Rajeev Bakshi

Founder @CurvSort · LexiLensAI (Agent Runtime Intelligence & Observability) & LexemeAI (Multimodal Instructional Intelligence)

[Rajeev Bakshi @CurvSort]

AI agents introduce new runtime failure modes: recursive loops, token explosions, retry storms, and opaque execution paths.

Exploring these challenges with LexiLensAI — runtime intelligence, observability, anomaly detection, and cost attribution for enterprise AI agents.

https://t.co/D7s2WdySOs

[Rajeev Bakshi @CurvSort]

Classroom video analysis is now live in LexemeAI.
Upload a recorded session → get structured classroom observation output.
No generic summaries.
Clear teaching patterns, engagement signals, session structure.
Built for schools, coaching institutes & academic teams.
App Store (macOS / iOS): https://t.co/nVZdK628a3
Website: https://t.co/6Eu1VXZaHa
Feedback welcome.

[Rajeev Bakshi @CurvSort]

LexemeAI v1.1 is live.

A shift from a general AI tool to domain-structured intelligence — built for dense documents, audio, video, and high-context workflows.

macOS & iOS.

Video-based session analysis coming next.

https://t.co/mFON68xaxX

[Rajeev Bakshi @CurvSort]

For those curious about the mechanics:

Supervised fine-tuning and reinforcement learning mostly adapt model behaviour - they don’t encode new core knowledge.

Continued pre-training can add knowledge, but without original training data it risks catastrophic forgetting - losing basic skills like instruction following and planning.

[Rajeev Bakshi @CurvSort]

Fine-tuning AI models helps them behave better - it doesn’t really teach them new knowledge.

Re-training them on new data can help, but if you’re not careful, they forget what they were originally good at.

That’s why large companies build AI models from scratch.

The catch? It’s extremely expensive and slow.

The smarter path: start from strong pre-trained models and inject domain knowledge at the right stage, without breaking what already works.

[Rajeev Bakshi @CurvSort]

Thrilled to share LexemeAI — a macOS & iOS app that helps professionals extract clarity from dense audio & documents using AI.
Feedback & feature requests welcome!
Discord: https://t.co/4F4Yb6948x
Demo: https://t.co/TngoLhAWFT
App: https://t.co/vvQ5RSBIx8

[Rajeev Bakshi @CurvSort]

Why your AI product team needs to master 'Experimentation Science'. From my experience building multi-lingual platforms at CurvSort, using methods like CUPED and Sequential Testing is critical for managing real-world noise.
Check out https://t.co/EVMa4P9idM.

[Rajeev Bakshi @CurvSort]

@JohnCena This is interesting

[Rajeev Bakshi @CurvSort]

@awscloud Awesome, can be helpful in aligning with architectural constraints and also an eye on the cost.

[Rajeev Bakshi @CurvSort]

@marktenenholtz @AnthropicAI Worth a read

[Rajeev Bakshi @CurvSort]

Today was one of productive days. Avoiding interruptions helps.

[Rajeev Bakshi @CurvSort]

@fjmellq Looks nice

[Rajeev Bakshi @CurvSort]

@FitFounder Right

[Rajeev Bakshi @CurvSort]

@bryan_johnson Right. It can be standing in a hurricane. Emotions and feelings are oceans of nature in a very real senses.

[Rajeev Bakshi @CurvSort]

Jokes are a fantastic way to recharge your mind’s energy. Playfully tease a buddy, engage in a sport when possible, and keep strengthening your journey with fun and laughter!

[Rajeev Bakshi @CurvSort]

@elonmusk Brexit reflected a manifestation of pessimism and fear about Europe’s future.