◇ AMIT KUMAR — SYSTEMS & AI ENGINEERING
Remote · Available for consulting

Fullstack & AI systems that scale.

I'm Amit Kumar, a software engineer and IIT Bombay graduate who designs and ships the infrastructure underneath high-stakes products — building AI agent harnesses, guardrails, and the backend systems running behind them.

Start a project → Review case studies

Tech Stack

An AI-powered stack built for production.

FIG. 01 — CAPABILITY MAP

Languages

TypeScriptPythonGolangJavaJavaScriptNode.jsSQLC++

AI & Tooling

Claude APIOpenAI APIMulti-agent orchestrationRAGPrompt engineering & guardrailsMCP serversAI gatewaysn8nClaude CodeCursor

Frameworks & Infra

ReactFastAPIDjangoSpring BootSQLAlchemyAWSDockerKubernetesTerraformPostgreSQLMySQLRedisDynamoDBMongoDBDuckDBMicroservicesCI/CD

Track record

FIG. 02 — DEPLOYMENT HISTORY

Blink Health

Software Engineer (IC4) · Feb 2024 – Mar 2026 · Remote (New York, U.S.)

Handled fulfillment, identity, and the LLM orchestrator behind Blink's AI workflows.

HealthTechLLM orchestrationCDC streamingCI/CD

Stable Money

Backend Developer, Founding Team · May 2023 – Feb 2024 · Bangalore

Shipped the app from 0 to 1 — built the payment gateway and an AI advisor chatbot.

FintechRAGPaymentsBanking APIs

DealShare

Software Development Engineer – 1 · Mar 2022 – May 2023 · Bangalore

Part of customer engagement and retention — built features that increased engagement.

EcommerceGrowthAutomationMicroservices

General Electric

Project Engineer / Engineering Intern · May 2019 – Mar 2022 · Bangalore

Embedded-systems security.

EmbeddedBLESecurity

Indian Institute of Technology Bombay

2016 – 2021 · Mumbai

Bachelor's and Master's in Engineering

Education

Case studies

FIG. 03 — SELECTED ENGAGEMENTS · CLICK A CARD TO EXPAND
CASE 01Agentic tool company

Agent Trace: An Observability Engine for LLM Agents

  • Captures every LLM call, tool use & retry as structured events
  • Natural-language queries answered with no LLM in the path
  • Embedded DuckDB — zero ops overhead
  • Batched SDK ingestion with automatic retry & backoff

A TypeScript SDK batches events — 50 events or 5 seconds, whichever first — with exponential backoff, and posts them to a /capture endpoint that enqueues to an async DuckDB writer. Natural-language questions map to one of 8 named SQL queries against a single wide event table, so answering them never costs a model call.

analytics-engine · capture
Agent TypeScript SDK
↓
/captureBatched · async DuckDB writer
↓
KPI Dashboard
Trace Explorer
NL Query
Illustrative architecture — not to scale
LLM observabilityDuckDBFastAPIReact
CASE 02Disco

Ad Campaign Planner: One Brief In, a Full Campaign Out

  • Two-stage Claude pipeline: brief parsing → campaign planning
  • Publisher scoring across five weighted dimensions
  • Persona-targeted ad copy generated per campaign
  • Provider-agnostic — Claude or OpenAI behind one interface

Claude Haiku extracts structured attributes from the raw brief — category, audience, price tier, channel fit — with JSON retry logic. Claude Sonnet then scores publishers across five weighted dimensions and drafts persona-specific creative, behind an abstract LLMClient interface so Claude or OpenAI can sit behind the same pipeline.

ad-campaign · /campaign/plan
Advertiser Brief
↓
Parser · Claude Haiku → Planner · Claude SonnetProvider-swappable via LLMClient
↓
Publisher Ranking
Persona Creative
Budget Split
Illustrative architecture — not to scale
LLM pipelineClaude APIFastAPIDocker
CASE 03Blink Health

Multi-Agent LLM Orchestration with HIPAA-Grade Guardrails

  • Built the multi-agent LLM orchestrator powering BlinkRx end-to-end
  • Guardrails and an immutable audit log on every automated decision
  • HIPAA-grade schema validation with human escalation thresholds
  • Routes prescriptions across Amazon Pharmacy, Uber Direct & ScriptDrop

Agents act independently on low-risk steps but escalate above defined risk thresholds — every decision constrained to a validated schema and written to an immutable audit log. Keeping agents HIPAA-compliant meant schema-validated outputs and full auditability on every action, plus human escalation for high-risk steps, without collapsing response latency.

orchestrator · blinkrx-prod
Prescription Event
↓
LLM Agent OrchestratorGuardrails · Audit log · Human escalation
↓
Amazon Pharmacy
Uber Direct
ScriptDrop
1M+DAILY PRESCRIPTIONS ROUTED
Illustrative architecture — not to scale
LLM orchestrationGuardrailsPythonHealthcare
CASE 04Blink Health

Replacing Batch Reporting with Near-Real-Time Data Pipelines

  • Replaced nightly batch jobs with near-real-time CDC streaming
  • Ordered, replay-safe event pipeline straight off production Postgres
  • Standardized the dev → staging → prod release pipeline
  • Cut deployment incidents with one consistent promotion path

Built the streaming pipeline off the transactional databases, and separately designed dev → staging → prod promotion pipelines so every change followed the same release path. Streaming changes out of live tables without adding load or risking data drift — ordering guarantees and replay-safety had to hold up under real transaction volume, not a demo.

pipelines · blink-data
Transactional DBPostgreSQL
↓
CDC Event StreamOrdered · replay-safe
↓
Dashboards & Downstream Services
Batch → LiveREPORTING LATENCY
↓DEPLOYMENT INCIDENTS
Illustrative architecture — not to scale
CDCEvent streamingCI/CDPostgreSQL
CASE 05Stable Money

Stable Advisor: An Agentic RAG Chatbot for Investment Guidance

  • Built the agentic RAG chatbot end-to-end as founding engineer
  • Retrieval pipeline over live rates from 11+ banking partners
  • Deterministic fallbacks so return figures never hallucinate
  • Live across 3,500+ Indian cities

Built the chatbot end-to-end as part of the founding team — a retrieval pipeline over live bank rate and product data, with a conversation flow that never oversteps into unlicensed advice. Bank rates and terms change daily, so a hallucinated return figure is a real trust and compliance risk — solved with RAG over a continuously-synced catalog plus deterministic fallbacks for numeric claims.

stable-advisor · chat
Best 1-yr FD rate for ₹1L?
Compared 11 banks — top pick: 7.8% p.a., ₹5L DICGC-insured.
Is it safe?
Yes — insured up to ₹5L per bank by RBI's DICGC.
11+BANKS COMPARED
3,500+CITIES SERVED
Illustrative UI — not an actual product screenshot
RAGOpenAI APIFintechConversational AI
CASE 06Stable Money

Payment Service & Multi-Bank Integrations for Fixed-Income Booking

  • Owned the payment service from scratch — idempotent and reconciled
  • Multi-bank integration pattern for fixed-income booking
  • Encrypted data handling across every banking partner
  • Zero double-bookings despite differing settlement timing

Designed and owned the payment service from scratch — idempotent processing, reconciliation, encrypted data handling — and led the integration pattern each new banking partner plugged into. Different APIs, failure modes, and settlement timing per partner meant the reconciliation and retry layer had to be robust enough to guarantee correct money movement, with no double bookings and no silent failures.

payments · stablemoney-prod
Investor App
↓
Payment ServiceIdempotent · Reconciled · Encrypted
↓
Bank Partner A
Bank Partner B
Bank Partner C
Idempotent Reconciled Encrypted
Illustrative architecture — not to scale
PaymentsBanking APIsReconciliationFintech
CASE 07DealShare

Gamezop: A Gamification Platform Driving 3x Engagement

  • Shipped the gamification platform end-to-end
  • Real-time price sync across microservices
  • 3x lift in daily engagement
  • Cut expired-deal errors by 90%

Designed and shipped the platform end-to-end, then built automation that kept selling prices and deal listings in sync across microservices the moment buying prices changed. Deal state lived across multiple microservices at once — a stale sync meant customers could buy an expired deal, so the fix had to be automated and real-time.

gamezop · rewards
Rewards 6-day streak
780 ptsNext: 1,000
3xENGAGEMENT LIFT
90%↓EXPIRED-DEAL ERRORS
Illustrative product UI — not an actual screenshot
ProductGrowthAutomationMicroservices
CASE 08Ecommerce Client

Cart Management API: Built to Extend, Not Rewrite

  • Layered FastAPI architecture — routes, services, repositories
  • Strategy-pattern promotion engine — extend with zero rewrites
  • Delivery-aware pricing with best-discount-wins logic
  • Inventory validation with clear, actionable error messages

Each promotion type — percentage, fixed — is its own class registered in a lookup table, so adding a new type means one file and one registry line; the cart service, routes, and database never change. Promotions are checked against the delivery date, not "now," and when multiple promos match a line, best-discount-wins avoids stacking into a negative total.

cart-api · promotions
POST /cart/items
↓
CartServiceDelivery-aware · Inventory-checked
↓
PERCENTAGE
FIXED
+ new types
4ARCHITECTURE LAYERS
ZeroROUTE CHANGES TO EXTEND
Illustrative architecture — not to scale
FastAPISQLAlchemyStrategy patternEcommerce

Contact

Have a system that needs to scale, or a project that needs to ship?

Open to consulting engagements in backend infrastructure, AI automations, and production LLM systems.

amitkr.iitbombay@gmail.com
Email me →

Book a discovery call

FIG. 04 — 15-MIN INTRO CALL
cal.com · amitkr/15min