so you can see exactly where we fit.
Browse all playbooksNearly 100 AI workflows. Two lanes — Turbo for instant format conversion, Goal for multi-step intelligent processing. Free tier, no credit card required.
Try Convilyn FreeFree tier · No credit card needed.
