IdeaWell

Project ideas from Hacker News discussions.

Decisions API is in public beta

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Four dominant themes in the discussion

  1. Price & performance advantage of competing models
  2. “$0.10/mm input vs. $0.042/mm input. Both free output.” – jerrygenser
  3. “In the same bench a full Jev run cost USD 0.0192,‑ vs Luna at USD 0.06,‑ … about 3× in favour of Jev.” – Topfi

  4. “Why would I run it myself? It's $0.10 per million tokens. Dirt cheap. (Jev is even cheaper.)” – mediaman

  5. Commoditization and the race to the bottom on price

  6. “The response to Jev should be the nail in the coffin over whether or not the AI business is a commodity market.” – TSiege
  7. “If I were OpenAI or Anthropic I’d be racing to make their products as sticky as possible bc ppl will flock to what’s cheapest otherwise.” – TSiege

  8. “It takes me all of 2 keypresses to switch models. I don't know of a less sticky product.” – tripleee

  9. Enterprise factors: existing contracts, compliance, and switching cost

  10. “If you are a business dealing with anything remotely sensitive then this is not so easy and you are basically forced to do business with a big player.” – OutOfHere
  11. “You've not seen how long it takes to switch an enterprise claude subscription to github copilot or vice versa with all the compliance and shareholders.” – schleck8

  12. “If you work for a company that has a 3 to 6 month onboarding period for new vendors and a lifetime commitment to maintain a whole bunch of vendor management horseshit…” – jcims

  13. Technical capabilities and limitations (multimodal, zero‑shot classification, confidence scores, latency)

  14. “You can give it a CCTV image … and ask it to quickly decide actions such as triggering an automated auditory alert… ” – fennecfoxy
  15. “Decisions can take image inputs, which is a pretty common need.” – OutOfHere
  16. “It’s much faster and cheaper (an order of magnitude). And theoretically will give you better answers statistically as it's calibrated.” – Closi
  17. “The probabilities I am seeing so far do not correspond with figures the business would find very agreeable.” – bob1029 (referring to confidence output)

These themes capture the core concerns: cost competitiveness, market commoditization, enterprise adoption barriers, and the practical trade‑offs of the new decision‑model API.

🚀 Project Ideas

BenchDecide

Summary

  • Provides an automated benchmarking suite to compare decision models (OpenAI Decisions, Jev, Mercury Decide, etc.) on latency, cost, calibration, and accuracy across user-provided datasets.
  • Core value: one-click model selection and switching for production workloads, eliminating the manual eval overhead that slows adoption.

Details

Key Value
Target Audience ML engineers, product teams using decision models for classification, routing, moderation
Core Feature Run side‑by‑side evaluations with configurable metrics, export reports, and integrate via SDK or CLI
Tech Stack Python, FastAPI, Pandas, OpenAPI, Docker, optional HuggingFace evals
Difficulty Medium
Monetization Revenue-ready: SaaS subscription tiered by evaluation runs per month

Notes

  • HN commenters stressed the need for evals: “Anyone using a decision model like this is going to have to spin up their own evals” (simonw) and “How else do you establish your model and prompt combination works?” (ajmurmann).
  • Addresses vendor‑lock‑in concerns by making model switches trivial: “If you work for a company that has a 3 to 6 month onboarding period for new vendors…” (jcims).

ClassiFast

Summary

  • Zero‑shot classification API that invokes decision models, returns calibrated probabilities, and automatically escalates low‑confidence inputs to a larger LLM for verification.
  • Core value: fast, cheap, and confident classification with a built‑in safety net, removing the need for manual prompt engineering.

Details

Key Value
Target Audience Developers building moderation, content tagging, intent detection pipelines
Core Feature API endpoint that takes text/image, runs a decision model, checks a confidence threshold, and falls back to GPT‑4o/mini if uncertain
Tech Stack Node.js/Python, OpenAI API, Typesafe Jev SDK, Redis for caching, Prometheus metrics
Difficulty Medium
Monetization Revenue-ready: Pay‑per‑call with a free tier

Notes

  • Users asked for scenarios where confidence matters: “Any use case where a confidence level is desired and you want that number to actually mean something.” (grosswait)
  • Highlighted speed/price gains: “It is much faster and cheaper (an order of magnitude).” (Closi)
  • Removes prompt‑engineering friction: “You could do this before” … “No, you couldn't, this gives you something new” (weird‑eye‑issue).

MemDecide

Summary

  • Adds persistent per‑user or per‑session memory to decision model calls via a lightweight sidecar that stores context in an encrypted KV store, enabling sticky applications without retraining models.
  • Core value: vendor lock‑in through memory, reducing churn and allowing personalized decision‑making.

Details

Key Value
Target Audience Enterprises wanting to differentiate AI offerings, SaaS platforms using decision models
Core Feature Transparent middleware that prefixes decision model input with relevant memory tokens retrieved from a store and updates memory after each call
Tech Stack Go/Rust sidecar, gRPC, AES‑256 encrypted SQLite or DynamoDB, OpenAPI wrapper
Difficulty High
Monetization Revenue-ready: Enterprise license per seat

Notes

  • HN discussion highlighted stickiness: “If I were OpenAI or Anthropic I’d be racing to make their products as sticky as possible bc ppl will flock to what’s cheapest otherwise.” (jiggawatts) and suggestion of per‑user memory as lock‑in.
  • Directly mirrors a user idea: “My approach would be per‑user (or per‑project) ‘memory’.” (jiggawatts)
  • Memory can be configured for zero‑data‑retention to satisfy enterprise compliance concerns.

DecideFlow

Summary

  • Event‑driven orchestrator that ingests multimedia streams (images, video frames, audio) and uses decision models to trigger automated actions (alerts, routing to LLMs, logging) with safety checks and audit trails.
  • Core value: enables real‑time, low‑latent decision‑making for security, moderation, and industrial IoT without building custom pipelines.

Details

Key Value
Target Audience DevOps, security teams, product builders needing real‑time image/video moderation or trigger systems
Core Feature Configurable pipelines: input → decision model (e.g., “Is this frame showing a safety hazard?”) → action (webhook, LLM call, storage) with retry and dead‑letter queue
Tech Stack Python (FastAPI + Celery), OpenAI Decisions API, FFmpeg for frame extraction, WebSocket UI, Prometheus/Grafana
Difficulty Medium
Monetization Revenue-ready: Usage‑based pricing per processed frame

Notes

  • Users praised multimodal input: “You can give it a CCTV image … and ask it to quickly decide actions such as triggering an automated auditory alert, deferring to a larger model for more detailed analysis” (fennecfoxy).
  • Addresses compliance and safety needs by logging decisions for audit: “Deterministic predictions are sought by those looking to offload their decision responsibility to AI…” (OutOfHere).

Read Later