Harsha Guntaka
Menu

Warren, NJ · Software Engineer

Harsha Guntaka

Software EngineerFull Stack AI ProductsAgentic AI

Software Engineer with 4+ years building full stack AI products and backend services. I own delivery from ambiguous requirements and prototypes through React/TypeScript interfaces, Python APIs, PostgreSQL models, and deployment. I use RAG, agents, product feedback, evaluations, and performance data to improve usability and reliability.

View projects Résumé Get in touch

01 · About

Products that measure themselves.

I build full stack AI products and the backend services behind them. At Airspan Networks I own an incident-investigation workspace end to end, from product discovery to production, using React/TypeScript, Python/FastAPI and PostgreSQL.

I treat AI features as products that need measurement. I iterate with pilot feedback and evaluate retrieval quality, agent tool selection and argument validity on held-out tasks.

Base
Warren, NJ
Core stack
React, TypeScript, Python, FastAPI, PostgreSQL
AI tooling
RAG, LangGraph, tool calling, evaluations
Degree
M.S. Computer Science, University of Bridgeport
Certificate
Data Science, IIIT Bangalore

02 · Experience

Two backend teams, one habit: prove it works.

Airspan Networks logo

Airspan Networks

Software Engineer, AI Product Engineering

– In progress

Works on the Network Investigation Workspace, an AI product for incident investigation, evidence retrieval, and saved findings.

  • Owned full stack delivery from product discovery to production, building React/TypeScript interfaces, Python/FastAPI services, and PostgreSQL data models for evidence-backed incident investigations.
  • Iterated on streaming answers, inspectable citations, and visible progress using pilot feedback; increased weekly active usage from 38% to 71% of enabled users.
  • Built LangGraph workflows for evidence retrieval and diagnostic tool calls with durable checkpoints, approval validation, and stable operation identifiers to avoid duplicate execution.
  • Evaluated retrieval quality, agent tool selection, and argument validity on held-out tasks; improved context selection and validation, increasing task completion from 82% to 94%.
  • Delivered AWS releases with Docker, GitHub Actions, and regression tests; monitored errors and latency and verified AI-assisted code changes through review and integration testing.
  • React
  • TypeScript
  • Python
  • FastAPI
  • PostgreSQL
  • LangGraph
  • AWS
  • Docker
  • GitHub Actions
  • Integration testing
Rakuten Symphony logo

Rakuten Symphony

Member of Technical Staff, Backend Systems

–

Worked on the Infrastructure Investigation Dashboard, a web product for event visibility, diagnostics, and investigation workflows.

  • Built Python REST APIs and asynchronous workers to ingest infrastructure events, validate schemas, and preserve searchable PostgreSQL event histories.
  • Developed React/TypeScript interfaces with live processing status, event filters, and failure drilldowns to streamline operational investigations.
  • Designed restart-safe processing with stable event identifiers, PostgreSQL uniqueness constraints, and bounded retry handling.
  • Investigated inefficient SQL queries and missing indexes, reducing p95 API latency from 800 ms to 310 ms under the measured workload.
  • Added pytest integration tests and automated release checks to catch regressions in event ingestion, persistence, and API behavior.
  • Python
  • REST APIs
  • Asynchronous workers
  • PostgreSQL
  • SQL
  • React
  • TypeScript
  • pytest

03 · Projects

Working notes, drawn as traces.

Each drawing shows what the project does. The notes list only what its README or the record states.

AI / RAG · Featured

3GPP RAG

A local question-answering system over 3GPP technical specifications. It combines BM25 sparse retrieval and ChromaDB vector search, fused with Reciprocal Rank Fusion, and answers with a local LLM through Ollama.

  • Parses .docx and .pdf specs into 1,000-character chunks with 200 overlap, then indexes them for both retrievers.
  • Fetches 6 candidates per retriever and sends the top 5 chunks to the LLM after fusion.
  • Runs fully locally with Ollama (llama3.2:3b for answers, nomic-embed-text for embeddings).
  • Python
  • ChromaDB
  • BM25
  • Ollama
  • LangChain
  • RAG

Source

AI / RAG

RAG Fundamentals Lab

A collection of Python utilities and demos covering each stage of a retrieval-augmented generation pipeline: document parsers, chunkers, vector stores, hybrid search and end-to-end RAG with Ollama.

  • A chunking tool with 8 strategies, from fixed-size and sliding window to section-based and token-based.
  • Parsers for PDF, Word, CSV and plain text, with metadata and JSON output for ingestion.
  • Progressively richer RAG apps with source citation and BM25 plus vector search fused by RRF.
  • Python
  • ChromaDB
  • FAISS
  • Ollama
  • BM25
  • LangChain

Source

Real-time systems

resumable

Server-sent events with a real Last-Event-ID contract over a bounded replay buffer. When the resume point has been evicted, the server sends an explicit GAP signal, so a reconnecting client never silently skips events.

  • In a seeded demo, 24,540 events were published during repeated reconnects; this implementation delivered all of them.
  • A test suite checks 1,000 scripted disconnects for zero lost events, zero duplicates and strict ordering.
  • Three distinct resume outcomes: FRESH, RESUMED and GAP.
  • TypeScript
  • Node.js
  • Server-Sent Events
  • Last-Event-ID
  • GitHub Actions

SourceLive demo

Developer tools

locksmith

A migration linter that parses PostgreSQL migration SQL and predicts which lock each statement takes, whether it blocks reads or writes on the target server version, and roughly how long it holds the lock on a table of a given size.

  • Classifies statements into 18 operation kinds and encodes the lock conflict matrix as data.
  • In the demo, a 400M-row table's total write-blocking time is 56.7 min on PostgreSQL 10 and 6.7 min on PostgreSQL 16.
  • 41 tests; unrecognised statements are treated as worst case, never as safe.
  • TypeScript
  • PostgreSQL
  • SQL
  • Node.js
  • GitHub Actions

Source

Developer tools

onespec

One declarative spec produces a compiled client validator, an interpreted server validator and SQL CHECK constraints. A property-based parity suite shows all three accept and reject exactly the same rows.

  • 10,000-row three-way parity test across client, server and SQL targets.
  • Cross-field rules (requiredIf, lessThanOrEqual) that hold in all three targets.
  • Three deliberately different implementations, and a mutation test showing the parity check detects drift.
  • TypeScript
  • SQL
  • SQLite
  • Property-based testing
  • Node.js

Source

Security & data

consentpath

Clinical trial data access where authorisation is decided by the data layer: security views keyed on session claims, with a negative-authorisation test matrix. The reference design is PostgreSQL row-level security; this repo uses SQLite views to build the same shape.

  • Site isolation, column-level blinding of arm assignment, delegated monitor grants and consent-gated export.
  • A 32-case matrix (4 roles x 2 sites x 4 participants) asserts denials as well as permits.
  • Append-only access log enforced by triggers; 16 tests.
  • TypeScript
  • SQL
  • SQLite
  • Authorisation
  • GitHub Actions

SourceLive demo

Real-time systems

gridboard

A backpressure-aware telemetry buffer for grid dispatch consoles. Under overload it downsamples visibly, always keeps the newest sample, and marks values STALE instead of showing old data as live.

  • Under a 10x overload the bounded buffer detected 4/4 injected faults; an unbounded queue detected 0/4.
  • Peak queue of 64 samples versus 18,001 for the unbounded baseline.
  • A seeded fault simulator makes every run reproducible; 16 tests.
  • TypeScript
  • Node.js
  • Backpressure
  • Simulation
  • GitHub Actions

SourceLive demo

Backend & data modeling

ratevault

A motor insurance quote engine with bitemporal rate tables, so any past quote can be repriced exactly using the rates and rules in force at the time. Each rate carries a valid-time and a transaction-time range, and rows are closed, never updated or deleted.

  • Separates what was quoted (GBP 1995.19 in the demo) from what should have been quoted after a backdated correction (GBP 1671.64).
  • A database trigger rejects overlapping valid-time ranges, and issuance is idempotent on a client token.
  • Every price comes back with the multiplier and running total of each rating factor.
  • TypeScript
  • SQLite
  • Bitemporal data
  • Node.js
  • GitHub Actions

SourceLive demo

Security

blastgraph

Cloud IAM analysis that models an estate as a privilege graph and finds multi-hop escalation paths to crown-jewel resources. It ranks which single permission to remove to break the most paths.

  • On a sample estate of 14 principals, 19 edges and 3 crown jewels, it reports 6 paths for an outside attacker.
  • Removing the top-ranked edge eliminates 6/6 paths (100%).
  • Reachability is a single recursive SQL query with a cycle guard and hop limit.
  • TypeScript
  • SQLite
  • Recursive SQL
  • Graph analysis
  • GitHub Actions

SourceLive demo

Reliability

lastknowngood

A backup recovery drill harness that measures real RTO and RPO by running restores and tests backup immutability by attempting deletion in a sealed lab.

  • Drill orchestrator with per-phase timing, reporting the measured RTO against the claimed one.
  • Immutability prober that tries to delete and overwrite each backup, rather than reading a configuration flag.
  • Dependency-order restore that detects credential cycles and names them in the report.
  • Python
  • Standard library
  • Disaster recovery

Source

Privacy engineering

erasure

GDPR erasure driven by column-level lineage across multiple data stores, verified by an adversary that tries to re-identify the person you just deleted.

  • Lineage discovery follows derived columns, so new derived tables are found without editing a table list.
  • Dry-run by default, with a certificate of tables touched, rows affected and what was retained.
  • Retention obligations recorded with the specific legal basis, not a boolean.
  • TypeScript
  • SQLite
  • Data lineage
  • GitHub Actions

SourceLive demo

Full stack AI

supportiq-analytics

Support ticket analytics: AI-assisted categorisation, sentiment tracking and a ticket volume forecast, shown on a dashboard for support teams.

  • Dashboard with real-time metrics, sentiment trend charts, a volume forecast and an agent leaderboard.
  • Ticket workflow with categorisation, sentiment scoring, priority, agent assignment and message threading.
  • TypeScript
  • React
  • Node.js
  • PostgreSQL
  • Redis
  • FastAPI
  • Docker

Source

04 · Skills

Where each skill ran.

Hover or tap a skill to light up the projects and roles where it was used.

Languages & Frontend

  • Python
  • TypeScript
  • JavaScript
  • SQL
  • React
  • Responsive interfaces
  • Server-Sent Events (SSE)

Backend & Data

  • FastAPI
  • REST APIs
  • PostgreSQL
  • Redis
  • pgvector
  • Data modeling
  • Asynchronous workers

AI Product Engineering

  • LLM APIs
  • RAG
  • LangGraph
  • Tool calling
  • Agent workflows
  • Prompt orchestration
  • Evaluations

Cloud & Quality

  • AWS
  • Docker
  • GitHub Actions
  • CI/CD
  • pytest
  • Integration tests
  • Observability
  • Latency optimization
  • AI coding agents

05 · Education

Trained in two countries.

  1. August 2024

    M.S. Computer Science

    University of Bridgeport, CT

  2. 2021

    B.Tech. Electronics & Communication Engineering

    Amity University, India

  3. 2022–2023

    Advanced Certificate in Data Science

    IIIT Bangalore

06 · Contact

Let’s build the next one.

Interested in full stack AI products, agentic workflows or backend systems? Send me a message and I'll get back to you.

harsha.guntaka31@gmail.com