Muhammad Jehanzaib

Agentic platform that determines whether research performed under a client's contracts qualifies for IRC §41 R&D credits — automating the §41(d)(4)(H) "funded research" analysis that normally takes hours of manual legal review per contract. A multi-agentic LangGraph pipeline (document triage, rights-gate, contingency/financial-risk, jurisdiction & state-conformity, web search, adversarial self-review, IRS counter-argument simulation) yields a qualified / not-qualified determination with confidence scoring, evidence citations, and per-prong contract references. Bulk upload with native + vision-OCR text extraction, deterministic file bundling, full batch orchestration with cancellation, and an agentic chat that feeds a correction-driven re-run — grounded by a case-law & state-conformity knowledge base. Built at Ventive for a U.S. tax-advisory firm.

Form-aware extraction pipeline for payroll/wage documents (W-2, W-3, census) that routes each page chunk to a dedicated per-form handler, then verifies with dual-model consensus (Gemini 2.5 Flash + Grok 4.1 Fast) and a Claude verify pass, with an AWS Textract path for scanned docs. Handles the messy realities of real packets — multi-employer batches, multi-page/2-up/4-up forms, and per-state W-2 splits. An automated W-2 ↔ W-3 reconciliation engine keyed per (EIN, tax year) cross-checks employee count, federal Box 1, and state Box 16 wages with cents-level tolerance, emitting PASS / FAIL / PARTIAL / UNVERIFIABLE verdicts plus a roster and human-readable diagnostics; structures extracted entities into a Neo4j knowledge graph. Built at Ventive.

A GPT-4.1 + LangGraph agentic system that takes a tax scenario, asks structured clarifying questions, then generates optimised strategies with simulated returns for each and a comparative trade-off analysis — integrated with tax-filing APIs for execution.

Bank-statement analysis pipeline with OCR + LLM extraction, ML transaction categorisation, anomaly detection, and advice generation, visualised through FastAPI and Plotly.

An AI tenant-communication layer on top of Buildium that autonomously handles tenant SMS — rent follow-ups, maintenance intake (auto-creates work orders), lease-renewal outreach with PDF generation, and smart escalation — with per-tenant conversation memory. Includes a grounded Portfolio Copilot, a Buildium-mirrored dashboard, a WhatsApp-style inbox, role-based admin, and dark mode.

AI layer connecting four platforms that do not talk to each other for a high-volume real-estate team: Gmail, Follow Up Boss (people), Monday.com (deals), and Slack. Reads the agent's inbox in real time through Gmail Pub/Sub push, runs Claude over every email to classify it, identify each person and their role (buyer, seller, agent, vendor, co-owner), extract purchase-agreement terms from attached PDFs, and keep contact tags, deal stages, and activity notes in sync across the CRM and the deal board, with an SLA state machine that escalates missed VIP messages. Thin fast-ack webhook edges, fat Celery workers: every LLM call and platform write runs as a task, with nightly reconciliation and an LLM-judged property matcher behind a confidence threshold. All client-specific behavior lives in one policy file, so the same codebase deploys to the next agent by reconfiguring. Validated by month-by-month dry-run replays of 500 to 800 real emails per month with per-email verdicts before any live write. ~1,150 tests.

Underwrites any property deal with a deterministic engine (cap rate, cash-on-cash, NOI, comps) plus an AI Underwriting Agent that streams a multi-strategy fit assessment (buy-and-hold / BRRRR / flip / wholesale) and a Negotiation Strategist that produces an offer ladder and scripts. Kanban pipeline, portfolio dashboard, side-by-side compare, CSV import.

Agentic system that ingests leases, amendments, prior true-ups, and current invoices; extracts and interprets clauses to validate line items; cross-checks against external data (weather, utility benchmarks — e.g. snow-removal fees vs. actual snowfall); performs year-over-year anomaly detection; and gives clause-cited justification for every decision, with a conversational agent to question and override findings. Automated ~90% of the manual audit process.

A founder-facing decision-support platform for Amazon-led D2C brands: ingests commerce data into a canonical warehouse, applies deterministic scoring plus AI workflows, and surfaces "what changed yesterday, what to worry about, what to do" via dashboard, a tool-calling chat agent, real-time alerts, and a daily Telegram digest. Lead-time-aware reorder recommendations, a True-Profit calculator, Buy-Box monitoring, and returns/refunds tracking.

A multi-agent system that classifies and parses retailer emails (order confirmation, shipping, returns), extracts Order ID / Tracking Number / SKU, and presents a unified end-to-end order-tracking view.

A multi-agent solution that analyses buyer-seller conversations for intent, interest, and temporal patterns to enable intelligent real-time matchmaking; containerised and deployed on AWS.

An OCR + LLM pipeline that extracts and structures product details from retail labels (including Icelandic text) in real time to power personalised recommendations and dynamic retail updates.

An AI-driven operations platform for moving companies: a lead pipeline with an event-triggered automation engine, an AI Move Estimator that streams reasoning and returns structured inventory/volume/crew/price, lead-intelligence scoring, jobs & scheduling, invoicing, and a public quote flow. Full light/dark theming and role-based access.

The problem. A consultancy in the spinal biologics market answered client questions such as "which of these 49 products are cleared as 510(k), PMA, or 361 HCT/P, and since when?" by reading FDA records, trial registries, journals, and SEC filings by hand. Days per question, and no way to show where each fact came from.

What I built, as sole engineer from first commit to production handover. A research assistant that answers such questions in about 2 minutes for about $1 of model spend, with every sentence tied to a passage the reviewer can open. The ingestion pipeline keeps a library current from five public sources (FDA 510(k)/PMA/openFDA, ClinicalTrials.gov, PubMed and PMC, SEC EDGAR, DailyMed) plus uploads: PDF parsing with GROBID, chunking, embedding, and product and company entity resolution across 33,800 documents, 668,000 passages, and 711 tracked products. Retrieval runs three lanes (vector, keyword, and sentence windows that surface one decisive sentence buried in boilerplate) merged by rank fusion over a pgvectorscale DiskANN filtered index. A five-stage LangGraph workflow then plans the searches, has a judge model read every candidate passage and keep only what supports an answer, writes, and checks each sentence against its citation before the reader sees it, refusing when the corpus does not hold the answer.

Results. A hard question went from 6 minutes and $4 to 2 minutes and $1 after retrieval and judge optimisation (filtered index, judging during retrieval, a small judge model on a fast host with fallback). On the client acceptance set, every regulatory number in the answers was verified correct against the FDA record. Each answer ships with a live progress view, a source panel per citation, an audit trail of what was searched and kept, and its cost. Delivered with an admin console (product registry, source schedules, user approval, per-stage model and reasoning settings, cost reporting), scheduled backups, and a full operations runbook, administrator guide, illustrated user guide, and cost sheet for handover.

A multi-portal platform connecting the whole care loop across six role-scoped dashboards — booking → visit → e-prescription → pharmacy dispensing → insurance adjudication → financing — with AI clinical support (symptom triage, a streaming Care Navigator, clinical-note drafting via a tool-calling agent) and in-browser WebRTC telehealth. 46 API endpoints, 24 role-scoped screens.

Depth-camera keypoint detection and pose estimation for real-time gait analysis across 15 clinical parameters (step length, knee-flexion, cadence), with historical comparison and personalised suggestions for rehab professionals and athletic trainers.

A lightweight exercise-recognition and rep-counting system from video with real-time form feedback and personalised workout recommendations, deployable on ARM64 edge devices.

Real-time recognition of Activities of Daily Living and risky transitions (sit-to-stand, lying-to-sit) to flag unsafe movements and trigger timely interventions in elderly-care facilities and home monitoring — reducing fall risk and improving caregiver response.

Pose-estimation detection of risky movements (improper knee angle, hip flexion) for hip/knee-surgery patients, integrated with healthcare systems for predictive fall-risk forecasting, personalised prevention plans, and real-time alerts.

A PoseC3D-based system that identifies agitation-related behaviours with real-time monitoring, historical trend analysis, and personalised care suggestions — reducing caregiver burden and response time.

DeepFace deployed directly on Jetson Xavier for on-device facial recognition, mapping patient activities to identities for secure, low-latency, individualised behaviour tracking in care facilities.

A BLIP-based scene-description module that captions images from care environments in real time to support monitoring and documentation; Dockerised and deployed on AWS.

A planner-based chatbot that integrates vital signs, medical history, and real-time voice + facial emotion recognition, with patient-persona profiling to generate empathetic, personalised responses.

An LLM-powered chatbot that generates concise, relevant summaries of patient medical histories for clinicians, with modular integration into existing patient-data sources.

Scraped large medical datasets and built ETL pipelines for clinical-trial extraction, expert identification, and social-media sentiment analysis on drug trials/conferences, visualised in Power BI via a full-stack app with auth on AWS.

One lead-qualification engine behind three doorways — website chat (GoHighLevel Conversation AI), a web and phone voice agent (Vapi), and web forms — that answers within seconds, any hour. Each lead is scored by Claude against rules the business owns in Airtable (weights, threshold, disqualifiers — change a cell, behaviour changes, no deploy), enriched from the lead's own website, and the assistant books a discovery call on the live calendar mid-conversation. Contact, scored opportunity, note and follow-up task land in whichever CRM the client already uses — GoHighLevel or HubSpot, switched by one field. Engineered for production: idempotent intake with a replay guard, a dead-letter path where every failure answers 500 rather than a silent 200, per-contact opportunity upsert, an outbound Google Maps prospecting lane, a nightly metrics roll-up, and a client-facing Lead Desk dashboard. Chat-to-booked-call in about a minute; form-to-CRM in ~20 seconds.

An outreach pipeline that uses LinkedIn data (via ProxyCurl), generates personalised emails with GPT (content in S3), and tracks emails and history in DynamoDB.

Designed and built end to end a white-labelled, multi-tenant AI SaaS for a telecom operator (World Mobile). Users talk or type to "William," which 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 six services around one agent brain: a FastAPI gateway that every chat and voice request passes through (Redis sessions and history, SSE streaming, LangGraph onboarding flows with human-in-the-loop pauses); a standalone LangGraph graph server so the agent ships independently of the API; a LiveKit voice worker; a pgvector RAG service with neighbouring-passage context; a multi-agent orchestrator; and a Next.js studio.

Agentic core: the LangGraph agent classifies every message and routes it to dedicated email and calendar orchestrators, knowledge-base search, cross-session memory recall with safeguards against fabricated recall, or a tool-calling responder (knowledge, web search, time, task creation), streaming tokens as they are generated, including the RAG answer stream relayed through the graph. Tasks extracted from each call run 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. Voice: built the real-time voice agent (LiveKit Agents, Deepgram speech-to-text, Cartesia text-to-speech, Silero VAD, noise cancellation) on the same brain as chat, patching the LangChain adapter so the graph controls what is spoken; clear product questions go straight to retrieval behind an instant spoken filler, warm workers let the agent join a call in about 10 s, and each call's transcript becomes a titled session of linked tasks.

Knowledge and models: pgvector RAG with neighbouring-passage context and multilingual retrieval; every model routed through OpenRouter with automatic fallbacks and per-provider token and latency tracking; token streaming fixed for LangGraph on Python 3.10, cutting the first chat reply from 35 s to under 10 s. Platform: designed and implemented the multi-tenancy model (organization-scoped state, role-to-scope RBAC, scope-filtered tool binding) and authored the architecture plans the team adopted for serving many clients from one codebase, so a new client is configuration rather than a fork. Shipped to GCP through GitHub Actions and later as 8 services on Railway with auto-deploy from main and Playwright-verified releases.

An end-to-end e-learning app with AI agents for content creation, automated marking, exam creation and feedback, and a content-aware chatbot; plus user management, AI games, quizzes, a practice trainer, and an admin dashboard.

A full-stack app that turns a single keyword (plus tone and length) into a publish-ready SEO blog post — meta descriptions, title suggestions, keyword analysis, structured outline — in under 60 seconds via a 4-step AI pipeline, with a content library, inline editing, DOCX export, and Google SERP preview.

An open-source framework that turns fine-tuning into a compression pipeline, so models come out smaller and ready for production. One config-driven API covers multiple techniques — pruning (magnitude, Wanda, SparseGPT), quantization, LoRA, and MLA — on a dynamic graph execution model that adapts to the target hardware. Scales across multi-GPU setups (NVIDIA and Intel) via PyTorch DDP, with SageMaker JumpStart pipelines and lm-eval benchmarking built in. Also built the supporting GPU infrastructure: a control plane that rents cost-capped Vast.ai GPUs into a K3s cluster (Ray, FastAPI job API, Tailscale networking) and serves the compressed models for self-hosted, optimised inference via Text Generation Inference. Led a team of 4 — architecture, CI/CD, and releases.

Onboarding operations delivered two ways from one codebase: a web console for people, and a Model Context Protocol (MCP) server that Claude Desktop, Claude in Slack, or any agent can call with a personal token. One instruction starts an employee onboarding — person record, task list from a template, document requests with upload links, account provisioning — and tasks that need paperwork stay closed until the document is actually accepted; client onboarding (contract, workspace, billing, security review, kickoff, go-live) runs on the same state machine. Policy questions are answered only from the company's own versioned handbook, with a citation for every claim and a refusal when the documents are silent. The design centre is 25 narrow tools behind a single guard that runs role check → idempotent replay → handler → audit write, on success and on failure, so the console, the in-app chat agent, and the MCP endpoint cannot change data without a record. Five roles with row-level scoping, restricted personal data behind a separate audited call that is never exposed to chat or MCP, and every model call priced at write time from a versioned price table with daily/weekly/monthly cost views. 24 tools, 3 resources, and 3 prompts over a hand-rolled JSON-RPC 2.0 MCP endpoint; the seed data is produced by driving the real tools, so the audit log is genuine.