Background
Afraz Khan is the founder and lead AI engineer at EpochC, an AI development agency building systems that run in production rather than demos that impress in a meeting. He has delivered more than twenty contracts at a 5.0 average client rating, working across healthcare, financial services and document automation.
His work is the work described across this site. He built the LangGraph orchestrator that unified four clinical sub-agents behind a single conversational API for OphyCare, removing roughly 80% of the code in the process and carrying zero citation hallucinations. He built the KYC pipeline for Ace Money Transfer that pairs OCR with facial recognition and removed €40,000 a year of manual review at 98% field-detection accuracy, on 70% less compute than the baseline. He built the graph-plus-vector retrieval platform for EdgeCIO that lifted query accuracy 40% and answer relevance 35% while holding end-to-end latency under 200ms. And he built the AI medical scribe that turns a consultation into more than ten completed mandatory clinical forms.
He writes the engineering posts on this site. Where a post states an accuracy figure, a latency number or a cost saving, it comes from a system he shipped and can describe in detail — not from a benchmark or a vendor claim.
His practical bias is toward systems that know what they do not know. Confidence scoring, grounded retrieval and human review appear in nearly everything he builds, because in the domains he works in a confident wrong answer costs more than no answer at all.
What he builds
- LLM agents & multi-agent orchestration
- Retrieval-augmented generation (RAG)
- Knowledge graphs & hybrid retrieval
- OCR & intelligent document processing
- Clinical documentation & AI scribing
- KYC & identity verification automation
- Speech-to-text & audio ML
- Production ML systems, observability & CI/CD
Stack
- Python
- LangGraph
- LangChain
- GPT-4o
- RAG
- pgvector
- Neo4j
- MongoDB
- Redis
- PaddleOCR
- YOLOv8
- Deepgram
- FastAPI
- gRPC
- PostgreSQL
- Docker
- AWS
- GitHub Actions
- Langfuse
Systems delivered
-
OphyCare · Healthcare
One conversational API over four agents
Four clinical sub-agents behind one conversational API
-
Ace Money Transfer · FinTech
KYC onboarding, fully automated
€40,000/year removed from KYC review
-
EdgeCIO · FinTech
Graph + vector retrieval platform
+40% query accuracy on hybrid retrieval
-
OphyCare · Healthcare
AI medical scribe & form autofill
10+ clinical forms auto-filled per consultation
Client feedback
- Committed to Quality
- Clear Communicator
- Detail Oriented
- Collaborative
- Accountable for Outcomes
Selected by clients on Upwork contracts across 4+ delivered systems, at a 5.0 average rating over 20+ contracts.
Articles by Afraz Khan 55
- AI automation agency: what they actually do, and how to pick one The category runs from three-person Zapier shops to engineering teams shipping production systems. How to tell them apart before you sign anything.
- AI document management workflow: from filing cabinet to acted-on data A document management system stores and routes. AI changes what the system can understand about a document, which changes which workflows are worth automating.
- DAX Copilot explained: what it is, what it does, and the rename to Dragon Copilot Nuance DAX Copilot became Dragon Copilot in March 2025. What the Dragon Ambient eXperience does, how it differs from DAX dictation, and where it stops.
- Epic AI scribe integration: what it actually takes to write into the chart Getting an ambient scribe writing into Epic is an integration project, not a model problem. The routes in, what each costs, and where builds stall.
- On-premise RAG: building retrieval that never leaves your network When documents cannot reach a third party, the architecture is decided before anything else. Self-hosted models, the hybrid middle ground, and what each costs.
- RAG consulting: what to ask for when your system already exists Most RAG that disappoints is failing at retrieval, not generation. What a short consulting engagement should establish before anyone proposes a rebuild.
- RAG for life sciences: provenance, versioning and validated retrieval Life-sciences corpora need per-claim provenance and reproducible answers. What changes when an answer may reach a regulatory submission.
- Text classification techniques for document AI: what still works Classical classifiers, fine-tuned transformers and zero-shot LLMs each win in different places. How to pick, and why classification beats extraction.
- Agentic RAG: self-RAG, corrective RAG and when the loop is worth it Agentic RAG makes retrieval part of the reasoning loop, not a fixed first step. The patterns that work, what they cost, and when plain retrieval wins.
- Multimodal RAG: retrieval over PDFs, tables and diagrams Text-only pipelines silently drop the tables and figures where the answer lives. How multimodal RAG indexes visual content, and when page images beat parsing.
- RAG chunking strategies: semantic, fixed and structure-aware compared Chunking sets the ceiling on retrieval quality, and most teams tune the model instead. Fixed, recursive, semantic and structure-aware splitting compared.
- RAG evaluation: metrics, Ragas and the golden set you need first Without an evaluation set every RAG change is an opinion. The metrics that separate retrieval failure from generation failure, and how to build the golden set.
- RAG vs CAG: when to stop retrieving and cache the whole corpus Cache-augmented generation preloads the corpus into context and skips retrieval. Where that wins on latency, where it breaks, and how to route between them.
- Agentic AI vs generative AI: the difference that actually matters Generative AI produces output. Agentic AI decides and acts across steps. The distinction changes what the engineering costs and what can go wrong.
- AI agent platforms vs building your own: the four real options Agent builders, SDK harnesses, managed agents and manual loops differ on two axes: who supplies the harness, and who supplies the deployment. How to pick.
- Buying business process automation: what to look for and what to avoid EY puts initial RPA failure at 30-50%, and 70% of programmes plateau below 50 bots. How to evaluate a provider and scope a programme that survives contact.
- Chatbot development frameworks: what to use in 2026 Eleven chatbot frameworks compared: Rasa, Dialogflow CX, Lex, Bot Framework, LangGraph, Semantic Kernel and more. Which family to pick, and when to skip both.
- Conversational AI design: the principles that survive production What conversational AI design means once the model writes the words: boundaries, grounding, repair paths and handoff. With a worked example of each.
- Customer service chatbots: containment, deflection and what actually resolves Vendors promise 50% deflection. Gartner puts RAG containment at 40-65% and median deflection at 22%. What the metrics mean and how to build for resolution.
- Document workflow automation: what it costs and what it actually saves Manual document handling costs $5-25 per document. Vendors quote 200-400% first-year ROI. How to build the business case from your numbers instead of theirs.
- Choosing an embedding model for RAG: what MTEB does not tell you MTEB is a prior, not an oracle. Why leaderboard rank rarely survives your corpus, and the four constraints that should decide your embedding model.
- Google Document AI vs a custom IDP pipeline: how to choose Google Document AI is $1.50 per 1,000 pages for OCR, $30 for custom extraction. Where managed processors win, where they stop, and what to build instead.
- IT process automation: which tickets actually automate, and which do not Service desk tickets cost $15-22 each. Gartner puts deflection at 20-30% against vendor claims of 50-75%. Which L1 categories genuinely automate.
- Medical dictation software: Dragon, ambient scribes and when to build Dragon Medical One runs $79-99 per user monthly plus implementation. Ambient scribes error at 1-3% against 7-11% for dictation ASR. Which one you actually need.
- No-code automation platforms vs custom builds: where the line actually falls Zapier bills per task, n8n per execution. A 10-step workflow at 10k runs is 100,000 tasks or 10,000 executions. Where custom builds actually cross over.
- Virtual scribe services vs AI scribes: cost, accuracy and when each wins Human virtual scribes run $1,200-4,000 per provider monthly. AI scribes run $39-700. Where the 10x gap is justified and where it is not.
- AI chatbot development: what it costs, how long it takes, and what to actually build A practical guide to AI chatbot development: the three shapes a chatbot can take, what drives cost, realistic timelines, and when custom beats a platform.
- What AI medical scribes actually cost, and when building your own gets cheaper What ambient AI scribes cost per clinician, why per-seat pricing compounds, and the headcount at which building your own becomes the cheaper option.
- Ambient clinical intelligence: what ambient AI scribes actually do Ambient clinical intelligence, ambient scribes and ambient clinical documentation explained: what passive capture solves, and where the difficulty really sits.
- Enterprise workflow automation: what breaks at scale and how to design for it Enterprise workflow automation is mostly reconciliation between systems that disagree. State modelling, idempotency, partial failure and exception lanes.
- Intelligent document automation: from extraction to the action that follows Intelligent document automation is the full loop: classify, extract, validate, then act. Why projects that stop at extraction just hand the business a CSV.
- RAG as a service vs building your own: which one you actually need When a hosted RAG product is the right call and when it stops being one. Permissions, retrieval logic, per-query pricing and the three constraints that decide.
- Can you dictate SOAP notes from an Apple Watch? What actually works for ambient capture Whether a smartwatch is a viable capture device for clinical dictation, what the pipeline needs to produce a usable SOAP note, and which setups hold up.
- HIPAA-compliant AI architecture: PHI boundaries, retention and audit trails Architecting an LLM system that handles PHI: where the boundary sits, what a BAA changes about model choice, and why vector stores break retention policy.
- LLM observability in production: traces, evals and regression gates What to instrument in an LLM system so you can debug it, prove it improved and catch regressions: span tracing, evaluation sets and CI regression gates.
- Production LangGraph agents: checkpointing, cost ceilings and deterministic fallbacks What separates a LangGraph agent that demos well from one that survives production: durable checkpointing, token ceilings, bounded recursion, fallbacks.
- Build vs buy an AI medical scribe: an honest comparison When a custom AI medical scribe beats Abridge, Nuance DAX or Suki, and when it does not. Cost, timelines, form coverage, EMR integration and residency.
- Custom KYC vs Sumsub, AU10TIX and Onfido: build vs buy compared When building your own KYC and identity verification beats a vendor, and when it does not. Per-check pricing, coverage, residency and thresholds compared.
- AI clinical documentation: how to automate notes without adding risk How AI-generated clinical notes work, what automating them actually saves, the grounding requirements, and how to evaluate a system before it touches a chart.
- Intelligent document processing: OCR, IDP and document intelligence explained What intelligent document processing involves, how IDP differs from OCR, where accuracy is really won, and how to scope a project that survives production.
- Enterprise RAG architecture: permissions, freshness and evaluation What separates an enterprise RAG deployment from a demo: entitlement-aware retrieval, incremental ingestion, hybrid search and an evaluation harness.
- What is an AI medical scribe? A clinician's guide A plain-language guide to AI medical scribes: what they are, how ambient scribing works, the benefits and limits, and how to choose one you can trust in real care.
- AI for medical notes: how it works, what it saves, and what to watch A guide to AI for medical notes: the pipeline stage by stage, what it saves clinicians, the risks to manage, and how to tell a real system from a demo.
- AI medical scribe notes: a real example (audio to SOAP note) What an AI medical scribe actually produces: a patient-doctor conversation turned into a structured SOAP note, and how speech maps to charting fields.
- How to reduce LLM costs in production Practical ways to cut the running cost of an LLM application without hurting quality: model routing, caching, retrieval and smaller models where they fit.
- LangGraph vs LangChain: which should you use? LangChain and LangGraph are not competitors. Here is how they fit together, and when you should reach for LangGraph's graph model over a simple LangChain chain.
- How to choose an AI development company A checklist for hiring an AI development partner: what to look for, the red flags that predict a failed project, and the questions that separate builders.
- Building an AI medical scribe: from audio to structured notes How to turn a patient-doctor conversation into accurate, structured clinical documentation with speech-to-text, entity extraction and confidence-scored form filling.
- Knowledge graphs vs vector search for RAG Vector search finds similar text. Knowledge graphs capture relationships. Here is when each wins, and why the best retrieval systems use both.
- What is agentic AI? A practical guide for businesses Agentic AI explained without the hype: what it is, how it differs from a chatbot, where it creates real value, and how to start safely.
- How much does it cost to build an AI agent? A practical breakdown of what actually drives the cost of building a production AI agent, why cheap prototypes fail, and how to budget for something that ships.
- RAG vs fine-tuning: which do you actually need? RAG and fine-tuning solve different problems. A decision framework for choosing between them, and why teams reach for fine-tuning when they need retrieval.
- Multi-agent systems with LangGraph: when you need them (and when you don't) Multi-agent architectures are powerful and easy to over-use. How to tell when a LangGraph multi-agent system is the right call, and how to build one.
- OCR for KYC: building document verification that actually works What it takes to build automated identity verification with OCR and face matching, from field accuracy to edge deployment, on a system that saved EUR 40k.
- How to stop your RAG system from hallucinating Most RAG hallucinations are a retrieval problem, not a model problem. How to build retrieval that cites its sources and refuses to make things up.
Working on something like this?
Tell me the problem. I will tell you honestly whether it is worth building.