01 / Infrastructure
scaleserve · remote runner
ScaleServe
Your entire fleet, one command away
Flutter desktopFastAPITailscale
Available for AI & automation workPortfolio · 2026
AI & Automation Engineer · Melbourne, Australia
I build AI and automation systems that run in production: website generators, voice agents, fleet tooling and real-time scam detection on IBM Z.
- Right now
- AI & Automation EngineerSoudCoh · part-time
- Master's of Data ScienceRMIT University · GPA 3.5 / 4.0
- IBM Z Student AmbassadorRMIT Mainframe Club
0
Projects shipped
0
Internships
0
Part-time role
0/1000
AWS cert score
01 / Selected work
Ten projects across AI automation, infrastructure, ML and full-stack. Open any card for the full story, screenshots and results.
02 / AI & Automation · Full-stack
WebAuto
Finished small-business websites in under 10 minutes
Claude agentsReactBuilt at SoudCoh
03 / ML & Data · AI & Automation
ScamShield
Real-time scam and mule-account interception on IBM LinuxONE
1st place · IBM Z Datathon 2026IBM z17 · Telum IITeam lead
04 / AI & Automation · Full-stack
IncAutomate
AI front-desk automations for any business, self-hosted
n8n + Hugging FacePostgres + QdrantMulti-tenant
05 / AI & Automation
“Hi, thanks for calling…”
Conversa
An AI receptionist on your phone line
Voice AITwilio media streamsOpenAI / Claude
06 / AI & Automation · ML & Data · Infrastructure
LLM Experiment Bench
A live test bench for squeezing 20B coding LLMs onto one 16 GB GPU
Local LLMsllama.cpp · CUDALive test bench
07 / ML & Data
PredStroke
Stroke-risk prediction in a Flutter app
Flutter appML classifier
08 / AI & Automation
SimpleAutomate
Tap any element, get its XPath
FlutterXPath engine
09 / AI & Automation · ML & Data
Finance Agent
An agent that reads your bank statements
LLM pipelineLive
10 / Full-stack · ML & Data
RentMyHouse
Rental operations with ML rent benchmarking
React + FlaskLive
11 / ML & Data · Full-stack
Folio Pulse
Portfolio analytics from ECAS PDFs
Built at HDFC AMCLive
03 / ML & Data · AI & Automation
Overview
Australians lost $2.03B to scams in 2024, and the money moves through mule accounts within minutes. ScamShield checks every payment before the money moves, in milliseconds, on an IBM z17: most go straight through, risky ones get a clear warning, and mule-pattern payments are paused for a human. I led Team IBMBoB (with Vaishnavi Joshi) to first place at the Federation University hub of the IBM Z Datathon 2026, a global IBM event, and built the LinuxONE deployment, Telum II benchmarks, scoring API, security and audit chain.
Stack
PythonFastAPIPyTorchXGBoostSHAPIBM Granitellama.cppDockerIBM LinuxONE
What it does
- 01One streaming feature engine over payment, payer and receiving-account history, O(1) per payment
- 02Customer-behaviour autoencoder on the Telum II on-chip AI accelerator (NNPA) plus an XGBoost risk model with SHAP reasons
- 03IBM Granite (llama.cpp + zDNN) writes analyst briefs; a live dashboard has analyst, customer and auditor views
- 04Security: HMAC pseudonymisation, AES-256-GCM on CPACF, JWT roles, and a hash-chained audit log that flags any edited row
Measured
4.4 ms
p50 server latency
274/s
live on 3 vCPUs
1.4×
Telum II NNPA vs CPU
0.074%
of payments paused
Two unseen test days, scored on the IBM z17
Risk model ROC-AUC on 862,792 unseen payments0.984
Laundering schemes broken (≥1 payment paused)147 / 174
Laundering payments warned or paused82%
Paused payments that were laundering73%
Laundered money paused before it moved$269M / $424M
How it decides
1
ALLOW · 98.26%
Nothing unusual: the payment goes straight through.
2
WARN · 1.66%
Risky, out of character, or a big first payment to a new payee. The customer gets a specific warning and decides; warnings never block.
3
HOLD · 0.074%
Strong mule-account match. Paused while the bank calls back on a known number; an analyst gets an IBM Granite evidence brief and a human decides.
Honest limits: the public IBM AMLworld data labels laundering, not scam victims, so results measure breaking mule chains. 95% of the money paused is ten very large transfers; the median paused laundering payment is $11,266.
02 / Experience
Built WebAuto, a website-generation pipeline that turns a website-type wireframe, a client's intake answers and photos into a production React (Vite) site, cutting build time from 2–3+ hours to under 10 minutes per site.
03 / Community
IBM Z Student Ambassador at RMIT University, building the Mainframe Club and getting students started on enterprise computing.
Competition · 17–18 Oct 2026
🏆 First place · Federation University hubIBM Z Datathon 2026
A 24-hour, students-only IBM Z data competition run by IBM, at Federation University's Berwick campus. I led Team IBMBoB in Theme 1, Real-Time AI for Critical Decisions, and we took first place at the Federation University hub with ScamShield: real-time scam and mule-account interception on an IBM z17.
IBM Z Student Ambassador · RMIT
Mainframe Club
Club site, IBM Z Xplore referral flow with poster QR codes, RMIT member sign-up, events, a Discord community bot and privacy-friendly analytics, all self-hosted.
17 Oct 2026Won first place
IBM Z Datathon 2026
Federation University, Berwick
Led Team IBMBoB to first place at the Federation University hub with ScamShield, real-time scam interception on an IBM z17. Also promoted the event to RMIT students through the club.
19 Nov 2026Promoting
Asia Pacific IBM Z Day: Special Edition
Online, live
IBM's regional IBM Z day, shared with club members as a watch-along.
20 Nov 2026Organising
IBM Z Career Connection @ RMIT Grad Bootcamp
RMIT Megaflex, Building 8
Hands-on IBM workshop: getting started on IBM Z Xplore, mainframes in government and finance, enterprise AI skills and IBM Z careers.
04 / About
I build data-driven systems and AI automation: pipelines, backends and agents that turn messy inputs into reliable products.
At HDFC Asset Management I worked on end-to-end data systems involving unstructured PDF extraction, aggregation and analytical visualisation. In Melbourne I've built AI voice automations at Conversion Assist AI, website generation at SoudCoh, and real-time scam detection on IBM Z.
I completed a BTech in Computer Science Engineering at VIT and I'm now doing a Master's of Data Science at RMIT University, focused on distributed systems, statistical thinking and scalable data infrastructure.
Education
2025 – 2027 (expected)
RMIT University
Master's of Data Science
Melbourne, Australia · GPA 3.5 / 4.0 · Semester 1 complete, Semester 2 nearly done
2021 – 2025
Vellore Institute of Technology
BTech Computer Science Engineering
Vellore, India · GPA 7.81 / 10
2019 – 2021
Nirmala Memorial Foundation College
Higher Secondary Certificate (HSC)
Mumbai, India · Score 84.83%
Certifications
AWS Certified Cloud Practitioner
Amazon Web Services
Score 916 / 1000 · valid till January 2027
Google Cloud Digital Leader
Google Cloud
Valid till February 2027
05 / Toolbox
01AI & Automation
Claude APIClaude Code agentsMulti-agent pipelinesn8nMakeHugging FaceRAG / QdrantIBM GraniteTwilio voiceElevenLabsPower Automate
02Data & ML
PythonPandasPyTorchXGBoostScikit-learnSHAPData pipelinesPDF extractionRechartsPower BI
03Backend & Databases
FastAPIFlaskPostgreSQLSQLAlchemyREST APIsWebSocketsJWT / MFA authSQLite
04Frontend & Mobile
React / ViteTypeScriptTailwind CSSFlutterDartJavaScript
05Cloud & DevOps
DockerLinuxNginxGitAWS (S3, Lambda, QuickSight)TailscaleIBM LinuxONESSH
06 / Quick answers
01Who is Soham Athawale?
Soham Athawale is an AI and automation engineer in Melbourne, Australia. He works part-time as an AI and Automation Engineer at SoudCoh, studies a Master's of Data Science at RMIT University (GPA 3.5 / 4.0), and is an IBM Z Student Ambassador who runs the RMIT Mainframe Club. His team won first place at the Federation University hub of the IBM Z Datathon 2026.
02What does Soham build?
Production AI and automation systems: WebAuto (finished small-business websites in under 10 minutes), Conversa (an AI phone receptionist), IncAutomate (self-hosted AI front-desk automations), ScaleServe (fleet control for Tailscale-connected machines), ScamShield (real-time scam interception on IBM LinuxONE, first place at the IBM Z Datathon 2026, Federation University hub) and LLM Experiment Bench (a live test bench for quantizing and benchmarking local coding LLMs).
03What is WebAuto?
WebAuto turns a website-type wireframe, a client's answers to business questions and their photos into a finished, production-quality React site in under 10 minutes, replacing an AI-writes-everything approach that took 2–3+ hours per site. Soham built it at SoudCoh.
04What is ScamShield?
ScamShield checks every bank payment before the money moves, in milliseconds, on an IBM z17. It reached a risk-model ROC-AUC of 0.984 on 862,792 unseen payments and broke 147 of 174 laundering schemes. Soham led Team IBMBoB to build it at the IBM Z Datathon 2026, where it won first place at the Federation University hub.
05What is ScaleServe?
ScaleServe is a cross-platform desktop app for managing Tailscale-connected machines: it runs remote SSH commands, streams local scripts to any runtime and fans work out across devices, with MFA-protected central auth.
06Is Soham available for work?
Yes. Soham is open to AI and automation roles, internships, collaborations and interesting projects in Melbourne or remote. Email sohamathawale2003@gmail.com.
07 / Contact
Open to AI and automation roles, internships, collaborations and interesting projects in Melbourne or remote.