Ideas

Round 1 §

• Using Zero Shot Object Detection
  • They are generally slow and non-specific
• English to no-sql converters
  • Breaking into chunks
  • Cypher queries/ mongo
• GANs for LLMs
  • Solving the Hallucination problem
  • Have a generator and a discriminator
  • [here]
• Bench marking LLMs
  • https://scholar.google.com/citations?view_op=view_citation&hl=en&user=Bs9mrwwAAAAJ&citation_for_view=Bs9mrwwAAAAJ:Wp0gIr-vW9MC
• LLM Compression (quantisation)
  • LLM - KICK over SoTA
• Generating TSA data using GANs that have good temporal dynamics
  • https://towardsdatascience.com/synthetic-time-series-data-a-gan-approach-869a984f2239
• QRs - https://www.qrcode-tiger.com/the-best-ai-generated-qr-code-art-a-must-see
• Meet summerizer
• Auto Bill approval
• Smart alert slack bot - potential misunderstandings, forecast problems, suggest interesting things
• Campaign schedular
  1. Given a campaign description (target audience, mediums, ) it should give you a timeline of doing things
  2. But how will we check accuracy? What is objectively considered correct?

---

Round 2 §

1. Personal Assistant
2. LLM for law
3. LLM for generating promQL (over prometheus) / English to NOSQL converter
4. Smart alert slack bot
5. LLM for monitoring and optimally configuring a system (Cassandra.yaml -> config)

---

Round 3 §

1. Personal Assistant
   • Too many tools out there
   • What will be the USP?
   • Will anyone actually use it? We already have Google Assistant or Siri?
2. LLM for law
   • Too many tools out there for E-discovery and Legal research - Legal NER and IITK
3. LLM for generating promQL
4. Smart alert slack bot
   • Obtaining data is a challenge
   • Objectives aren’t clear - it’s an agent
5. LLM for monitoring and optimally configuring a system
   • Where will we get the data?

Criterion	promQL	optimal config
Input	English sentences	Config files and logs
Output	promQL query	Config files
Data availability	?	?
Training	Easy (match against the correct query)	Hard? (Do we have to know the optimal config? Can we train the model by rewarding it based on the direction of change of parameters of interest?)

---

LLM for generating promQL