[HN Gopher] Show HN: I built a serverless data API builder - no ...
___________________________________________________________________
Show HN: I built a serverless data API builder - no storage, low
latency
Hey HN, I'm Bo, cofounder of Fleak.ai. Over the past several
months, our team has been hard at work developing Fleak, a data API
backend builder, and we would love your feedback on what we've
built so far. What Fleak Does: Fleak simplifies the process of
building and deploying API backends. It features a no-code IDE UI
that lets you create workflows by chaining together steps such as
native SQL transformations, calling LLM models, AWS Lambda
functions, and more. With a single click, you can deploy these
workflows to a production endpoint (during test, we are able to
handle 5000 QPS without LLM node). Key Features: Native SQL
Transformation: Work with SQL directly, without the need for a
storage layer. Simplified Setup: No more cluster configuration and
script juggling. Efficient Deployment: Streamlines the deployment
cycle and reduces the need for constant monitoring. The inspiration
for Fleak came from my frustration with maintaining production data
pipelines, especially with the added complexity brought by LLMs.
Fleak aims to simplify life for data practitioners. Known
restrictions: 1. not all SQL syntax is supported, we are adding
daily 2. free version has rate limit due to cost 3. LLM node
latency is not optimized due to cost, but our base rate limit is
higher than out of box models and can be elastic. Our codebase is
meticulously maintained, though it's not open source yet as we're
still considering which parts to release. You can check out Fleak
at www.fleak.ai. We are hosting a small product feedback lunch on
Stanford campus this thursday 8/8 1:30pm-2:30pm. Come join us!
https://lu.ma/0beq21pd We'd love to hear your thoughts: Does this
sound useful to you? What features would you like to see? Any
advice on open-sourcing? Thanks!
Author : pinkfluffymochi
Score : 51 points
Date : 2024-08-06 15:49 UTC (7 hours ago)
(HTM) web link (fleak.ai)
(TXT) w3m dump (fleak.ai)
| bolei wrote:
| What are the best use cases you have built?
| pinkfluffymochi wrote:
| The common use cases we found so far are two: data tagging see
| https://docs.fleak.ai/1.0/tutorials/sentiment-labeling and
| embedding pipeline into vector store, see:
| https://docs.fleak.ai/1.0/tutorials/pinecone-embedding
| junjieshen wrote:
| I think there can be batch processing capabilities, just like
| OpenAI
| catapart wrote:
| I'm just one person, and I'm probably not even the target
| demographic, but if you're taking feedback:
|
| I full-on laughed out loud reading your landing page for the "old
| way"/"new way" comparison.
|
| It reads, to me, like "old way: easy to follow, if cumbersome,
| flowchart", "new way: 4 separate abstract drawings that are
| actually impossible to follow and do not imply any specific
| process is occurring."
|
| Obviously, I'm probably in the minority here. Just some food for
| thought, if you're interested in that kind of thing.
| JOnAgain wrote:
| You're not alone. I don't get the contrast either.
| localfirst wrote:
| no its what i thought too and i felt like this is a tough
| problem especially when you get into enterprise space with a
| dozen heterogeneous stacks and outdated vendor APIs
|
| if it was this easy then we would see a ton of layoffs but you
| can't because all these individual pieces and their
| idiosyncrasies cannot be streamlined and its cheaper to just
| have human interfaces.
|
| So many cases I've seen where "hey this doesn't work the way it
| should because the vendor hired bunch of Indians who are paid
| by the hour didn't bother to spellcheck or use proper patterns
| because why would they when you pay them to bill more hours"
|
| We'll see more and more result of negligent managers throwing
| third world labor at critical systems and its impact on the
| economy down the road. This is something only fixable by other
| humans and not any AI or SaaS
| bolei1007 wrote:
| Thanks for the feedback! It's harsh but very valuable!
| tazu wrote:
| Disappointed I can't use an arbitrary S3 URL. I'd like to use
| Cloudflare R2.
| balls187 wrote:
| > Does this sound useful to you?
|
| Who is your target user?
| bolei wrote:
| data scientists and data engineers who do not want to deal with
| the server management or deploy multithreading jobs when
| creating data ingestion APIs.
| abraxas wrote:
| I'd love something like this but with the option to bring in my
| own model(s). In my case I need to run BLIP2 to generate the
| kinds of embeddings I care about.
|
| Ideally I'd love to be able to run arbitrary Python code in a
| node together with custom pip scripts to install the libraries I
| care about. We do some image processing steps and looks like this
| is not something you support.
|
| The bias here is clearly towards text processing but I think more
| and more companies like this one should start thinking about
| multimodal pipelines.
|
| One last point, I did not get far enough in my tests to see if I
| can publish a public API point secured by an API key. That's
| absolutely a must as having to mess around with a gateway myself
| to access this would nullify most of the benefits of this
| platform.
| bolei1007 wrote:
| Thanks for the feedback! Currently we support connecting to AWS
| lambda functions. If you have one deployed we can definitely
| call it within the workflow
|
| Also, the published APIs are publicly available, no need to
| configure gateway for sure
| spencerchubb wrote:
| What is an AI workflow? What does "embed API endpoint" mean? What
| is a concrete example of something I would use it for? Why would
| I use it as opposed to other solutions?
| dsmurrell wrote:
| I'm having similar issues understanding how I might use this.
| bolei wrote:
| you may think of it as a cloud function, we mostly focus on
| HTTP based data ingestions without a storage layers. e.g.
| Your app is sending some data, Fleak can be used to create an
| API to process that data then return back to the app
| synchronously or save the processed data directly into a
| storage layer.
| samtp wrote:
| For someone who is not deep into AI lingo, this pretty much
| sounds like an ORM + Cloudflare Worker/AWS Lambda. But I'm
| probably really off here.
|
| What would really help would be a concrete business case
| for when someone would use this and how it helps vs other
| options
| ricardobeat wrote:
| The "Docs" link goes back to the homepage. I found it hard to
| tell what exactly this does - how can serverless reduce overhead
| (usually it's the other way around), how does it scale the LLM
| backend and so on.
| bolei wrote:
| sorry about that. Please try docs.fleak.ai directly? What Fleak
| offers is a simple way to capture the data processing logic,
| with or without LLM, and we make deployment to production with
| one click. That means as a user, no need to write multithread
| scripts, configuring clusters or create load balancers since
| Fleak's ingestion engine will optimize the latency across
| different processing unit (we call them nodes, whether it's llm
| or simple SQL). On the LLM side, we work with our partner to
| coordinate the throughput, so if there is a demand for 5000 bps
| Llm inferencing, we can certainly scale up to that. But for SQL
| functions, the users can basically get that level of
| autoscaling out of box.
| simple10 wrote:
| This looks useful for no-code team members to create embeddings
| for LLM.
|
| Since you're looking for feedback...
|
| The "New Way" section would benefit from simplifying into a
| single flow chart with just two steps. Then move the existing 4
| blocks to a features section.
|
| Also would be great to have simple use cases on the landing page
| and link to the current Use Cases page for details. On the Use
| Cases page, I would get rid of the graphics. They don't convey
| much info and distract from the screenshots.
|
| Maybe add a page for product comparisons. Without playing with
| the product first, it's a bit unclear to me how my no-code teams
| would prefer Fleak over their current tools (Make.com, n8n,
| Pabbly, Zapier, etc.).
|
| The pricing page is somewhat of a deal breaker for my team to
| even spend time testing the product. The only option (free tier)
| seems to have really limits like 1 request / sec. And there's not
| enough info to understand how max events and other rate limits
| play together. I would keep the 1 user, 5 pipeline, token limits,
| and 500 requests (pipeline executions?) per month. Then remove
| the other limits. Alternatively, offer a Pro tier with a free
| month trial with higher limits.
|
| Also on the pricing page, it would be great to list out a few use
| cases as examples of what's possible within the limits of each
| tier.
|
| Hope this helps! And best of luck. It looks like a promising
| product.
| bolei1007 wrote:
| Thank you for the feedback! They are truly insightful!
| nextworddev wrote:
| Kind of useful but I'd generally point people to AWS
| StepFunctions which is a strict superset of this product
| (assuming they are ok with using AWS)
| bolei wrote:
| that's really great observation. what kind of tasks are you
| building with StepFunctions? We are integration partner with
| AWS Lambda, so definitely curious about StepFunctions too!
| marban wrote:
| Isn't Langflow the better tool for LLMs?
| bolei wrote:
| Langflow is a great orchestration product! Our goal is to
| enable low latency data processing with and without LLMs
| through one click deployment into production ready APIs.
| gerritjvv wrote:
| I suggest you try Fleak, Langflow, and LangChain to compare.
| Neither Langflow nor LangChain are simple options, and they are
| more geared towards chatbot use cases.
___________________________________________________________________
(page generated 2024-08-06 23:00 UTC)