Oken Keithellakpam
Shifting the Paradigm
AI Specialist · Full-Stack Developer · Open Source Maintainer
Transmuting the structural foundation of system and chaos, to undergo metamorphosis through the lens of patterns and time.
Who Am I?
I like to make things, break things, and fix things but importantly learn. Understanding systems in a world of chaos, trying to make sense in a world of probabilities.
"Nevertheless, to enjoy the process, when living is the goal."> Core Modus Operandi
Where I've Worked
Navigating the constellation of my career—from open-source orbits to scaling AI galaxies.
CubeTen
AI Specialist
Developing AI powered systems involving document retrieval, chatbot assistants, and GeoAI workflows. Built vector and keyword search systems, contributed to AI solutions for the Manipur eSeaba government portal, and explored GIS, remote sensing, and segmentation based approaches for environmental and water encroachment analysis.
Outlier.ai
Freelance AI Evaluator
Evaluated large language model outputs for alignment, reasoning quality, truthfulness, instruction following, and safety. Contributed high quality annotations and feedback for improving model performance.
Nibiaa Devices
AI Engineer
Developed and optimized computer vision systems for real time EV bus monitoring on NVIDIA Jetson Nano devices. Delivered mobile applications integrated with AI powered monitoring workflows.
NIT Manipur
Summer Research Intern
Worked on English–Manipuri translation datasets through proofreading, preprocessing, and quality validation. Applied NLP techniques to improve dataset reliability and consistency.
Freelance Developer
Full Stack & ML Developer
Built and deployed full stack applications using MERN, Django, and FastAPI for multiple clients. Developed machine learning features including NLP systems, classification models, and computer vision workflows while mentoring learners in ML development and deployment.
Creative Space
A collection of production-grade AI systems, research experiments, and full-stack architectures.
e-Gazette RAG System
Built a hybrid document intelligence system for a government client over a 14GB+ corpus of official gazette records. Historical documents with physical wear made standard OCR unreliable, so the pipeline uses Dolphin VLM for degraded text extraction, feeding into a hybrid BM25 + pgvector retrieval layer with reciprocal rank fusion. Returns summarized answers with source citations to specific gazette sections.
Manipur Jal Sampada
Geospatial water encroachment detection system built for a client from scratch — including learning GIS from zero. Initial approach using NDWI/NDBI/NDVI index thresholding proved too error-prone on real terrain, so switched to the FLAIR land classification model (ResNet34-UNet) on Sentinel-2 multispectral imagery, which significantly improved classification accuracy. Generates semantic event masks (urban expansion, flooding, wetland conversion) exported as GIS-ready layers.
Face2Comic
A conditional GAN that translates real portrait photographs into comic-style illustrations. The conditional architecture enforces correspondence between input and output (unlike standard GANs), using a U-Net generator for spatial detail preservation and a PatchGAN discriminator for high-frequency texture realism. Trained model published on HuggingFace.
Image Captioning
An image captioning multimodal model that generates natural language captions for images using the Flickr8k dataset. CNN encoder extracts visual features; Transformer decoder attends over them to generate sequences. The cross-modal attention mechanism is what makes this architecturally interesting — the model learns which image regions matter for each word in the caption.
Nibiaa P2P App
Forked and modified the Meshtastic open-source codebase to build a mesh networking P2P app. Shipped to the Play Store in under 2 weeks with zero prior Android experience.
Graph RAG Experiment
Tested Graph RAG against naive vector retrieval on a PDF book corpus. Vectorized the document with an embedding model and explored how graph-based context retrieval affects summarization and question answering quality.
Local Study Partner AI Agent
Feed a PDF syllabus, get a study plan, timeline, MCQs, and analytics back. Built on a local LLM with RAG. Dropped when local inference got too slow — worth reviving with API.
SKOLAR ML Internship
3-month ML internship. Worked through data preprocessing pipelines (mean, median, KNN imputation), Lasso regression, KNN K-value tuning, Decision Tree vs Random Forest comparison, and PCA dimensionality reduction and reconstruction.
Milestones & Recognition
A record of hackathons, open source leadership, and community speaking engagements.
Speaker
March 1, 2025Delivering a speech on Open Source Contribution, Importance and Oppourtunites at Cyberlla 2025 held at Nielit, Imphal