Muhammad Sohail

The agent needed to hold context across a multi-turn sales conversation while staying grounded in real technical documentation: no room for hallucinated specs on a home battery install. I built session state in PostgreSQL so the agent could reference earlier turns, and layered a RAG pipeline over ingested manuals and pricing sheets so every answer traced back to source text. On sizing, it pulls live 3-phase load data from HomeWizard telemetry and runs it through hard-coded installation constraints: 5×4mm² cable cross-section requirements, 70kg-per-unit floor loading limits, so a recommendation is only ever one the install can actually support.

Built during the early months of COVID-19, when clinics were overwhelmed and many patients couldn't safely see a doctor in person. Patients complete a WhatsApp-based survey; the system scores it for risk and randomly assigns high-risk cases to one of a rotating pool of volunteer doctors for remote follow-up, spreading the caseload fairly instead of funneling everyone to the same few volunteers.

The original workflow was fully manual: someone reading a scanned pamphlet and retyping it into a form. I replaced that with a dual-view interface, the pamphlet image on one side, the OCR- and OpenAI-extracted fields pre-filled on the other, so an operator's job became confirm-and-submit instead of transcribe-from-scratch. That cut per-record handling time dramatically across a backlog of 50k+ pamphlets.

Streaming live call audio to Watson STT over WebSockets meant transcription had to keep pace with the call in real time, not batch-process it afterward. On top of the transcript stream, I added a lightweight rule layer that flags moments where a caller requests a passcode or other restricted data mid-call: useful for a compliance team to review, on calls worth a second look rather than buried in a full transcript archive.

Post-surgical infection tracking was being done without a shared system between hospitals. HCAI gave clinicians a Flutter app to log LAARC-standard lab evaluations directly from the ward, feeding a Vue.js dashboard that WHO and NIH collaborators used for performance reporting. I worked as solutions engineer on the architecture: a microservices split between the mobile intake app and the reporting backend, with the data model designed specifically around the surveillance protocols the WHO/NIH team needed, not a generic form builder retrofitted to fit.

TRA needed one real-time data layer under a handful of separate apps and external booking systems, all touching the same household and lead records. I co-architected the TRA API (Node.js/TypeScript, REST) and designed the TRADB MongoDB schema to hold that data at scale, then built the Manifest System on top to track bookings, agent performance, commissions, and where each lead sits in the sales lifecycle.

The hardest part wasn't the API, it was the data-merging logic: households and leads get created independently across systems and have to be reconciled without losing history, so I wrote extensive Mocha/Chai test suites covering 20+ edge cases (duplicate resolution, membership inheritance, primary vs. secondary contact filtering on meeting payloads) before any of it touched production. A message broker absorbs the high-volume call traffic into TRADB, WebSockets push real-time reservation updates, Elastic indices stay in sync for search, and re-tuning the MongoDB cluster configuration saved roughly $12,000/year in infra cost without touching performance.

The system needed duplicate-incident detection without slowing down the main API, so I split it into two Quarkus services: the core API stays synchronous (CRUD, JWT auth, hexagonal ports/adapters so persistence and messaging never leak into the domain layer), while an incident-created Kafka event triggers a separate analyzer service that runs Postgres pg_trgm similarity matching asynchronously and reports back over Kafka. Everything ships in Docker Compose locally and a Kubernetes manifest for the cluster.

For my final-year project I stood up Dionaea, Suricata, and Kippo honeypots to capture live attack traffic, then centralized everything in MongoDB so it could be queried and visualized as a single dataset rather than three disconnected log formats. The output maps attack intensity geographically, letting you see attack sourcing patterns at a glance instead of scrolling raw logs.