No. 01 · Lead
AI Engineer
Fine-tuned CLIP, hybrid retrieval and an AI sales copilot, all in production.
Proof: Toko MarcellMarcell Hermawan Kristianto
AI Engineer & Data Scientist
Intern candidate · Data Science at Binus. I build AI systems, then measure them against a baseline.
Roles I am applying for
No. 02
Data Scientist
Every retention strategy costed in dollars; the model spends 18% less than discounting everyone.
Proof: Telco churn
The folio
Two pieces of work you can open
I · The shop
Recall@126.5%39.2%539 held-out pairs · vs. baseline CLIP
Live
Toko Marcell
A storefront built as a test bench for retrieval.
I fine-tuned CLIP on 5,378 fashion image–text pairs and fused it with BM25 and dense text search in pgvector using reciprocal rank fusion. On top sits a LangChain copilot that works as the shop's salesperson: it finds products for a shopper, recommends and sells them, and answers sizing and store-policy questions.
- PyTorch
- CLIP
- pgvector
- FastAPI
- LangChain
- Next.js
Open the live store ↗ Read the code ↗
Backend sleeps on a free tier; the first search can take ~40 s.
II · The till
Cost / month$6,439$5,259986 held-out customers · vs. discounting everyone
Notebook · reproducible
Telco churn
Who should get the retention discount?
I costed six targeting strategies on the same held-out customers. Logistic regression with a threshold chosen by cost instead of the default 0.5 spends 18% less than discounting everyone. A one-line contract rule gets most of that gain, and the notebook says so, along with the offer success rate below which the model stops winning.
- scikit-learn
- imbalanced-learn
- XGBoost
- LightGBM
- pandas
The press
Published research
The ledger
ICOBAR 2025 · Springer Nature LNEE 1693 · in press
Model Comparison for Toxigenic Cyanobacteria Image Classification Using Transfer Learning and Post-quantization
B. Manafe*, M. Kristianto, D. Thio, G. Elwirehardja, B. Pardamean
| Backbone | Acc. |
|---|---|
| EfficientNetB1 | 96.72% |
| EfficientNetB0 | 95.37% |
| MobileNet (prior work) | 93.63% |
| ResNet50 | 92.47% |
FP16 25.5 MB → 12.8 MB · 45.3 ms CPU · Δacc 0.00
Chapter in press
2nd of 5 authors
ICOBAR 2025
Which CNN should run on a low-power microscope?
We compared six pretrained CNNs on microscope images of toxigenic cyanobacteria. EfficientNetB1 reached 96.72% test accuracy, and FP16 post-training quantization halved its size with no accuracy lost. The work was done with the BDSRC research lab and accepted at ICOBAR 2025.
- TensorFlow
- Keras
- TFLite FP16
- Transfer learning
The jacket
Patches earned on real work
Each patch names where I used it.
Models & evaluation
PyTorch · CLIPfine-tuned image–text retrieval
CLIP · VLMsimage–text search
TensorFlow · Keras6-CNN benchmark, FP16- scikit-learn · XGBoostchurn models, cost-chosen thresholds
EvaluationRecall@k · MRR · PR-AUC
LLM & agents
LangChain · Geminisales copilot, tool calling- NVIDIA NIM · RAGLLM app workshop, 9 h
- ONNX Runtimecross-encoder reranker
Hermes · Pi Agentdaily since Apr 2026
Data & serving
PostgreSQL · pgvectorHNSW + BM25 fusion
FastAPImodel serving
PySpark · Airflowbatch jobs, scheduling
Docker · Linuxcompose stacks
Next.js · Supabasestorefront, team portal- Python · R · SQLR for Bayesian stats
The campus walk
How I got here
Five stops since my first semester, and where I am now.
-
Binus University
Began a B.Sc. in Data Science. GPA 3.85 / 4.00.
-
Research Activist
Binus Data Science Club, with the BDSRC lab. Output: an ICOBAR 2025 paper, in press with Springer Nature.
Read the paper ↓ -
PIC, Data Seeker
Lead 15 delegates in 5 squads, picked through a Kaggle entrance test I set and judged. Built the portal that runs the team's claims and solution vault.
The team's portal ↗ -
Mentor Scholarship
SASC Binus, on a merit grant. Junior mentor, promoted to senior in Aug 2026. Weekly tutoring with study modules I design.
-
TA · NVIDIA DLI workshop
Teaching assistant for NVIDIA's 9-hour Rapid Application Development with LLMs workshop at Binus: poster, venue, registration, screening, and helping participants debug their labs.
The club's post ↗
The credentials book
Ten certificates, each one checkable
Open one to read it here, or follow it to the issuer.
PyTorch for Deep Learning
Generative AI with LLMs
Agentic AI
Rapid Application Development with LLMs
Fundamentals of Deep Learning
Introduction to LangGraph
Introduction to LangChain
AI RAG Systems
Agentic AI
Full-Stack Web Development