MUHAMMAD JEHANZAIB

Lead engineer on a five-service platform that automates IRC §41 R&D tax-credit qualification for a U.S. tax-advisory firm: an LLM contract-analysis engine, a FastAPI gateway, a React admin console, a tax-form extraction engine, and a §41 calculation and workpaper engine. Responsible for security, concurrency, and data-integrity hardening across all five services, including client data isolation.

• Architected the contract-analysis engine (Python, LangGraph, Celery, Redis) as a multi-agentic workflow in which specialized stages (triage, rights assessment, contingency, jurisdiction, adversarial review, IRS counter-argument, final determination) each evaluate one aspect of a contract against IRC §41 and §6662 standards. Each stage receives only the clauses relevant to it, and independent stages execute in parallel to reduce analysis time.
• Converted the model's raw confidence output into an auditable §6662 Tax Position score suitable for reviewer defense, with an automated approval gate that clears unambiguous cases and routes the remainder for human review.
• Built a streaming multi-LLM chat interface for reviewers, supporting model switching mid-analysis, clean cancellation of long-running jobs, and safe correction of results with protection against prompt injection.
• Designed server-authoritative resumable sessions so that long-running analyses persist across page reloads, navigation, and tab switches, with automatic repair of corrupted state and strict isolation between clients.
• Delivered a multi-tenant automation platform in which system events trigger configurable actions (for example, provisioning Google Drive folder structures) through a reliable event pipeline, with a no-code workflow editor and dashboards for runs and events.
• Built a tax-form extraction engine (AWS Textract with Claude post-processing) covering 29 IRS form types (W-2/W-3, 1099, K-1, 1120/1120-S/1065, 6765, 2553, state forms), with per-form document segmentation and LLM-based form classification.

Designed and built end to end "William," a multi-tenant agentic AI SaaS with real-time voice for a telecom client (World Mobile): users talk or type to an assistant that answers from the company knowledge base, acts in email and calendar, remembers earlier conversations, and turns a voice call into a project that a team of AI agents completes live.

• Architected the platform as six services around one agent brain: a FastAPI gateway through which every chat and voice request passes (Redis-backed sessions and history, streamed responses, LangGraph onboarding flows with pauses for user input), a standalone LangGraph graph server so the agent ships independently of the API, a LiveKit voice worker, a pgvector RAG service, a multi-agent orchestrator, and a Next.js studio.
• Built the LangGraph agent: it classifies each message, directs email and calendar requests to dedicated orchestrators, and routes all other requests to knowledge-base search, cross-session memory recall (with safeguards against fabricated recall), or a tool-enabled responder (knowledge search, web search, time, task creation), streaming tokens as they are generated.
• Built a multi-agent orchestrator that runs every task extracted from a call through five specialised agents (research, plan, build, review, write-up), with live WebSocket progress, per-agent output and token cost, instruction-driven re-runs, and downloadable reports.
• Built the real-time voice agent on LiveKit (Deepgram speech-to-text, Cartesia text-to-speech, Silero voice activity detection, noise cancellation) on the same agent core as chat, and patched the LiveKit LangChain adapter so that the agent, rather than the plugin, governs spoken output. Added a fast retrieval route with an instant spoken filler and warm workers so the agent joins a call in about 10 seconds; each call's transcript becomes a session of linked tasks.
• Built the knowledge and model layer: pgvector RAG with neighbouring-passage context and multilingual retrieval, every model routed through OpenRouter with automatic fallbacks, and token streaming fixed for LangGraph on Python 3.10, cutting the first reply from 35 seconds to under 10.
• Designed and implemented multi-tenancy: organization-level data separation in Redis, role-based permissions, and tool access restricted to each user's entitlements. Authored the architecture plans adopted by the team for serving multiple clients from a single codebase.
• Shipped to GCP through GitHub Actions and later as 8 services on Railway with auto-deploy from main and Playwright-verified releases, with per-provider token and latency tracking.

Founding engineer on OptiML, an open-source framework for LLM compression and fine-tuning that produces smaller, lower-cost models for production. Responsible for architecture, CI/CD, release management, and a team of 4.

• Designed a single config-driven API covering pruning (magnitude, Wanda, SparseGPT), quantization, LoRA fine-tuning, and MLA, including hardware-friendly 2:4 structured sparsity, with integrated lm-eval benchmarking.
• Scaled training and compression across multi-GPU NVIDIA and Intel environments with PyTorch DDP.
• Built the GPU infrastructure: cost-capped Vast.ai GPUs provisioned into a K3s cluster over Tailscale, Ray on AWS EC2, a FastAPI job API with per-user quotas, and self-hosted inference through Text Generation Inference.
• Implemented AWS SageMaker JumpStart pipelines and GitHub Actions CI/CD; managed sprints and releases.

Built the AI modules of "Ruyi," a robot that supports the recovery of elderly and post-surgery patients using personalized longitudinal data, together with a medical data bank for clinicians integrated with social-media intelligence.

• Developed the ADL (Activities of Daily Living) module, which recognizes patient activities and measures the effort and duration of each movement, running at 30 FPS on the device.
• Built an audio module recognizing 9 health-critical sounds, and a vision pipeline that converts robot sensor data into a structured patient record for recovery planning.
• Co-developed a gait analysis module tracking 15 clinical parameters, from step length and knee flexion to cadence.
• Deployed across NVIDIA Xavier edge devices, Microsoft Azure, and AWS.
• Built data scrapers, cleaning routines, and a generative-AI social-media intelligence layer supporting medical-data explainability.