PythonFastAPINext.jsReactTypeScriptPyTorchLangChainLangGraphTemporalpgvectorPostgreSQLAzureDockerNeMo GuardrailsPresidioGitHub ActionsPythonFastAPINext.jsReactTypeScriptPyTorchLangChainLangGraphTemporalpgvectorPostgreSQLAzureDockerNeMo GuardrailsPresidioGitHub Actions
01Projects

Projects

01 / 06
01

SecureSphere-LLM

Node.jsNeMo GuardrailsPresidioFastAPIReactLLMSecOps

Live telemetry: allowed, blocked and sanitized traffic with recent threats

An LLMSecOps platform that sits between VS Code and the AI coding assistants developers use every day. Every request to Copilot, Gemini, Cloud Code or Claude is intercepted, tied to the developer who sent it, screened by NeMo Guardrails, and logged before it reaches a model. Admins decide per team which protections apply and watch the traffic live from a dashboard. Built as a final year project across three services.

  • Interception proxy

    A Node.js man in the middle proxy that decrypts editor traffic, normalises prompts and attachments from four assistant protocols into one contract, resolves the developer and team from a shared identity store, and rolls guardrails out per provider and per model behind release gates.

  • Guardrails service

    A FastAPI server on NVIDIA NeMo Guardrails with Colang rails for prompt injection, jailbreaks, secrets such as cloud keys and tokens, dangerous commands, topic control and output checks, with Presidio masking personal data and a pytest suite covering each attack class.

  • Admin dashboard

    A React control plane with JWT and OAuth sign in, live request telemetry, an audit log viewer, a policy workbench for switching protections per team, and identity aware group management.

02Experience

Experience

  1. Jun 2026 to now
    Agento logo

    AI Security Research Lead at Agento

    Enterprise agentic workflow platform, Melbourne, Australia

    Lead AI security research for a platform where agents compose and run multi step business workflows: guardrails and runtime monitoring, a connector action policy engine, multi provider LLM routing, agent memory and ontology design, and the RAG knowledge base. Agento was selected for the Australian Government backed Buy Australian AI Partnership Program.

    • Audited 825 LLM call sites, closed the one unguarded bypass, and enforced 3 pre-dispatch invariants in Temporal under CI, then proved coverage by reintroducing the bypass and catching it at the exact call site.
    • Built a policy engine with per tool allow lists and RBAC gated approvals across 10 integrations, enforced at a single execution chokepoint.
    • Built complexity based LLM routing with semantic caching across OpenAI, Anthropic, Moonshot AI and Gemini, with up to 30x lower per token cost on simple tasks and no loss in output quality.
    • Architected a multi tenant RAG module shipped across 7 releases with 29 automated tests, using pgvector HNSW retrieval and row level security to isolate every tenant.
    • Hardened 6 public endpoints with spoof resistant per IP rate limiting and Turnstile CAPTCHA, and secured webhook ingress with constant time HMAC verification.
  2. Mar 2026 to Jun 2026
    Imperium Dynamics logo

    AI Software Engineer I at Imperium Dynamics

    Microsoft partner, Chicago, USA

    Built scalable LLM applications and LangChain agents with modular Python architectures, governed data pipelines on Microsoft Fabric, and AI workflows on Azure AI Foundry and Azure Pipelines.

    • Built LLM applications and LangChain agents on modular Python architectures designed to scale across client projects.
    • Designed Microsoft Fabric data pipelines on a medallion architecture, governed end to end with Microsoft Purview for data lineage and cataloguing.
    • Delivered AI workflows and cloud automation with Azure AI Foundry and Azure Pipelines, with secure backend authentication for AI services.
    • Standardised development with Git and PDM managed environments for reproducible builds and deployments.
  3. Feb 2025 to Feb 2026
    Analytic Insider logo

    Generative AI Associate Developer at Analytic Insider

    Generative AI applications, Dearborn, Michigan, USA

    Built generative AI products including a RAG based tax consultant bot with LangChain and Pinecone, and automated structured data pipelines across thousands of business directories.

    • Built a RAG tax consultant bot on LangChain and Pinecone that improved query retrieval efficiency by 12%.
    • Automated structured data pipelines covering 3,000+ business directories to feed LLM applications at scale.
    • Shipped 2 GPT-4 tax assistants, BotTax and US TaxPatriot, with voice input and source citations down to the page.
    • Migrated retrieval from local FAISS to Pinecone with 3072 dimension OpenAI embeddings so the bots could scale.
  4. Sep 2024 to Dec 2024
    10Pearls logo

    Data Science Intern at 10Pearls

    Forecasting and real time pipelines, Virginia, USA

    Developed an air quality prediction system with real time OpenWeather and Hopsworks pipelines, and automated CI/CD with GitHub Actions.

    • Built an AQI forecaster with Random Forest, Ridge and Logistic Regression that reached 78% accuracy.
    • Forecast air quality 72 hours ahead from a year of hourly OpenWeather pollution data.
    • Automated CI/CD with GitHub Actions, cutting deployment time by 30%, with hourly feature runs and daily retraining.
    • Explained predictions with SHAP and LIME, and served the forecast through a Streamlit dashboard with hazard alerts.
03Research

Research

Two first author manuscripts in preparation, with the honours, hackathons and teaching around them.

In preparation·Target: NAACL 2027 LangCode Workshop

SecureSphere-LLM: An Identity-Aware Interception Proxy with Programmable Guardrails for Governing AI Coding Assistants Inside the IDE

S. M. M. Abidi et al.

Governs 4 AI coding assistants inside VS Code through 8 programmable guardrails, and blocks 31 of 31 attack prompts in a labelled 78 prompt corpus at a 66 ms median, with fail closed enforcement and per request audit trails.

See the project
In preparation·Target: ACM RecSys 2027

NexusHire: A Candidate-Side Decision-Support Layer over a Five-Stage Job Recommendation Cascade

S. M. M. Abidi, A. Shah and S. M. Haider

A five stage cascade over 123K job postings with counterfactual skill gaps and an independent salary verdict. Cold start users stay within about 35% of warm users on Precision@10 instead of collapsing.

See the project
Honours and hackathons
  • AgentX Hackathon, top 5 teamsAstera Software, Westlake Village, California, USA
  • Nebius x NVIDIA Global AI Hackathon, 2026Personal AI track: built JComit, a private finance agent that redacts data locally and runs NVIDIA Nemotron inside a policy locked sandbox (team of 4)
  • Best Final Year Project nomination, 2026SecureSphere-LLM, nominated by an external jury from Turing
  • International Computer Science Competition, 2026Special honour in the qualification round, qualified for the final
  • Data Structures and AlgorithmsPerfect 50 out of 50 on the final examination
Upcoming. Nov to Dec 2026: International Machine Learning Challenge, CALICO Cup (UC Berkeley), MIT Informatics Cup and NASA Space Apps Challenge.
Teaching at FAST NUCES
  • Computer NetworksTeaching AssistantSpring 2026
  • Theory of AutomataTeaching AssistantFall 2025
  • Discrete StructuresTeaching AssistantFall 2025
  • Data Structures LabStudent Lab AssistantFall 2024
  • Computer Organization and Assembly LanguageTeaching AssistantSpring 2024

Classes averaged 61+ students.

04Skills

Skills

01Agentic AI
11 tools

Agents, retrieval and orchestration

LangChainLangGraphRAGMCPTemporaln8nLLM routingAgent memoryOpenAI APIAnthropic ClaudeHugging Face
02LLM security
10 tools

Guardrails and safe AI in production

NeMo GuardrailsPresidioPrompt injection defenceSecure tool usePolicy enforcementPII redactionTopical safetyInput and output filteringAudit loggingMicrosoft Purview
03Machine learning
7 tools

Modelling, evaluation and explainability

PyTorchTensorFlowKerasscikit-learnXGBoostSHAPLIME
04Backend and data
11 tools

APIs, databases and vector search

PythonSQLFastAPIDjangoFlaskPostgreSQLpgvectorRow level securityMongoDBPineconeFAISS
05Cloud and DevOps
9 tools

Shipping and running it

Azure Blob StorageApplication InsightsAzure AI FoundryMicrosoft FabricDockerGitHub ActionsJenkinsCI/CDGit
06Frontend and visualisation
7 tools

Interfaces and dashboards

Next.jsReactTypeScriptStreamlitMatplotlibPlotlySeaborn
05Education

Education

2022 to 2026

BS Computer Science

FAST NUCES, KarachiCGPA 3.85 / 4.0
Honours
  • Four time gold medalist
  • Rector's honor list, four times
  • Dean's honor list, three times
Relevant coursework
Deep LearningArtificial IntelligenceRecommender SystemsGenerative AIProbability and Statistics
Certifications
  • Supervised Machine Learning (99.66%)Stanford University
  • Elevating LLM Safety with NVIDIA NeMo GuardrailsCodeSignal
  • Convolutional Neural Networks (98.87%)DeepLearning.AI
  • Build AI Agents using MCPIBM
  • Copilot Studio Agent AcademyGlobal AI Community
  • Microsoft Applied Skills (2)Microsoft Learn
  • Developing LLM Applications with LangChainDataCamp
  • Large Language Models ConceptsDataCamp
Leadership
  • President · IEEE FAST NUCES Student Branch2025 to 2026

    Led 9 workshops and the IEEEXtreme 24 hour qualifier for 100+ participants. Taught a hands on NVIDIA NeMo Guardrails workshop to 160+ participants on IEEE Day 2025, and hosted a backend engineering session with Securiti AI for 110+.

  • Deputy, Data Science and ML · Google Developer Groups on Campus (GDSC)2024 to 2025

    Led 10+ workshops, hackathons and seminars engaging 500+ students.

06Contact

Contact

Let us build something
connected.

The fastest way to reach me is email. I reply within a working day.

Based in
Karachi, Pakistan
Education
BS Computer Science, CGPA 3.85, gold medalist
Focus
LLM security, AI safety, agentic AI
Leadership
President, IEEE student branch