Available now — AI/ML roles · Lahore / remote

Syed MuhammadMujtaba buildsgrounded AI.

AI engineer for LLM, agentic systems, RAG and real-time voice — self-correcting agents and production retrieval, shipped full-stack from the model to the interface for five companies.

See projects

Production RAG & agentic systemsReal-time voice AIEnterprise GenAI · shipped to customersFull-stack — MERN & FastAPI

systems

01 · How my RAG self-corrects

retrieve — hybrid dense + BM25

rerank — cross-encoder over Qdrant

grade — is each doc actually relevant?

↺ rewrite the query if it underperforms

fall back to live web search (Tavily)

✓answer only if it's grounded in sources

LangGraph state machinemax 3 loops

02 · What a voice turn costs

speech-to-textstreaming

LLM inferencefirst token

text-to-speechchunked

budget per turnunder 500 ms

Overlap the stages, or the human hears the seam.

03 · How I wire agent graphs

plannersplits the brief

researcherretrieval + web tools

writerdrafts to the brief's tone

criticrejects → sends back

assembleslides · curriculum · quiz

State, not prompt chainsretryable steps

04 · Ask a policy, get a citation

⌕

embed the question, not the keywords

FAISS over handbooks & compliance docs

answer carries the clause it came from

✓ cited, not guessed — plain language in, policy out

LangChain · FAISS · HuggingFace embeddings

Self-correcting RAG

Real-time Voice AI

MS AI · LUMS

The pitch

I don't ship AI that guesses. I build systems that grade their own answers, rewrite their own failures, and respond before you notice the pause — and I build the full-stack product around them, from the database to the interface.

Download CV LinkedIn ↗GitHub ↗mujtabashah1204@gmail.com · Lahore, Pakistan

How I think

My obsession is LLMs that stay grounded — not just impressive on benchmarks, but reliable when the retrieval fails, the query is ambiguous, or the user asks something the knowledge base was never built to answer.

01LangGraph · Qdrant · Cross-Encoder

RAG that self-corrects

Standard retrieval-augmented systems fail silently — bad retrieval still produces confident-sounding output. My systems grade their own answers, rewrite failing queries, and fall back gracefully. I pushed top-5 retrieval precision from 72% to 89% with cross-encoder reranking over Qdrant.

02Telnyx · Retell · Vapi · GPT-4o

Voice AI at sub-500ms

Real-time conversation is an orchestration problem: speech-to-text, LLM inference, text-to-speech, and routing must all happen before the user notices a pause. I've built and shipped voice agents on Telnyx and Genesys with end-to-end latency under 500ms, handling dynamic agent provisioning mid-call.

03FastAPI · Next.js · Docker · PostgreSQL

Full-stack engineering, not just models

From designing vector search architectures in the backend to building the React dashboards that surface insights — I own the full picture. I've delivered production systems on Google Cloud, FastAPI, and Next.js, with async orchestration through Make.com and Redis-backed caching.

I started as a full-stack developer at Hazel Mobile, moved into teaching the MERN stack at Knowledge Streams, and made a deliberate shift into AI when I joined Splendid Mark — where I built my first production RAG pipelines and NLP analytics dashboards.

At hytGenX I took it further: fine-tuning transformer models, quantizing for inference efficiency, and automating document-heavy workflows with LangGraph agents. Cutting manual processing time by 50% wasn't an accident — it was careful pipeline design.

My MS at LUMS, finished in June 2026, went at the edge cases of RAG: advanced retrieval strategies, query rewriting, and evaluation frameworks for grounding LLM outputs. Then IgniteTech, where that work met enterprise scale — GenAI embedded across products, shipped to live customers.

Years building AI

Companies shipped for

51+

Technologies in production

MS AI

LUMS, 2026

LangGraph ✦ Qdrant ✦ Claude API ✦ GPT-4o Realtime ✦ FastAPI ✦ Next.js ✦ Telnyx ✦ Cross-Encoder ✦ FAISS ✦ MCP ✦ PyTorch ✦ Docker ✦ Redis ✦ PostgreSQL ✦ Make.com ✦

Retrieval that
checks itself

Grading, query rewriting and fallbacks — instead of a confident guess

Conversation
without the lag

Voice pipelines built to a latency budget, stage by stage

Agents wired
as state

Graphs with retryable steps and a critic — not prompt chains

Shipped as
real products

Full-stack delivery: database, API, interface, deploy

01 — Projects

Projects, numbers first.

Six systems, in the order I'd defend them — mechanism first, then the number it produced.

Retrievalgrounded, or it refuses

QueryRetrieveGrade ✓Answer

↺ rewrites failing queries · up to ×3 loops · falls back to web search

Corrective RAG with self-reflective retrieval

A cyclic LangGraph state machine grades every retrieved document, rewrites failing queries and falls back to live web search — three self-correction loops before graceful degradation. Grounding checks before every answer.

LangGraph

GPT-4o

QdrantCross-Encoder

Code ↗Walk me through it ↗

Voiceno audible pause

real-time
turn budget

Real-time voice AI agent integration

Conversational platform on Telnyx and Genesys with GPT-4 dialogue management and Retell/Vapi speech — dynamic agent provisioning and live query routing across concurrent inbound calls.

Telnyx

GPT-4RetellVapi

Code ↗Hear a demo call ↗

Agentsdrafts a critic approved

reference.pdf⟶slides · quizper-instructor
tone & domain

Multi-agent content platform (planner → critic loop)

A LangGraph agent graph — not a prompt chain: a planner splits the brief, researcher and writer agents work against retrieved source material, and a critic rejects weak output back into the graph. Retryable steps, async orchestration via Make.com, concurrent multi-tenant sessions on Google Cloud.

LangGraph

GPT-4o

Claude

FastAPI

Next.js

Make.com

Code ↗See the graph ↗

Searchanswers with its source

⌕

✓ answer grounded in policy §4.2 — 2.4s

AI-powered HRMS with semantic policy search

Cloud-native HRMS with RAG-based conversational search across policies, handbooks and compliance guidelines — plain-language answers instead of minutes of PDF digging.

LangChain

HuggingFace

React

PostgreSQL

Code ↗Read the write-up ↗

Evaluation · the one everyone skipsmeasured, not vibes

Retrieval & grounding evaluation harness

Anyone can demo a RAG pipeline; the hard part is proving it got better. I keep a golden question set per knowledge base, score retrieval (hit-rate, precision@k, MRR) and generation (groundedness, citation coverage, refusal correctness), and run it as a regression gate — so a prompt tweak or an embedding swap can't quietly make the system worse.

This is where the 72% → 89% precision number came from: a measurement loop, not a guess.

PythonGolden setsLLM-as-judgeGroundedness scoring

CI gate

Code ↗See a sample report ↗

Eval run — before / after reranking

precision@5 · baseline0.72

precision@5 · + cross-encoder0.89

groundedness (judged)pass / fail per answer

Every change re-runs the set. Regressions block the merge.

Full-stack · before the AI workMERN → production

Full-stack web applications, end to end

Before AI, I shipped web products: React and Next.js frontends over Node/Express and FastAPI services, MongoDB and PostgreSQL schemas, JWT auth and role-based access, REST and WebSocket APIs, Dockerised deploys with CI. Dashboards, admin panels, HRMS modules and customer-facing apps — built, deployed and maintained.

That's why my AI work doesn't stop at a notebook: I can take a model from retrieval design through the API layer to the interface a real user clicks — and I taught this same stack to a full cohort of students.

React

Next.js

Node / Express

FastAPI

MongoDB

PostgreSQLJWT / RBAC

Docker

Code ↗See a live app ↗

One request, all the way down

◱React / Next.js UI

↓

⇄REST / WebSocket API

↓

⚙Node / FastAPI services · auth

↓

▤MongoDB / PostgreSQL · Redis

↓

☁Docker · CI · cloud deploy

02 — Experience

Where I've shipped.

Mar — Jul 2026

AI Innovation Specialist · IgniteTech (Crossover)

Owned a customer-onboarding automation platform for an enterprise GenAI product end-to-end — concept through live customer rollouts. AI-native daily: Claude Code, MCP, hands-on evals of Claude, GPT and Mistral.

Sep 2024 — Jun 2026