Backend & Search Engineer · Gurugram, IN

Abhinav Tyagi.

I teach machines what people mean — not just what they type. Five years of search platforms and ranking models that stay calm at 25K requests a minute.

Let’s talk →Resume ↓

01 — Live · query understanding

I’m not a list of frameworks. I’m the person who needs to know why the third result outranked the first.

Search sits where language meets systems, and that’s exactly where I like to live. Query understanding, ranking models, ingestion pipelines that never sleep — the result is platforms serving 25K+ requests a minute across a 10-million-product catalog that still feel personal.

Search something. Watch it think.

Type anything and it parses your words into entities, recalls candidates with BM25, then re-ranks them — every row tells you why it landed where it did.

query understandingidle

parsewords → entities
recallBM25 candidates
rescoresignals decide

Start typing to search 162 products.

// real BM25 + rule-based NER, in your browser · CTR/ATC are synthetic priors, not telemetry

Lexical match gets you candidates. Synonyms, attribute fit and demand signals decide who actually deserves the top slot.

25K+

Search requests / min

~10M

Products indexed, near-realtime

−45%

Search latency, Redis-backed

Years shipping to production

02 — Experience

Four roles · one rail

03 — Skills

Three things I know cold. The rest I know well.

Depth beats breadth in search work — most of my day happens inside the first three. Everything below them is a tool I have shipped, not a tool I have read about.

Relevance

Elasticsearch

Relevance tuning, the DSL by hand, and two 6.x → 8.x migrations that stayed online.

10MDocs indexed
Near-realtime

Throughput

Java · Spring

Search services that hold their nerve under load, with the hot path kept in Redis.

25KRequests / min
−45% latency

Ranking

LTR · XGBoost

Rescore pipelines, feature logging, and the judgement work that decides what better means.

3 yrsIn production
Rescoring live

Also in the toolbox

Shipped, not skimmed

Search & data

Kafkaingestion pipelines
Redishot-path caching
MongoDBaggregation pipelines

Language

NERsizes, units, brands
Vector searchdense retrieval, ANN
Embeddingstwo-tower, tagging

Backend

Pythontooling & ML pipelines
Microservicesevent-driven
LLM toolingMCP, agents

Infra

Kubernetesorchestration, rollouts
Dockercontainers & CI
AWS · GCP · Azuremulti-cloud deploys
OpenTelemetrytracing

Anything on this page, I can walk you through in the code.

Full stack on the CV ↓

04 — Lab

Where I build things that don’t have a brief yet

Right now that means one thing: giving Elasticsearch a conversation layer. An agent that holds a real connection to my cluster, understands what I’m asking in English, and does the indexing work itself.

In progressFlagship

Natural query

An Elasticsearch agent that speaks plain English. Ask it a question and it plans the query, runs it against my live cluster and answers. Describe a schema and it writes the mapping and creates the index — no console, no JSON by hand.

How it connects

Plain Englishprompt

Agentplan · tools

MCP connectorelastic · custom

Elastic clusterlive index

Writes mappings

Creates indices and field types straight on the cluster.

Answers in English

Reads back results instead of raw hit arrays.

Tool-scoped

Every cluster action is an explicit MCP tool call.

Java 17Spring Boot 3.xMCP Java SDKElasticsearch 8.x