Suraj Prasai — Software / AI Engineer

featured projects

What I've built

GeoTALOS

Geospatial AI Platform

AOI-driven platform for discovering imagery, running detection/segmentation models, and automating monitoring workflows through an interactive map and visual pipeline builder. Owned the FastAPI + PostGIS/pgSTAC backend; cut peak segmentation memory 8x (66.5GB → 8.4GB) via tiled raster streaming.

PythonFastAPIPostgreSQL/PostGISReactDocker Compose

Backend → Frontend →

GeoTALOS demo

AwakeForest (prototype)

KnowThyself

Agentic LLM Interpretability · AAAI 2026 Demo

LangGraph orchestrator-router agent that consolidates 5+ fragmented LLM interpretability tools into a single chat interface, routing natural-language queries to specialized analysis modules. Upload a model, interrogate its behavior conversationally.

PythonLangGraphLLaMA 70BNext.js

Backend → Frontend → Paper →

KnowThyself demo

more projects

WakeWander

Personal · Agentic Planning

8-node LangGraph plan-and-execute travel planner: human-in-the-loop interrupts, budget tracking, SSE progress streaming.

PythonLangGraphGemini 2.5

GitHub →

Cloud Speech Memo Analyzer

Personal · Production-style Cloud Service

3-stage Azure pipeline (Speech-to-Text → Language analysis → Neural TTS) deployed to Azure App Service. Instrumented end-to-end with OpenTelemetry distributed tracing and custom metrics queryable via KQL in Application Insights.

PythonFastAPIAzure AIOpenTelemetryAzure App Insights

GitHub →

Telephony RAG Assistant

Production · Nepal Telecom

National-scale RAG chatbot (hybrid retrieval) plus real-time ASR → LLM → TTS telephony pipeline on Kafka. Proposal secured £150K GSMA funding.

PythonLangChainKafka

Low-Resource ASR & LLM Fine-tuning

Speech & NLP · COLING 2025

LoRA fine-tuning for Whisper (WER 22→12.8) and Wav2Vec2 + n-gram/Transformer LM fusion (WER 54→18.3).

PyTorchHuggingFaceLoRA

Paper →

experience

Where I've worked

Graduate Research Assistant · Wake Forest University

Aug 2025 — Present

Winston-Salem, NC

Designed and engineered GeoTALOS, an AOI-driven geospatial AI platform (evolved from an earlier prototype, AwakeForest, accepted as an IJCAI-ECAI 2026 demo paper) replacing ad hoc, duplicated-TIFF workflows with one centralized system; owned the backend and spatial schema, cutting peak memory usage up to 8x (66.5GB → 8.4GB on SAM3 segmentation) and data-load latency to near-zero via tiled raster streaming — submitted to ACM SIGSPATIAL 2026.
Built a no-code visual workflow builder so non-programmers can run and automate ML inference pipelines (detection, segmentation, temporal NDVI monitoring) from a map UI; independently built the full React frontend.
Built KnowThyself (AAAI 2026 Demo Track), an agentic LLM interpretability assistant with an orchestrator-router pipeline (LangGraph + LLaMA 70B) that routes natural-language queries to specialized analysis modules, consolidating 5+ fragmented interpretability tools into one chat interface.

AI Engineer · Wiseyak Solutions

Aug 2020 — May 2025

Kathmandu, Nepal

Architected and shipped to production a RAG chatbot (hybrid dense + sparse retrieval, LangChain) for Nepal Telecom serving real-time customer queries at national scale; co-built a real-time telephony AI pipeline (ASR → LLM → TTS) on Kafka producer-consumer streaming and authored the technical proposal that secured £150K in GSMA international program funding.
Curated and augmented ASR datasets (300 hrs supervised, ~2,000 hrs unsupervised); designed and ran LoRA fine-tuning pipelines for Whisper, improving WER from 22 to 12.8 on Nepali transcription.
Built a Nepali instruction-tuning corpus via translation models; fine-tuned LoRA-based LLaMA3 8B and analyzed cross-lingual representations in low-resource settings — published at COLING 2025 (CHiPSAL Workshop).
Interviewed and grew the ML engineering team from scratch to 10 engineers; stood up on-prem GPU infrastructure (3x RTX 4090) and LLM serving pipelines (Docker, quantization) for enterprise clients; led and mentored the team throughout.

ML Research Intern · Integrated ICT Pvt. Ltd.

Jul 2019 — Mar 2020

Kathmandu, Nepal

Built an end-to-end Nepali ASR pipeline: cleaned a Devanagari speech corpus (font normalization, non-UTF filtering, spelling correction), fine-tuned Wav2Vec2, and fused an n-gram + Transformer LM, improving WER from 54 to 18.3; corpus work published in Nepalese Linguistics (2022).