Gaurav Ghop
Menu

Sheet 00 · Profile

GauravGhop

Backend Software Engineer, Distributed Systems & AI Agents

I build backend systems that hold up in production: tool-calling LLM agents with evals and monitoring, event-driven services on Kafka and AWS, and Postgres-backed data layers. I'm a founding engineer at OurFreedom.ai and I am working toward an MS in Computer Engineering at NYU.

View projects Resume (PDF) Get in touch
Portrait of Gaurav Ghop
Fig. A · New York, United States
agent requests a day at OurFreedom.ai
1,000
hallucination rate cut at OurFreedom.ai
16% → <5%
part records reconciled at eQ Technologic
40M+
fewer peak DB connections at Duende
85%

Sheet 01 · Experience

Where I've shipped

  1. Jun 2026 – Present

    In progress

    OurFreedom.ai

    Software Engineer, Founding Team

    New York, NY

    Founding engineer on the agent layer and the AWS production backend, shipping GTM MVP features with the founders.

    • Led Carla, a LangGraph tool-calling agent across 6 tools, with Pydantic-validated outputs, retry and fallback handling and a human review queue, serving about 1,000 requests a day.
    • Designed the eval pipeline: a 150-case gold set run nightly with Langfuse tracing. Hallucinations fell from 16% to under 5%, and tool-selection accuracy rose to 91%.
    • Added Prometheus and Grafana monitoring for p95 latency, tool failure rate and cost per request, with canary rollouts for every change.
    • LangGraph
    • Pydantic
    • Langfuse
    • AWS
    • Prometheus
    • Grafana

  2. Jun 2025 – Aug 2025

    Tiny Archives

    AI Software Engineer, Summer Intern

    New York, NY

    Built the ingestion and retrieval backbone for answering questions over archival collections.

    • Shipped a Python ingestion pipeline that normalized archival exports (EAD/XML, CSV, PDF) into a unified Pydantic schema, replacing 3 parsers.
    • Built a RAG pipeline with LangChain and pgvector that grounded answers in 40,000 indexed documents.
    • Cut the hallucination rate from 18% to 4%, measured by a weekly eval set of 500 queries.
    • Python
    • Pydantic
    • LangChain
    • pgvector
    • RAG

  3. Jul 2023 – Aug 2024

    Duende Ingenious

    Founding Engineer

    Pune, India

    Early engineer responsible for backend reliability, order processing and release automation.

    • Introduced PgBouncer connection pooling with PostgreSQL, reducing peak connection count by 85% and eliminating connection-exhaustion incidents.
    • Built an asynchronous order processing queue in Node.js on AWS SQS with exponential backoff retries and a dead letter queue, cutting silent order failures from 0.5% to 0.3% of daily volume.
    • Containerized backend services with Docker and automated CI/CD through GitHub Actions with integration tests, cutting the release cycle from 2 weeks to 1 week.
    • PostgreSQL
    • PgBouncer
    • Node.js
    • AWS SQS
    • Docker
    • GitHub Actions

  4. Aug 2021 – Jun 2023

    eQ Technologic

    Software Engineer I

    Pune, India

    Backend engineer on PLM, ERP and SAP integration products for enterprise and US defense customers.

    • Built idempotent Kafka consumers with automated reconciliation across 40M+ part records, detecting and resolving 1,200+ revision mismatches per month between PLM and ERP systems.
    • Developed Java Spring Boot microservices for a low-code API Gateway, adding retry logic, circuit breakers and JUnit integration tests. This reduced failed SAP and Teamcenter sync jobs by 40%.
    • Built a Kubernetes operator that provisions multi-tenant connector deployments from CRDs, eliminating 200+ hand-maintained manifests.
    • Java
    • Spring Boot
    • Kafka
    • JUnit
    • Kubernetes
    • SAML SSO

Sheet 02 · Projects

Things I've built to learn how systems fail

Featured

Dispatch Engine Simulator drawing

Backend systems

Dispatch Engine Simulator

Real-time dispatch for delivery-style matching: Redis GEO lookups with a PostGIS fallback, scored assignment, state-machine retries and Kafka events.

  • 5K-request bursty load test: about 180ms p95 dispatch latency and 92% dispatch success.
  • Shipped as freelance work for a food catering business: 300 orders/day and 90% fewer delivery-status inquiries.
  • Java
  • Spring Boot
  • PostgreSQL
  • Redis

Source on GitHub

TinyLSM schematic drawing

Storage engines

TinyLSM

An LSM-tree key/value store in C++17 with a memtable, write-ahead log, bloom-filtered SSTables and leveled compaction. Small enough to read in one sitting.

  • Demo writes 200k keys in 1,553 ms and reads back 50,000 of 50,000 random keys.
  • Crash recovery replays the WAL and scans SSTables on reopen, recovering all 50,000 sampled keys.
  • C++17
  • Bloom filters
  • WAL
  • SSTables

Source on GitHub

University Advisor Chatbot drawing

AI agents

University Advisor Chatbot

An agentic chatbot that gives international students source-cited university recommendations. Claude Sonnet runs a tool-calling loop over 5 custom tools on a FastAPI backend.

  • Sold and integrated into an Indian study-abroad consultancy's web app, serving 500+ students weekly.
  • Every recommendation cites its sources.
  • Python
  • FastAPI
  • Next.js
  • AWS Bedrock

Source on GitHub

locksmith drawing

Developer tools

locksmith

A migration linter that predicts which PostgreSQL lock a DDL statement takes, whether it blocks reads or writes, and how long it is held on a table of a given size.

  • Classifies statements into 18 operation kinds against the lock conflict matrix for PG 10, 11 and 12+.
  • Demo: a 400M-row table on PG 10 shows a 50.0 min write block, versus 5 ms on PG 16.
  • TypeScript
  • PostgreSQL
  • Node.js
  • SQL parsing

Source on GitHub

Also built

  • sysprobe drawing

    Observability

    sysprobe

    eBPF profiler for Linux: syscall, run-queue and TCP connect latency, exported to Prometheus.

    Source

  • gridboard drawing

    Real-time systems

    gridboard

    Backpressure-aware telemetry buffer that marks overloaded panels STALE instead of live.

    SourceLive demo

  • consentgraph drawing

    Data systems

    consentgraph

    Identity resolution where per-purpose consent gates the graph join itself.

    SourceLive demo

  • schemadrift drawing

    Developer tools

    schemadrift

    GraphQL breaking-change gate that uses real field-usage telemetry to set severity.

    SourceLive demo

  • dmarcwatch drawing

    Observability

    dmarcwatch

    Deliverability monitoring from DMARC reports that names the root cause of alignment failures.

    SourceLive demo

  • REQREQ, amendedREQREQTEST, passingTEST, failing COVEREDSTALE_COVEREDFAILINGUNCOVERED

    Requirements

    traceboard

    Traceability tool linking requirements, tests, code and defects; flags stale coverage.

    SourceLive demo

  • TOTP coderelayedassertionrelayedbank acceptsorigin checked LOGGED INREFUSED clientDataJSON.origin

    Security

    originbound

    A lab where one relay succeeds against TOTP and is refused by WebAuthn.

    SourceLive demo

  • fixed-horizon, peeked dailyalways-validthresholdlooks, day by day

    Statistics

    alwaysvalid

    Sequential A/B test calculator with always-valid inference and an A/A peeking simulation.

    SourceLive demo

  • ACCEPTEDstandardise onPostgreSQLACCEPTEDnew services defaultto DynamoDB no supersedes edge same subject · overlapping scope

    Developer tools

    adrlint

    Decision-record linter that flags unresolved contradictions between architecture decisions.

    Source

Sheet 03 · Stack

Parts list

Languages
Java · Python · TypeScript · JavaScript · C++ · SQL
Backend & APIs
Spring Boot · Node.js · FastAPI · Next.js · REST · Webhooks · Microservices · OAuth · JUnit · React Native
Data
PostgreSQL · PgBouncer · PostGIS · Redis · Cassandra · LSM-trees · Bloom filters · InfluxDB · MongoDB
Cloud & Infra
AWS (SQS, S3, EC2) · Docker · Kubernetes · Kafka · Flink · GitHub Actions · CI/CD · Grafana · Sentry
AI & LLM
Agent orchestration · RAG · LLM guardrails · LangGraph · LangChain · Langfuse · LoRA · MLX · Ollama

Sheet 04 · About

How I work

Outside work I write storage and GPU code from scratch: an LSM-tree key/value store in C++ and fused attention kernels in Triton and CUDA.

Each project in this sheet is a small system built around one way real systems fail.

Awards
Runner-Up, AWS GenAI Hackathon 2025
Awards
Finalist, Google DeepMind Hackathon 2025

Education

Master of Science · Computer Engineering

New York University

In progress ends May 2026 New York, NY

  • Relevant coursework: Distributed Systems, High-Performance Machine Learning, Design & Analysis of Algorithms
  • Merit Scholarship for graduate tuition

Bachelor of Engineering · Electronics & Telecommunications Engineering

Vishwakarma Institute of Information Technology

Ended May 2021 Pune, India

Sheet 05 · Contact

Let's talk backend

I'm always glad to talk about backend systems, LLM agents in production, or founding-team engineering roles. Email is the fastest way to reach me.

Open to backend engineering roles