The Problem
You've got a slow service. You run perf record, or go tool pprof, or Java Flight Recorder. Now you have a 50MB file full of stack traces. You open it in a flame graph tool and... stare at colored rectangles, hoping the bottleneck jumps out at you.
I've been there. Too many times.
The Solution: AIProfiler
Upload any profiler file -> pick an AI provider (OpenAI, Anthropic, NVIDIA) -> paste your API key -> get a ranked list of bottlenecks with actual code fixes.
Supported Formats
- Linux
perf.data/perf scriptoutput - Go
pprof(CPU, heap, block, mutex) - Java Flight Recorder (
.jfr) - Chrome DevTools (
.cpuprofile,.trace) - Speedscope (
.speedscope.json) - Python
cProfile/py-spy - .NET (
.nettrace) - Node.js
--cpu-prof - And more...
What You Get
- Severity-ranked bottlenecks -- not just "this function is hot" but "this loop allocates 40% of memory"
- Profile evidence -- exact lines, call stacks, sample counts
- Copy-paste fixes -- in your language, with explanation
- No dashboard to learn -- upload, read, apply
Privacy & Cost
- Free -- no paid tiers, no signup
- Your API key -> forwarded once to provider -> never logged or stored
- Profiles auto-deleted after 7 days
- In-memory processing -- nothing written to disk
Example Output
🔴 CRITICAL: json.Marshal in hot path (23% CPU)
Evidence: 1,847 samples in encode.go:42 -> json.Encoder.Encode
Fix: Replace with json-iterator/go or pre-allocate buffer
go
// Before
json.NewEncoder(w).Encode(data)
// After
jsoniter.ConfigFast.Marshal(data)
---
## Why This Exists
Profiling tools show *where* time is spent. Developers need to know *why* and *how to fix it*. LLMs are surprisingly good at bridging that gap when given structured profile data.
## Try It
1. Grab a profile from your production service (or generate one locally)
2. Go to **https://aiprofiler.samma.app/**
3. Drop the file, pick a provider, add your API key
3. Read the fixes
No account, no tracking, no upsell. Just paste a profile, get fixes.
---
**Feedback wanted:** Does the analysis quality hold up for your stack? What profiler formats are missing? What would make you use this regularly?
*Built by a developer who got tired of guessing.*
Top comments (0)