I care about two things that sound different but feel the same to me:

1. what a product is actually doing when a user clicks something
2. what a system is actually doing in the two seconds before the answer shows up

Most tutorials show you how to call an API. I want to know why the pipeline failed, which layer lied, and whether the feature was the right thing to ship.

I am an Analyst at MiQ, working across the MENA markets. Previously, at Flipkart, I worked on seller funnel analytics and search personalization — behavioral data into product decisions. Outside work I build and test intelligent systems: retrieval, memory, forecasts and source-aware agents.

MiQ · Analyst · MENA markets
Explore my Systris portfolio →

AI systems · retrieval · inference
memory · evaluation · product experiments

if it isn't merged, it isn't here.

magpie · every agent's model, one place - Codex on DeepSeek, Claude Code on Kimi, from the menu bar

I fixed the Codex prompt cache so two accounts saving in the same mtime tick both keep their instructions. merged in yetone/magpie#74 → · Sep 2026

I stopped disabled keys from being fetched during model refresh, with a regression test proving zero requests to them. merged in yetone/magpie#76 → · Sep 2026

I fixed the TestSmartRouting hour-boundary flake by anchoring its reset fixture within one hour, so the five-hour tie-breaker stays deterministic. merged in yetone/magpie#77 → · Sep 2026

01 Where does the latency actually go?
02 What should an AI system remember - and prove it remembered from the source?
03 When is a forecast wrong because of the model, and when because of the place?
04 Why do technically good products still fail distribution?
05 Which problems are worth automating, and which only look that way?

IntelliRAG
I wanted to know where RAG actually breaks.

ingestion → chunking → hybrid retrieval → rerank → citations → evaluation → observability
The live browser lab exposes issue ingestion, cited retrieval traces and a lexical knowledge graph. The public demo currently uses temporary keyword/extractive mode; persistent hybrid retrieval and real-provider evaluation are next. The Python platform has separate deterministic CI benchmarks.

try the live lab → · read the measured audit →

memoRABLE
What if documents became memory?

Six source-linked blocks. Click a memory, the original lines light up. Publish once to email / web / doc without rewriting the truth. Local-first.

try it →

ThermoSense · hyperlocal temperature forecasting
Can a forecast know your rooftop?

Ground truth → commercial API bias → ensemble forecast → public leaderboard → retrain. The product is the loop, not the model name.

see the experiment →

Finsight
Can an AI answer also explain how much it should be trusted?

Multi-agent research with freshness, source agreement, versioned knowledge, and a confidence score you can inspect.

inspect the system →

infer-tab
What changes between prefill and decode?

CPU-only KV-cache experiments write traces of attention arithmetic and tensor bytes; a Next.js visualizer replays them. No model download needed.

run the experiments →

Search - what actually happens between a query and the ranked result (Flipkart search personalization was the first place this got real for me)
Funnels - where discovery leaks: the step users drop, not the dashboard average
Trust UX - when a product should show confidence, provenance, or “I don’t know yet”
Distribution - why a technically solid system still fails to get used
The invisible middle - the 2 seconds you never see: auth, memory, routing, the work that makes complexity feel effortless

I go system → product → business. Same habit: open the black box, name the failure mode, then decide what to ship.

The 2 seconds you never see
I thought I knew what happened after you hit enter. I was wrong.

More when I have something worth saying → Substack

measure → build → break → learn → repeat

Write the tradeoff down. Keep eval next to the code. Prefer systems that fail in known ways. Same rule for product: if I can’t explain the funnel step, I don’t trust the feature yet.

→ what actually determines RAG latency (retrieval vs rerank vs generation vs cold start)
→ how memory systems should preserve provenance without becoming another summary blob
→ when local inference is the right constraint vs when it just feels pure
→ closed-loop evaluation: leaderboards that force the model to face ground truth
→ why technically good products fail distribution

I update this when the questions change.

GitHub · Substack · Portfolio · Email

_{BITS Pilani · dual degree · class of 2026 · MiQ analyst, MENA · ex-Flipkart product analytics
Older experiments stay public. The four above are the ones that still feel like me.}