Member of Technical Staff · Lyzr AI

Sheet ii · Method

How I tend to build, across these files.

  1. i.Draw it first

    I sketch the whole system before writing code.

    → every file
  2. ii.Keep the hot path small

    The request path does only what it must; the rest goes async.

    → № 04
  3. iii.Interfaces between parts

    Stages talk through interfaces, so parts swap without rewrites.

    → № 03
  4. iv.Check with evals

    Answers get scored, so changes are judged by results.

    → № 02

File № 01 · Engineering practice

Cortex: practice for design decisions.

Problem
Architecture judgment is learned on real systems, where mistakes are costly.
Constraint
Score the reasoning, not the final diagram.
Outcome
Tasks and drills by role, scored answers, and progress over time.
taskyour answerAI scoringprogressfeedback · 5 min4 parts5 dimensions
Drawn by
Rahul Rawat
Stack
Next.js · FastAPI
Year
2026
Status
Live ↗
Sheet
1 / 4

File № 02 · Retrieval

Cognesys: RAG that checks its own answers.

Problem
RAG quality drifts, and nobody notices until users do.
Constraint
Checking must not slow down the answer.
Outcome
Every answer gets scored; config changes are A/B tested.
queryhybrid searchrerankanswerevaluatorsoptimizerA/B testbackgroundnew configthe loop runs after the user has the answer
Drawn by
Rahul Rawat
Stack
FastAPI · Qdrant
Year
2026
Status
Live ↗
Sheet
2 / 4

File № 03 · Document intelligence

Cognita: turning messy documents into clean data.

Problem
PDFs, Office files and scans, made usable for search and LLMs.
Constraint
Plain Go. No PDF library, no ML models.
Outcome
One API to parse, chunk, export and store. Same input, same output.
fileslayoutOCR, scans onlyIRchunks · mdevery arrow is an interface
Drawn by
Rahul Rawat
Stack
Go · OCR
Year
2026
Status
Live ↗
Sheet
3 / 4

File № 04 · Links & analytics

Launchpad: short links that know what they earned.

Problem
Knowing which click led to a signup or a sale.
Constraint
A redirect shouldn't wait on the database.
Outcome
Click → lead → sale attribution, sent on to Meta and TikTok.
clickresolverredirectRedis streamPostgresinsightshot path · cache firstbatched, off the hot path
Drawn by
Rahul Rawat
Stack
Go · Redis · Postgres
Year
2026
Status
Live ↗
Sheet
4 / 4

Appendix A · Experience

Experience

Lyzr AI

Full-time · Bengaluru, India

Member of Technical StaffJun 2026 — Present
Building production-grade AI systems end to end with customers — from backend infrastructure to the interfaces on top.
Backend Engineer InternDec 2025 — May 2026
Architected production-ready AI infrastructure and multi-agent orchestration systems.

Appendix B · Skills

Skills

A pragmatic toolkit with a bias toward backend, data and applied AI.

Backend
Node.js · FastAPI · Python · Go
AI & Agents
LLMs · LangChain · RAG · Vector search
Data
PostgreSQL · MongoDB · Redis · ClickHouse
Infra
AWS · GCP · Docker · Kubernetes
Frontend
Next.js · React · TypeScript · Tailwind

Appendix D · Reading

Reading

Papers, posts and write-ups I've been reading lately.

Loading…

Appendix E · Contact

Contact

Let's build something worth shipping.

A backend that needs to scale, an AI system to design, or just a good conversation about engineering.