I build voice agents and the infrastructure behind them.

1. 01 Open source

   Voice Agent Starter Kit Lythe Labs

   A complete real-time voice loop, microphone to spoken reply, with the latency work already done and the ASR, LLM and TTS providers left swappable.

   Role

   Designed and shipped it. Pipeline, transport, adapters, telemetry.

   Problem

   A phone conversation tolerates well under a second of silence before it stops feeling like a conversation, and real callers interrupt. Every new voice project was rebuilding the same latency-critical scaffolding to get there: audio transport, turn detection, barge-in, streaming across three providers that each want to be awaited differently.

   Approach

   Stream everything. Audio moves over a WebSocket, the LLM emits text deltas, and TTS chunks start playing before the sentence is finished, so time-to-first-audio stops depending on total response length. Client-side VAD decides when the caller has actually stopped rather than waiting on a fixed timeout, and doubles as barge-in: the moment the caller starts speaking over the agent, queued audio is cut and playback stops, so the agent yields the turn instead of talking over the person it is meant to be listening to. The whole path stays async so no provider call can block another, and every stage is instrumented so a regression shows up as a number instead of a complaint.

   Outcome

   Open-sourced and Dockerised. New voice projects start from a working loop with a latency summary endpoint, instead of from an empty main.py.

   • FastAPI
   • WebSockets
   • Deepgram
   • Groq
   • Cartesia
   • VAD + barge-in
   • asyncio
   • Docker

   Source
2. 02 On PyPI

   Relay Arclat

   An MCP server that audits Python voice-agent codebases for the async mistakes that quietly add latency to a phone call.

   Role

   Author. AST analysis, MCP server, packaging and release.

   Problem

   Voice agents live or die on response time, and the worst offenders are invisible in review: one blocking call inside an async path, awaits that should have run in parallel, a tool chain doing three round trips where one would do.

   Approach

   Static analysis at the AST level rather than runtime profiling, exposed as an MCP server so the audit runs inside the assistant's own tool loop, so the agent can inspect the codebase it's working on.

   Outcome

   Published to PyPI as relay-arclat. Latency review becomes an automated pass instead of a careful read-through.

   • Python
   • MCP
   • AST analysis
   • asyncio
   • PyPI

   Source relay-arclat
3. 03 Open source

   Crucible

   A multi-agent framework that runs entirely on your machine: hybrid retrieval over SQLite, local models through Ollama, zero core dependencies.

   Role

   Author. Framework design, memory layer, retrieval.

   Problem

   Most agent frameworks assume a hosted vector database and a deep dependency tree. That's fine for a demo and painful for anything that needs to be reproducible, private, or still working in six months.

   Approach

   Memory in plain SQLite with hybrid BM25 + embedding retrieval, so keyword and semantic matches both land; local model support via Ollama; and a hard rule of no core dependencies.

   Outcome

   Agent workflows that are self-hosted end to end and reproducible. No external services to sign up for, nothing to drift out from under the run.

   • Python
   • SQLite
   • BM25 + embeddings
   • Ollama
   • Multi-agent

   Source
4. 04 Open source

   Agent Testing Suite Lythe Labs

   An evaluation dashboard that makes multi-agent pipelines legible: where the time goes, where the money goes, which tools actually get called.

   Role

   Built at Lythe Labs. Instrumentation and dashboard.

   Problem

   Multi-agent systems are opaque. When a pipeline is slow or expensive, there's rarely evidence pointing at which agent or tool call is responsible, so tuning turns into guesswork.

   Approach

   Instrument every run for latency, API cost and tool-call frequency, then render the pipeline as an interactive flow so a bottleneck is something you can see rather than suspect.

   Outcome

   Agent tuning becomes data-driven: regressions surface as numbers, and optimisation targets the step that's actually costing.

   • Python
   • LangChain
   • Streamlit
   • Observability

   Source

Also built

• ai-cache

  Open source

  Automatic caching of LLM responses across OpenAI, Anthropic and Gemini behind one interface. Repeat calls resolve locally, cutting spend and latency on repeated NLP queries.

  • Python
  • OpenAI
  • Anthropic
  • Gemini
  Source

• Long-Running Agent

  Open source

  Autonomous agent built for weeks of continuous runtime: LLM planning loops, sandboxed code execution, persistent memory and resumable sessions that survive a restart.

  • Python
  • LangChain
  • FastAPI
  Source

• mcpie

  Open source

  Security scanner for MCP servers, in the shape of npm audit. Scans the servers you install and the ones you build for tool poisoning, over-broad permissions, credential exposure and rug pulls. Zero runtime dependencies.

  • TypeScript
  • MCP
  • Static analysis
  Source

1. Mar 2026 - PresentSan Francisco · Remote

   Member of Technical Staff Lythe Labs

   Building and shipping production voice agents, multi-agent systems, and the infrastructure that runs them: system design through deployment, solo, across a distributed remote team.

   • Designed and shipped the open-source Voice Agent Starter Kit: a streaming ASR → LLM → TTS loop over WebSockets with client-side VAD, barge-in and built-in latency telemetry, cutting voice agent integration time on new projects.
   • Built real-time voice pipelines and tool-calling agent infrastructure against a sub-800 ms time-to-first-audio budget, keeping the whole path async so no provider call blocks another.
   • Engineered a long-running autonomous agent supporting weeks of continuous runtime, using LLM planning, sandboxed execution, persistent memory, resumable sessions.
   • Own CI/CD infrastructure with automated test integration, the AWS production stack, and MCP server development end-to-end.
   • Maintain test plans for voice pipelines and agent systems; triage failures across telephony tooling and LLM integrations via Jira and Linear.

   • Voice agents
   • LangChain
   • FastAPI
   • AWS
   • CI/CD
   • MCP
2. Jan - Mar 2026Georgia, USA · Remote

   AI Intern School of AI

   Delivered 25+ AI agent systems independently across healthcare, finance, cybersecurity and education, each end-to-end with minimal supervision.

   • Built full-stack multi-agent platforms with persistent memory, voice control, workflow automation and document analysis, cutting manual document processing time by an estimated 70% in pilot tests.
   • Implemented agent-to-agent communication, context-aware task coordination and structured LLM outputs in LangChain.
   • Shipped a medical note structurer (100% of unformatted clinical notes converted to standard templates), an earnings call analyzer, an incident response assistant, an AI tutoring platform and a collaborative document analyzer.
   • Reported system designs and progress weekly to cross-functional leads across three time zones.

   • LangChain
   • Multi-agent
   • Python
3. Nov 2025 - Jan 2026Lahore, Pakistan

   AI Intern AABS

   Built and optimised production RAG pipelines on Groq and ChromaDB, holding low latency at scale without losing retrieval accuracy.

   • Improved retrieval precision by 40% through embedding pipeline tuning and relevance optimisation.
   • Developed PDF chatbots and memory-augmented conversational agents; integrated PostgreSQL for session and metadata storage in multi-agent MLOps workflows.
   • Applied user-intent classification and multi-agent orchestration to route queries accurately, reducing irrelevant LLM responses.

   • Groq
   • ChromaDB
   • PostgreSQL
   • RAG
4. Sep 2025Cairo, Egypt · Remote

   Machine Learning Intern Elevvo Pathways

   Six distinct ML projects delivered independently in one month, each through the full lifecycle: cleaning, feature engineering, training, evaluation and visualisation.

   • Linear and logistic regression, K-Means clustering, CNNs, XGBoost, and a collaborative-filtering recommender.

   • Scikit-learn
   • XGBoost
   • TensorFlow
5. Aug - Sep 2025Dubai, UAE · Remote

   Data Analyst Intern Excelerate

   Led a team of four interns analysing social media marketing data.

   • Identified three underperforming ad campaigns and presented cost-reduction recommendations that the finance team adopted.
   • Built interactive Looker Studio dashboards on Pandas and NumPy pipelines, with SQL extracting and aggregating campaign data across sources.

   • SQL
   • Pandas
   • Looker Studio
6. Sep - Nov 2024Islamabad, Pakistan

   Data Science Intern Securely Innovations

   Data work and QA for a security analytics team.

   • Preprocessed large structured and unstructured datasets in Pandas, NumPy and SQL; automated reporting workflows that cut manual processing time by 30%.
   • Applied statistical modelling and data mining to surface actionable patterns for analysts.
   • Ran functional, regression and mobile-responsiveness QA across the company's main web platform, tracking defects in Jira.

   • Pandas
   • SQL
   • QA
   • Jira

• 

  AI & machine learning

  What I learned about RAG models as an AI intern

  Practical lessons from implementing retrieval-augmented generation in a startup: how retrieval quality, preprocessing and system design decide whether an LLM app is reliable.

  Read on Medium
• 

  Web scraping

  My journey into ethical web scraping with Python

  Respecting robots.txt, rate limiting and the legal considerations that separate responsible data collection from the kind that gets you blocked.

  Read on Medium