Writing

A flat store of facts gets corrections, refinements and polite agreement all wrong. Separating when something was true from when the system learned it fixes most of it.

• LLM
• Memory
• Data Modelling

Label blind, and treat contradiction as a gate rather than a term in the score. Recall and precision alone will let a system buy recall with wrong facts.

• Evaluation
• LLM
• Retrieval

Hand-tuned boosts and Reciprocal Rank Fusion are the usual answers for combining retrieval branches. Fitted on held-out data, both lost to plain cosine similarity.

• Retrieval
• Evaluation
• LLM

No amount of tuning top-k or thresholds moved it, because a third of the right answers were never candidates in the first place. How to tell a candidate-generation failure from a ranking failure.

• Retrieval
• LLM
• Evaluation

The dashboards were green because the system was logging its own failure as a user hangup. The cause was one timer, shared where it should have been per-channel.

• WebRTC
• Debugging
• Real-Time

I joined as a senior engineer. In my first month I rewrote the data layer and built a versioned deploy system. Neither was on my JD. Here's why I did them anyway.

• React Native
• TanStack
• DevOps

When a microservice goes down at 3am, the slow part isn't fixing it — it's figuring out which service is actually broken. We built an LLM agent that does the triage.

• Agents
• MCP
• On-Call
• Observability

When users pay per minute for a call, the wallet is on the critical path. Here's how we used Redis-level reservations and locks to prevent double-billing and negative balances.

• Redis
• Distributed Systems
• Payments

An LLM that does astrology readings sounds like a gimmick. Building one that feels like a different astrologer for every user is a real engineering problem.

• LLM
• Personalization
• Product

Two React Native apps, one CometChat SDK, and a long list of edge cases. Here's what we learned shipping live consultations between users and astrologers.

• Real-Time
• React Native
• CometChat

Six weeks ago I was building systems where a 2% error rate meant 2,400 patients got denied care. Today I'm building astrology consultations. Here's what that switch actually feels like.

• Career
• AI
• Consumer Tech

The messy, complex, and rewarding reality of building AI systems that process 120k+ medical procedures daily.

• AI/ML
• Healthcare
• Engineering

How I transitioned from building a freelancing platform to revolutionizing healthcare automation with LLM technologies.

• Career
• Entrepreneurship
• AI

A deep dive into the code and concepts behind our 98% accurate medical data extraction pipeline.

• LLMs
• Evaluation
• Healthcare Tech

Everyone told us not to use serverless for a production startup. We did it anyway. Here is what broke, and how we fixed it.

• AWS
• Serverless
• Architecture