Karachi · AI systems engineer
I build the system around the model — not just the integration. The part that holds.
Multi-agent pipelines, context architecture, evaluation frameworks, and the full production backend. Four years shipping. Lately: a self-evolving MCP agent that learns from every failure and ships its own fixes.
A recorded lesson goes in. A score a school can defend comes out. Every finding is quoted and timestamped. A statistical audit — weighted rubric, standard deviation, z-score against that teacher's own history, Cronbach's alpha — gates the report before it reaches anyone's file. The model does not touch the math. English, Urdu, and Arabic.
Multi-tenant classroom platform — live class, attendance, gradebook, and parent portal. I own tenancy, auth, and the live-class path. The API decides who joins and issues the token. Self-hosted LiveKit carries the media. Scale does not become a per-minute managed-video bill.
I led the architecture. One project holds the PDFs, call recordings, images, and handwritten notes of a matter. Every answer keeps the source passage attached. Ambiguous sections are flagged for a person to resolve — not smoothed over by the model.
A self-evolving outreach agent with no LLM of its own — your coding agent is the brain, it is the hands. An MCP server with 31 tools that drives your logged-in Chrome to apply across four channels (email, LinkedIn Easy Apply, ATS forms, connection notes), verifies every action, and learns from each failure: it records the blocker with a screenshot, harvests answers you've already given, asks you once, remembers forever, and files its own fixes back to the repo. The codebase improves itself, session over session. Open source, MIT.
Open source, on PyPI. A local evaluation harness for speech-to-text — WER and semantic WER, because fluent text can still be factually wrong. Used as the accuracy gate for Xyova. A companion downloader pulls real classroom audio from TalkBank ClassBank for benchmarking.
The engineering contract my coding agents work inside. Evidence-backed workflows over MCP — for Claude Code, Codex, Grok, and Kiro. Agents implement. I own the architecture and the review. The contract is the guardrail.
An AI content engine built as an agent-native tool — research, write, anti-AI filter, publish. Works inside Claude Code, Codex, Cursor, and Kiro over MCP. The pipeline learns your voice, blocks 50+ robotic patterns automatically, and adapts one post across LinkedIn, X, and YouTube. Zero code to run it.
A memory aide on Telegram. Markdown files are the source of truth. The model is a window into them. It cannot send mail, book a calendar, browse, or open a shell — those tools were never built. The refusal is the architecture.
Three layers, and they do not do each other's job. The model reads and cites. A separate audit measures and can force human review. A person sets the standard. I do not ship a system where the output reaches anyone unsupervised.
Context is not free. I scope what each agent can read — the right portion, the relevant history, nothing that adds tokens without adding signal. Retrieval layers filter and rank before the reasoning layer sees anything.
Transcribe Studio scores transcripts with WER and semantic WER. A grammatically perfect output that contradicts the source still fails. I apply the same principle to every pipeline: the eval has to catch the failure mode that looks like success.
Checkpointed state means a pipeline can stop and resume. Human escalation is a designed output, not a fallback. Tools the agent does not need are never built. The refusal is the design, not an afterthought.
S3 multipart with chunk-level recovery. Liveness is real data progress — not a heartbeat, because a heartbeat cannot arrive during the outage you are recovering from. A partial file is still processed. I wrote this up publicly as the decision not to reach for Kinesis or WebSockets.
Self-hosted LiveKit on Mishkat VLE. The API authorizes and issues tokens. The SFU carries the media. Authorization never moves onto the media server. Chosen because school-scale video should not be an open meter.
Scope is enforced where rows are read and written — not in a hidden button. A front-desk account must fail to write a lead-only discount when calling the API directly. Handlers stay thin. Services own decisions.
PRs run format, lint, and types. OIDC short-lived credentials, no long-lived cloud keys in the pipeline. The box installs, starts, and hits a health check. A failed check rolls the release back automatically.
Four years across these tools in production. Click a category to see what I reach for.
Spec first. Agents implement. I own the architecture.
The spec is written before the first line of code. Coding agents — Claude Code, Codex, Kiro — work inside a written contract that defines what they can touch and what they cannot. I review every change. I own the decisions that survive the demo.
On pipelines: context is scoped before the model sees it. Evaluation is built before the pipeline is trusted. Failure modes are designed for, not discovered after launch. A 48-hour proof of concept is what won a major client — because the architecture held under real conditions, not just demo conditions.
Crash-safe live upload without Kinesis or WebSockets. A public API on FastAPI and AWS. Session validation. Thinking in public, with the code on GitHub.
The layout I reuse: router, service, crud, model. HTTP authenticates and translates. Business rules do not live in the handler. Scoped queries throughout.
Pulls TalkBank ClassBank English classroom audio and CHAT transcripts so a speech benchmark has something real to fail against.
Contact