3 ms·

[stub]

by tomhow 1y ago

[stub]

Abir-01 1y ago
[dead]
- nityasha 1y ago
  Thank you so much for using Nityasha, Abir! Really appreciate you being an early user. We spent a lot of time on the UI trying to make it feel natural and not overwhelming. My dad kept saying "if a 13-year-old can't understand it instantly, we failed" Out of curiosity - what features do you use most? And what would you like to see added? Also, when you say "better than some new gen slms" - are there specific areas where you find it more helpful? Always trying to understand what's working so we can double down on it. Thanks again for the support!
Ayish_Shaikh 1y ago
[flagged]
- nityasha 1y ago
  Thanks Ayish! Glad it's working well for you. Quick clarification though - Nityasha isn't actually an SLM (Small Language Model) itself. We use existing large models (like GPT-4, Claude) but add a persistent memory layer on top + specialized features like Study Mode. The "showing results" part is what we're most proud of - we tried to focus on actually being useful for daily tasks rather than just being impressive in demos. What kind of tasks are you using it for? Would love to hear more about your workflow!
asknraaj 1y ago
[dead]
nityasha 1y ago
[dead]
40yoccoder 1y ago
[dead]
mdrzn 1y ago
So a wrapper around an existing LLM? even chatgpt can reference previous conversations and remember context. Also the fake comments (or friends comments?) are not really liked on Show HNs. Congrats on building it.
- nityasha 1y ago
  Fair point on the "wrapper" label. Let me clarify what we're building on top of base LLMs: Yes, we use OpenAI/Anthropic APIs - we're not training models from scratch (like you said, neither does Perplexity, Jasper, or most AI tools). What we add (technical details): 1. Persistent Memory Architecture - Vector embeddings of user context stored in Pinecone - Semantic search across past conversations (not just in-session) - Retrieval pipeline: query → embed → cosine similarity → top-k memories → inject in prompt - Challenge: Managing token costs while maintaining context 2. Socratic Teaching System (Study Mode) - Question analysis: detect knowledge gaps - Progressive hint generation (not just Q&A) - Tracks learning progression - Example: Instead of "here's binary search code", asks "what property of sorted arrays makes this possible?" 3. Unified Workflow Integration - Email parsing + calendar sync + task extraction - Single interface reduces context switching - Memory persists across all tools Architecture overhead: - User sends query - Retrieve relevant memories (vector search) - Build augmented context window - Send to LLM with enriched prompt - Generate + store new embeddings - ~200ms additional latency for memory operations You're right that the base intelligence is GPT-4/Claude. But saying "wrapper" feels like saying Notion is "just a wrapper around PostgreSQL" or Stripe is "just a wrapper around payment processors." The value is in the layer we built, not the underlying model. That said - we should've been clearer about this upfront. Our first comment didn't explain the technical depth. That's on us. Re: the fake comments - you're absolutely right to call that out. Those were friends/early users we asked to support the launch. That was a mistake and goes against HN's culture of authentic discussion. I'm 13 and this is our first HN launch. We didn't understand how much the community values genuine engagement over orchestrated support. Won't happen again. Apologies to the HN community for trying to game the system. Should've let the product speak for itself. Appreciate you taking time to give honest feedback instead of just downvoting. This is exactly why we launched here - to learn from people who know better. What would make this feel genuinely useful vs "just another wrapper" to you?

[stub]

by tomhow 1y ago

[stub]

Abir-01 1y ago

[dead]

nityasha 1y ago
Thank you so much for using Nityasha, Abir! Really appreciate you being an early user. We spent a lot of time on the UI trying to make it feel natural and not overwhelming. My dad kept saying "if a 13-year-old can't understand it instantly, we failed" Out of curiosity - what features do you use most? And what would you like to see added? Also, when you say "better than some new gen slms" - are there specific areas where you find it more helpful? Always trying to understand what's working so we can double down on it. Thanks again for the support!

Ayish_Shaikh 1y ago

[flagged]

nityasha 1y ago
Thanks Ayish! Glad it's working well for you. Quick clarification though - Nityasha isn't actually an SLM (Small Language Model) itself. We use existing large models (like GPT-4, Claude) but add a persistent memory layer on top + specialized features like Study Mode. The "showing results" part is what we're most proud of - we tried to focus on actually being useful for daily tasks rather than just being impressive in demos. What kind of tasks are you using it for? Would love to hear more about your workflow!

asknraaj 1y ago

[dead]

nityasha 1y ago

[dead]

40yoccoder 1y ago

[dead]

mdrzn 1y ago

So a wrapper around an existing LLM? even chatgpt can reference previous conversations and remember context. Also the fake comments (or friends comments?) are not really liked on Show HNs. Congrats on building it.

nityasha 1y ago
Fair point on the "wrapper" label. Let me clarify what we're building on top of base LLMs: Yes, we use OpenAI/Anthropic APIs - we're not training models from scratch (like you said, neither does Perplexity, Jasper, or most AI tools). What we add (technical details): 1. Persistent Memory Architecture - Vector embeddings of user context stored in Pinecone - Semantic search across past conversations (not just in-session) - Retrieval pipeline: query → embed → cosine similarity → top-k memories → inject in prompt - Challenge: Managing token costs while maintaining context 2. Socratic Teaching System (Study Mode) - Question analysis: detect knowledge gaps - Progressive hint generation (not just Q&A) - Tracks learning progression - Example: Instead of "here's binary search code", asks "what property of sorted arrays makes this possible?" 3. Unified Workflow Integration - Email parsing + calendar sync + task extraction - Single interface reduces context switching - Memory persists across all tools Architecture overhead: - User sends query - Retrieve relevant memories (vector search) - Build augmented context window - Send to LLM with enriched prompt - Generate + store new embeddings - ~200ms additional latency for memory operations You're right that the base intelligence is GPT-4/Claude. But saying "wrapper" feels like saying Notion is "just a wrapper around PostgreSQL" or Stripe is "just a wrapper around payment processors." The value is in the layer we built, not the underlying model. That said - we should've been clearer about this upfront. Our first comment didn't explain the technical depth. That's on us. Re: the fake comments - you're absolutely right to call that out. Those were friends/early users we asked to support the launch. That was a mistake and goes against HN's culture of authentic discussion. I'm 13 and this is our first HN launch. We didn't understand how much the community values genuine engagement over orchestrated support. Won't happen again. Apologies to the HN community for trying to game the system. Should've let the product speak for itself. Appreciate you taking time to give honest feedback instead of just downvoting. This is exactly why we launched here - to learn from people who know better. What would make this feel genuinely useful vs "just another wrapper" to you?