Abira Wisnunggal

I build AI systems that drive real outcomes

AI Engineer · ML Engineer · Data Scientist

AI Engineer in training focused on production-ready RAG and machine learning workflows. I build AI assistants and data pipelines that reduce repetitive support work, improve response consistency, and integrate cleanly with backend services.

Open to AI Engineering Internship · Freelance Collaboration

View Flagship Case Study → Hire Me →
Experience & recognition at
  • Nawatech
  • GAOTek
  • Home Credit Indonesia
  • Bronze · Pesta Data Nasional 2025
Abira Wisnunggal profile photo Upload Your Profile Photo
6+
Roles & Internships
6
Projects Built

About

01   About

Abira Wisnunggal Call me Abi.

I'm a final-year Data Science undergraduate at Nusa Mandiri University (GPA 3.73 / 4.00). For the past 1.5 years I've been building LLM applications, mostly RAG and Text-to-SQL, alongside classical ML models for classification and forecasting.

Most of that work is backend: retrieval pipelines (hybrid RAG, knowledge graphs, pgvector), Text-to-SQL with validation rules, and async services on FastAPI, Celery, RabbitMQ and Docker. I've also produced ground-truth annotation for autonomous driving and built BI dashboards in Looker Studio and Tableau.

I'm currently an AI Engineer Intern at BRIN, building question answering over a bibliometrics database. I'm looking for an AI engineering internship or junior role.

Current Tech Stack
AI / ML
  • Scikit-learn
  • XGBoost
  • YOLOv8
  • TensorFlow
  • PyTorch
LLM Systems
  • RAG
  • Text-to-SQL
  • Knowledge Graphs
  • Azure OpenAI
Data Stack
  • Pandas
  • NumPy
  • Tableau
  • Looker Studio
Storage & Retrieval
  • PostgreSQL
  • pgvector
  • MySQL
  • Redis
Infra & Deployment
  • FastAPI
  • Docker
  • Celery
  • RabbitMQ

Experience

02   Experience

Six roles, most recent first. Open a role to see what I did there.

What I did
  • Built question-answering over 20 publications, 138 authors, and 344 keywords — evaluated on 94 test queries with 82.6% search relevance and 88.6% retrieval success
  • Built a Text-to-SQL pipeline with Qwen2.5-Coder, sqlglot, and PostgreSQL with validation rules for natural-language bibliometric questions
  • Added safeguards: read-only DB access, SQL-injection protection, 10s query timeouts, 60 req/min rate limiting, and citation verification — 726 unit/integration, 38 live E2E, and 12 retrieval regression tests passing
What I did
  • Annotated real road scenes for lanes, road borders and edges, stop lines, road markers, parking slots, and other road structures
  • Labeled lane centerlines by driving direction and lane type — one-way and two-way roads, general, emergency, bus, non-motorized, tram, and specialized lanes
  • Captured detailed marking boundaries at intersections (line intersections, colors, thickness, guidance and deceleration markings, faded markings) for consistent ground truth across road conditions
What I did
  • Built the backend with FastAPI, Celery, RabbitMQ, Redis, and Docker — 6 containers, 5 RabbitMQ queues, and 4 Celery workers for QnA, PDF, and SQL workloads
  • Developed the RAG and Text-to-SQL pipeline with pgvector and Azure OpenAI embeddings, including schema retrieval and few-shot SQL generation — 3.23s mean end-to-end latency on 60 benchmark queries (0.40s retrieval, 2.57s LLM generation)
  • Created an automated LLM evaluation and regression-testing setup with DeepEval and Azure OpenAI, generating Excel and Markdown reports across 60 test cases and difficulty levels
What I did
  • Built an AI chatbot using LLM-based conversation, intent classification, and context-aware responses — reaching around 88% intent classification accuracy
  • Implemented semantic search and embedding-based retrieval, including query preprocessing, document chunking, and similarity matching
  • Tuned prompts and retrieval across response accuracy, retrieval quality, and latency to bring average response time to ~2.5 seconds
What I did
  • Analyzed customer and operational data using Python and SQL to identify patterns and key metrics across multiple datasets
  • Prepared and validated 5+ datasets through data cleaning, transformation, and quality checks
What I did
  • Built 4 interactive Looker Studio dashboards to monitor syllabus coverage, student progress, and assessment results across multiple departments
  • Consolidated student progress data from spreadsheet-based masterfiles into a centralized reporting pipeline, covering 950+ records and reducing manual data preparation for recurring reports

Projects

03   Projects

Where a project has a measured result, it's shown with the test it came from. Projects without one show what was built and nothing more.

Demo Lab logo
001 · Flagship

DemoLabs AI Chatbot

Nawatech internship · Oct 2025 – Feb 2026
0.885 mean answer correctness on 60 DeepEval test cases

One chat interface for analytics questions. It routes each question to document QnA, RAG-based Text-to-SQL, PDF/OCR analysis or chart generation.

  • 3.23s mean end-to-end latency on 60 benchmark queries (0.40s retrieval, 2.57s LLM generation)
  • 6 containers, 5 RabbitMQ queues and 4 Celery workers keep QnA, PDF and SQL workloads apart

RAG · Text-to-SQL · FastAPI · Celery · pgvector

Case study Demo video (YouTube) GitHub
nuxio.ai logo
002

nuxio.ai

In progress · Sep 2026 – present

A financial planning app: multiple wallets, budgets, savings goals and scheduled transactions, with a Gemini-based copilot that returns structured outputs the planner can act on, not free text.

Next.js · TypeScript · Supabase · Gemini

Looker Studio dashboard showing one student's Qur'an and Hadits completion rates by subject
003

Student Progress Analytics Dashboard

Jan 2025 – Apr 2025

A Looker Studio dashboard tracking Qur'an and Hadits memorization progress. Per-student completion rates roll up into class and cohort summaries, built from consolidated spreadsheet masterfiles.

Looker Studio · Spreadsheet data pipeline

004

Sentiment Analysis of Telco Tweets

Jun 2025 – Sep 2025
0.82 F1-score, three-class sentiment on ~10,000 telecom tweets

TF-IDF features and logistic regression to classify customer sentiment, plus the most frequent keywords in negative tweets as input for service fixes.

Python · TF-IDF · Logistic Regression

GitHub
005

Car Sales Forecasting: Indonesia's Top 5 Brands

Jul 2024 – Sep 2024
<10% SMAPE on every one of the five brands

Trend and seasonality decomposition per brand, then SARIMA, Prophet and LSTM trained and compared side by side.

SARIMA · Prophet · LSTM

006

Traffic Monitoring with YOLOv8 + DeepSORT

Jun 2024 – Jul 2024

YOLOv8 detects vehicles and DeepSORT tracks them, so each vehicle keeps one ID across frames. Counting and flow analysis run on those tracks.

YOLOv8 · DeepSORT · Computer vision

Certificates

04   Certificates
BNSP logo
Professional certification
BNSP (National Professional Certification Agency)
Associate Data Scientist
Indonesia's national competency certification for associate data scientists.
Skilvul certificate of completion for Python Lanjutan (Python Advanced)
Skills credential
Skilvul
Python Advanced
Advanced Python for data processing and automation scripts.
View certificate
RevoU certificate of completion for Intro to Data Analytics
Course completion
RevoU
Intro to Data Analytics
Two-week online course: SQL fundamentals, dashboards and presenting analysis.
View certificate
Dicoding logo
Course completion
Dicoding Indonesia
Data Visualization
Chart selection and dashboard design.

Honors and Awards

05   Honors & Awards
Aug 2025
Bronze Medal, Pesta Data Nasional x APTIKOM 2025
APTIKOM
National data competition · 3rd place
View certificate
2025
Top 160 of 448 teams, Satria Data
Satria Data (National Data Competition)
Preliminary stage · 97.12% accuracy
Jun 2025
Qualified scientific paper, Bank Indonesia
Bank Indonesia
National scientific paper program
Proposed a Smart Invoice concept that uses OCR-assisted extraction to automate financial documentation.
2025
Top 5, abstract competition, Agentic AI Hackathon
Agentic AI Hackathon
Abstract for an applied agentic-AI concept
Dec 2023
3rd place, International Invitational Contest of IoT Technology
China-ASEAN Education Cooperation Week
Multi-country contest
View certificate

RAG Pipeline Demo

06   Demo

How a question moves through the pipeline

The same stages as the DemoLabs chatbot: embed the question, retrieve chunks from pgvector, assemble the context, generate an answer. Ask about my work to step through it.

Scripted simulation. No model is called; the answers are pre-written from the content of this page.

Q
Query
—
E
Embed
Azure OpenAI
R
Retrieve
—
C
Context
—
G
Generate
Azure OpenAI

Contact

07   Contact

Got a project
or opportunity?

I'm looking for an AI engineering internship or junior role, and I take on freelance work. Email and the form both reach me directly.

abirawisnu@gmail.com Email www.linkedin.com/in/abiwisnu LinkedIn github.com/CodeByAbi GitHub