Show HN: Lumina – a minimal AI reflection app (source code) https://bit.ly/3MN7APK

Show HN: Lumina – a minimal AI reflection app (source code) https://bit.ly/4j8wwNL December 23, 2025 at 01:43AM

Show HN: A repo to turn any model into a reasoning model without training https://bit.ly/4pUjAxs

Show HN: A repo to turn any model into a reasoning model without training Hey all, Training AI Models to reason is currently very expensive. You require a lot of data, tons of compute in Reinforcement Learning, and more. And the reasoning infrastructure is not reusable. On top of all this, we don't really have a way to improve targeted performance and personalize intelligence within the systems. Over the last year, I've been looking into latent space reasoning as a way to solve this. By doing this, we can create a different reasoning layer that is reusable across models. We created a simple layer for 50 cents, and it already improves performance. We're working with a few people across major AI Labs at exploring this, but I also wanted to open source because intelligence deserves to be open. To that end, our startup has even opened up a small monthly prize pool for top contributors to the repo. Would love to have you in there. Here is a report we did breaking down the core design philosophy here-- https://bit.ly/4jehbLw... https://bit.ly/4pKA5vS December 23, 2025 at 12:09AM

Show HN: I automated forensic accounting for divorce cases (3 min vs. 4 weeks) https://bit.ly/4j5GTSi

Show HN: I automated forensic accounting for divorce cases (3 min vs. 4 weeks) Burned about 1 weeks on this. Not sure if it's useful to anyone beyond my original use case, but figure I'd share. Friend went through a nasty divorce. Had $750k going into the marriage (inheritance), put it in a joint account like an idiot. Five years later, account's been up and down, money mixed with paychecks and mortgage payments. Lawyer says "you need a forensic accountant to trace what's still yours." Quote comes back: $5k, 4 weeks minimum. I'm sitting there thinking - this is just transaction categorization and some relatively simple math (the "Lowest Intermediate Balance Rule" if you want to google it). Why doesn't software exist for this? Turns out it kind of doesn't. There are $50k enterprise tools for big law firms, but nothing a normal person or small practice can actually use. So I built a Django app that takes bank statement PDFs, uses latest Mistral's OCR-3 to parse them (because real-world bank PDFs and shots are a nightmare - scanned, rotated, weird formats), then runs them through an LLM to categorize transactions and a Python implementation of the LIBR algorithm. Output is a court-usable report showing exactly how much of your "separate property" is still traceable, with visualizations and evidence logging (SHA-256 hashing for chain of custody, audit trails, the works). Its FREE and whole process takes about 3 minutes. I'm in India and honestly just want to see if people use it. What's really interesting: -Latest Mistral's document OCR-3 is genuinely impressive on messy banking PDFs. Tried Tesseract first, got maybe 60% accuracy. -The LIBR algorithm is conceptually simple but has some gnarly edge cases (what happens when account hits zero? how do you handle multiple deposits of separate property? etc.) -Evidence integrity was harder than expected. Lawyers care a LOT about proving a document hasn't been tampered with. -Used Celery because some statements have 10k+ transactions and you can't block the request Currently running on Render with Postgres. Code's not open source yet because honestly it's kind of a mess and I need to clean up some stuff, but might do that if there's interest. Things I'm unsure about: -Should it be free? Subscription? How much? Bring your won key? Cause I'm putting money out of my pocket. -B2C vs B2B - individuals might use this once, but lawyers could use it repeatedly. -How much do I need to worry about legal liability for the output? I have disclaimers everywhere but still Anyway, it's live: https://bit.ly/4j5GU8O . Would love feedback, especially if you've dealt with this problem before or know the family law space. December 22, 2025 at 01:22AM

Show HN: Real-time SF 911 dispatch feed (open source) https://bit.ly/4aywE6Z

Show HN: Real-time SF 911 dispatch feed (open source) I built an open-source alternative to Citizen App's paid 911 feed for San Francisco. It streams live dispatch data from SF's official open data portal, uses an LLM to translate police codes into readable summaries, and auto-redacts sensitive locations (shelters, hospitals, etc.). Built it at a hack night after getting annoyed that Citizen is the only real-time option and they paywall it. Repo: https://bit.ly/49dQK4a Discord: https://bit.ly/49dQKkG Happy to discuss the technical approach or take feedback. https://bit.ly/490k3a8 December 22, 2025 at 01:59AM

Show HN: Mactop v2.0.0 https://bit.ly/3MMaRi9

Show HN: Mactop v2.0.0 https://bit.ly/3Y3h4c1 December 22, 2025 at 01:44AM

Show HN: Pac-Man with Guns https://bit.ly/4ays0pz

Show HN: Pac-Man with Guns Title really says it all on this https://bit.ly/493v2j1 December 22, 2025 at 12:17AM

Show HN: ZXC – Asymmetric, +40% decode vs. LZ4 on ARM (C, BSD-3, Fuzzed) https://bit.ly/4s4fy7j

Show HN: ZXC – Asymmetric, +40% decode vs. LZ4 on ARM (C, BSD-3, Fuzzed) https://bit.ly/4p8yhvM December 17, 2025 at 02:18PM

Show HN: Сulsans – Thread-safe async-aware queue for Python https://bit.ly/4p6E1WH

Show HN: Сulsans – Thread-safe async-aware queue for Python In my previous post [0], I described how I came to create aiologic. Here, I want to do the same for a derivative library - Сulsans. In October 2024, I started thinking about how to present aiologic queues. Andrew Svetlov's Janus library [1] had been around for quite some time and was much more popular, so I knew that comparisons with it would be inevitable. However, Janus seemed to be in a suspended state: there had been no major changes for three years, and almost all commits during that period were made by Dependabot. So I asked a relevant question [2]. During the discussion, I pointed out Janus' performance issues and stated that they could be solved by implementing queues on top of my primitives. But since Janus is a mature library, such a radical change could not be accepted. Therefore, as proof of concept, I implemented a new library - Culsans. That is how its story began. Over time, both libraries underwent changes. Janus received significant performance improvements in 1.2.0, not least due to my PRs [3]. In 2.0.0, contrary to the above, backward compatibility was broken as a result of the implementation of shutdown methods. And Culsans became an independent library with its own features (which neither aiologic nor Janus have). So, what is Culsans? It is a library that provides a way to communicate within a single process between different threads, different tasks (including from different event loops; asyncio, Curio, Trio, AnyIO - whatever you want), and even different greenlets (eventlet/gevent), all in a single instance. Its queues are fully compatible with the standard queues via Janus-like interfaces (as well as with Janus itself) and provide additional features such as dynamic maxsize. In short, I invite you to try out my library and see for yourself. [0] https://bit.ly/4qk6k4S [1] https://bit.ly/4p53pvQ [2] https://bit.ly/44AvCDB [3] https://bit.ly/4p1Pirf https://bit.ly/4pWtsa8 December 20, 2025 at 09:25PM

Show HN: Jmail – Google Suite for Epstein files https://bit.ly/3YFewAW

Show HN: Jmail – Google Suite for Epstein files Hi everyone! My name's Luke and I made the original Jmail here alongside Riley Walz. We had a ton of friends collaborate on building out more of the app suite last night in lieue of DOJ's "Epstein files" release. Please AMA! https://bit.ly/48R9lnu December 20, 2025 at 10:00PM

Show HN: I made an AI agent to interact with resume and make changes as you ask https://bit.ly/4qjzfpF

Show HN: I made an AI agent to interact with resume and make changes as you ask I built an agent that understands your resume content and make changes as you prompt with the real time resume preview. https://bit.ly/3MSrbOk December 20, 2025 at 08:29AM

Show HN: I Built an Image Captioning Tool Using Llama.cpp https://bit.ly/4j6tyJB

Show HN: I Built an Image Captioning Tool Using Llama.cpp Frustrated with the apparent lack of tools for tagging and describing images locally, I built a quick and dirty little tool. You start it up, start up llama-server, and point it at a directory of photos. It scans through them, captioning them one at a time, and provides the captions and tags in an editable interface for you. When you're happy with them, you can save them, which writes them to the exif metadata of the image, and moves onto the next one https://bit.ly/4b6dS6S December 19, 2025 at 10:49PM

Show HN: Misata – synthetic data engine using LLM and Vectorized NumPy https://bit.ly/4j6VBIN

Show HN: Misata – synthetic data engine using LLM and Vectorized NumPy Hey HN, I’m the author. I built Misata because existing tools (Faker, Mimesis) are great for random rows but terrible for relational or temporal integrity. I needed to generate data for a dashboard where "Timesheets" must happen after "Project Start Date," and I wanted to define these rules via natural language. How it works: LLM Layer: Uses Groq/Llama-3.3 to parse a "story" into a JSON schema constraint config. Simulation Layer: Uses Vectorized NumPy (no loops) to generate data. It builds a DAG of tables to ensure parent rows exist before child rows (referential integrity). Performance: Generates ~250k rows/sec on my M1 Air. It’s early alpha. The "Graph Reverse Engineering" (describe a chart -> get data) is experimental but working for simple curves. pip install misata I’d love feedback on the simulator.py architecture—I’m currently keeping data in-memory (Pandas) which hits a ceiling at ~10M rows. Thinking of moving to DuckDB for out-of-core generation next. Thoughts? https://bit.ly/49kC2JI December 16, 2025 at 03:38PM

Show HN: Daily Set Puzzle – I rebuilt it after setgame.com's SSL cert expired https://bit.ly/4pLzg61

Show HN: Daily Set Puzzle – I rebuilt it after setgame.com's SSL cert expired https://bit.ly/4pJ9MpL December 19, 2025 at 03:36AM

Show HN: Bithoven – A high-level, imperative language for Bitcoin Smart Contract https://bit.ly/3Y3CnKy

Show HN: Bithoven – A high-level, imperative language for Bitcoin Smart Contract Hey HN! I’m a researcher working on Bitcoin smart contracts, and today I’m releasing Bithoven—a high-level imperative language that compiles to native Bitcoin Script (Legacy, SegWit, and Taproot). The Goal: Raw Bitcoin Script is notoriously difficult to reason about. Writing raw Bitcoin Script today feels like writing Assembly in the 1970s. You have to mentally juggle the stack (`OP_SWAP`, `OP_ROT`), manually manage distinct execution branches, and pray you didn't leave a stack item unconsumed (which crashes the script). My goal was to bridge the gap between complex contract logic and raw opcodes, allowing developers to write readable, compile-time-safe code. Key Features: - Imperative Syntax: Write logic using familiar if/else and return statements instead of mental stack juggling. - Type Safety: First-class support for bool, signature, string, and number types to prevent runtime errors. - Targeted Compilation: Support for Legacy, SegWit, and Taproot compilation targets. - Native Primitives: Built-in keywords for timelocks (older, after) and cryptography (sha256, checksig). You can try it in the browser here (runs via WASM): https://bit.ly/3Y3O9Ez Here is an example of a Hashed Time-Locked Contract (HTLC): (condition: bool, sig_alice: signature) (condition: bool, preimage: string, sig_bob: signature) { if condition { // Relative locktime (Sequence) older 1000; return checksig (sig_alice, alice_pk); } else { // Hashlock verification verify sha256 sha256 preimage == hash; return checksig (sig_bob, bob_pk); } } The project is free open source and the academic paper is currently under review. I’d love to hear any feedback. Thanks for checking it out! https://bit.ly/490qUAt December 15, 2025 at 01:54PM

Show HN: Learning a Language Using Only Words You Know https://bit.ly/4aVwhTU

Show HN: Learning a Language Using Only Words You Know A proof-of-concept language learning app that uses LLMs to generate definitions of unknown words using only previously mastered vocabulary. https://bit.ly/4jp3x8B December 15, 2025 at 02:32PM

Show HN: Proxylity – serverless UDP services on AWS (one year in production) https://bit.ly/44zlD1m

Show HN: Proxylity – serverless UDP services on AWS (one year in production) https://bit.ly/4p4qtei December 18, 2025 at 01:22AM

Show HN: Largest Public Dataset of Electronic Circuit Files https://bit.ly/4ahaLJ8

Show HN: Largest Public Dataset of Electronic Circuit Files Introducing Open-Schematics: a large public dataset of electronic schematics with rendered images and structured metadata for ML, circuit understanding, retrieval, and validation. https://bit.ly/48UL3HX December 18, 2025 at 03:04AM

Show HN: Prompt-refiner – Lightweight optimization for LLM inputs and RAG https://bit.ly/494weTv

Show HN: Prompt-refiner – Lightweight optimization for LLM inputs and RAG Hi HN, While building RAG agents, I noticed a lot of token budget was wasted on formatting overhead (HTML tags, JSON structure, whitespace). Existing solutions felt too heavy (often requiring torch/transformers), so I wrote this lightweight, zero-dependency library to solve it. It includes strategies for context packing, PII redaction, and tool output compression. Benchmarks show it can save ~15% of tokens with negligible latency overhead (<0.5ms). Happy to answer any questions! https://github.com/JacobHuang91/prompt-refiner December 17, 2025 at 11:40PM

Show HN: C-compiler to compile TCC for live-bootstrap https://bit.ly/4s4zJls

Show HN: C-compiler to compile TCC for live-bootstrap https://bit.ly/3L8SXpg December 18, 2025 at 12:34AM

Show HN: Python packages for FastAPI apps – auth, logging, config, LLM, more https://bit.ly/48K1Y16

Show HN: Python packages for FastAPI apps – auth, logging, config, LLM, more https://bit.ly/3KLxQJD December 17, 2025 at 01:04AM